Marketing with AI

Marketing with AI Weekly Roundup: Choose How AI Uses Your Content, Prepare for Google’s AI Mode Ads, and Edit Whole Decks with Gemini

Robert Gillespie — Sun, 01 Nov 2026 11:01:00 GMT

Welcome to issue #57 of the Marketing with AI Weekly Roundup. Published every Sunday, it shares the top stories, playbooks, research, data and quick hits from the world of Marketing with AI over the last 7 days.

Subscribe to receive the week's most important AI marketing stories direct to your inbox.

Subscribe now

For two years, keeping your content out of AI training meant blocking each company’s crawler by name and hoping your list was current. On 30 October Google, Microsoft, OpenAI and Anthropic said they will honour one shared line in robots.txt instead, and Microsoft went a step further by letting sites block training while staying eligible to be cited in AI answers. For B2B marketers, that is the decision that matters, and it now has two separate switches.

Google also expanded ads in AI Mode to 10 more countries, with the UK due before the end of the year, and admitted on its earnings call the same day that AI Mode ads currently earn less than ordinary search ads. Gemini can now edit an entire Google Slides deck from one prompt and create finished documents, spreadsheets and decks inside the chat. And Claude is folding skills, connectors and plugins into one Plugins page, with a November deadline for Team and Enterprise admins.

The Playbook has four things to try, all built on an assistant connected to the tools you already use. John Morabito explains how an agency group chains skills into workflows that send weak drafts back for another pass. Grace Leung builds a content agent in Claude Code that drafts posts from a Notion inbox every evening. Jeff Su writes a rulebook that makes AI build Notion pages your way every time. And Wajari Velásquez shows five prompts that turn Claude, Search Console and a crawler into an SEO diagnostic tool.

1. One line in robots.txt will soon tell Google, Microsoft, OpenAI and Anthropic not to train on your site, and Bing lets you stay in its AI answers

On 30 October Google, Microsoft, OpenAI and Anthropic each said they will support a shared AI training preference in robots.txt, the file that tells crawlers what they may do on your site. It comes from the AI Preferences working group at the Internet Engineering Task Force, and it works through a “policy token” written as if it were a crawler name:

User-Agent: __usage_ai-training

Disallow: /

That group tells every participating company that the site should not be used to train AI models, and ordinary path rules still apply, so you can opt out one section or allow one folder, such as a press kit. The token covers training only. It does not affect search indexing, and it does not decide whether an assistant can read a page to answer a question and cite it, which is known as grounding. It is also not yet a standard: the proposal is an individual draft, written by Kevin Kelley, a standards lead at Anthropic, that the working group still has to finish.

The timing differs by company. Google’s Search Central post, written by John Mueller, says Google aims to support the control “in 2027”. When a site opts out with the shared token and does not mention Google-Extended, Google’s existing AI training control, Google will treat that as a Google-Extended block. Company-specific rules keep working alongside the shared one, and Google adds that sites which allow Google-Extended “are eligible to appear in responses and receive clicks via grounding in Gemini Apps”. OpenAI says it will implement the token “in the coming months”, and Anthropic has reportedly already started. A Meta spokesperson called a common signal an important step but did not say whether Meta will honour it.

Microsoft split training from AI answers. Bing’s post, “Your content, your choice”, adds two independent preferences. AI Training controls whether your content trains models. AI Grounding controls whether it helps generate answers in Bing, Copilot and Microsoft 365, with a Limited Grounding option that supports Microsoft’s own products but not the grounding services it sells to other companies’ chatbots. Both default to Allow, a page-level setting overrides robots.txt, and Microsoft says ordinary Bing rankings are unaffected. Temporary grounding controls arrive in Bing Webmaster Tools on 15 December for sites using NoArchive or NoCache, Bing starts honouring the new preferences from 1 February 2027, and after 1 March 2027 those two old directives stop counting as AI preferences.

Why this matters. Most B2B sites want to be cited when a buyer asks an assistant who to shortlist, and many have blocked AI crawlers without realising some of those rules also take them out of AI answers. The question is no longer “do we block AI?” but two separate ones: are we happy for models to learn from our content, and do we want to be quoted in the answers?

What to do. Open your robots.txt and list every AI rule in it, such as GPTBot, ClaudeBot and Google-Extended, with who added it and why. Decide training and AI answers separately with whoever owns the website. If you want out of training, you can add the shared token now so it takes effect as each company switches it on, but keep the company-specific rules until all four have shipped support. If any pages use NoArchive or NoCache, put Bing’s three dates in the diary.

2. Google is taking ads in AI Mode to 10 more countries, the UK is next, and Google admits they earn less for now

On 28 October Google said it is scaling ads in AI Mode, the conversational version of Google Search, in Australia, Canada, India, Kenya, Malaysia, New Zealand, Nigeria, Pakistan, the Philippines and Singapore, adding that “in Q4 2026, we plan to expand to additional locales, including Brazil, Mexico, and the United Kingdom”, with no dates. In the US, Google is launching multiturn ads, which can appear on follow-up questions as a conversation continues, all labelled “Sponsored”. Google says AI Mode queries are on average three times longer than traditional searches.

Search Engine Journal, citing Google’s help documentation, set out what this means for advertisers. Existing Search and Performance Max campaigns are eligible by default, there is no separate opt-out, and you cannot target AI Overviews or AI Mode on their own, although negative keywords and brand exclusions still apply. AI Mode performance is currently reported inside standard campaign and search terms reports, so you cannot see it separately. On Google’s Ads Decoded podcast, Ginny Marvin, Google’s Ads Product Liaison, said these ads are matched to both the query and the content of the AI response.

The earnings call explained the timing. Alphabet’s third-quarter results, the same day, showed Search and other revenue up 12% to $63.3 billion, slower than the previous quarter. Philipp Schindler, Google’s chief business officer, said “the strong growth in AI Mode queries had an impact because of the lower monetization we were seeing in AI Mode”. AI Mode is taking a growing share of searches and shows fewer ads, so Google’s answer is more AI Mode with ads, in more countries and across more of each conversation.

Why this matters. If you run Search or Performance Max in any of the new markets, your ads may already be appearing next to AI answers, and no report will show it separately. For UK advertisers it is months away. The searches reaching these ads are longer and more specific than the keywords most B2B accounts were built around, which changes which landing page a click should go to.

What to do. Pull the last 30 days of search terms and sort by length, because the long, conversational queries are your closest preview of AI Mode traffic. Check whether each landing page answers the question asked, not just the keyword it matched. Review negative keywords and brand exclusions now, since they are the only controls that carry over. And if you report paid search to leadership, note that AI Mode results sit inside the totals and cannot yet be separated.

3. Gemini can now edit an entire Slides deck from one prompt, and create finished files inside the chat

On 30 October Google said Gemini in Google Slides can now edit multiple slides or an entire presentation at once. From the side panel, one prompt can update content throughout a deck, refresh visual styles in bulk or match the style of another presentation. Gemini decides which slides need to change and edits the native slides, so they stay editable. The rollout began on 29 October and may take more than 15 days to appear. It covers Workspace Business Standard and Plus, Enterprise Standard and Plus, and the consumer Google AI Pro and Ultra plans, and Workspace customers get higher promotional limits for multi-slide edits until at least 1 December.

On 29 October Google also said you can now generate, preview and edit documents, PDFs, presentations and spreadsheets directly in the Gemini app. You can edit a file full screen, highlight text to ask about it, or ask Gemini to refine it. Decks can be opened in Google Slides or downloaded as PowerPoint or PDF, spreadsheets open in Google Sheets, and documents can be saved to Drive. Google says it is available to all users, including Workspace customers and personal accounts, in every language Gemini supports. PDF editing works on web and mobile, and other formats are web only for now.

Why this matters. The slow part of most decks is not the first version. It is the rework: a new campaign name across 30 slides, a refreshed brand template, updated figures in every caption the week before a board meeting. Gemini now does that in one pass and leaves the result editable, and a brief or tracker can be drafted and refined without leaving the conversation where you planned it.

What to do. Copy one deck you update every quarter, such as a campaign review or sales presentation, and ask Gemini in Slides to update it for the new quarter or match it to your current template. Go through it slide by slide and note what it changed and missed before you trust it with a live deck. In the Gemini app, try one document you normally start from blank, such as a campaign brief or a content calendar in Sheets.

4. Claude is merging skills, connectors and plugins into one Plugins page, and Team and Enterprise admins have a deadline

Claude is replacing its separate Skills, Connectors and Plugins tabs with one Plugins page. According to Anthropic’s help centre, read on 31 October, the item that used to say Customize now says Plugins. Your own plugins, including your organisation’s and ones shared with you, sit above the ones you can browse, filters narrow the list to skills or connectors, and public plugins are labelled Verified or Community. To use a skill, click “+” or type “/” in a chat. Everything already installed stays installed, but removing a plugin that packages a skill and a connector does not disconnect the connector, so you have to do that separately.

It is rolling out to Free, Pro, Max, Ultra and Team plans from 29 October, becomes generally available on 9 November, and is on by default for Enterprise from 30 November.

The admin deadline. From 30 November, on Team and Enterprise plans, if Skills are switched off for your organisation, plugins also stop working in Claude Desktop, Cowork and the Code tab, because skills are now packaged inside plugins. Admins who want plugins to keep working need an owner to turn Skills on before then, under Organization settings, Library, then Policy, where cloud code execution and file creation must also be on. Anthropic notes that its HIPAA agreement does not cover connectors, plugins or skills.

Why this matters. Plugins are now the unit a marketing team shares: a brand voice skill, a CRM connector and a set of instructions packaged together. One page makes it easier to see what is installed and what each item can reach, which is the first question to ask of any AI tool with access to company data. The Community label matters too, because a community plugin can include a connector to a service nobody has approved.

What to do. When the page reaches you, open Your plugins, then Show all, and filter to list skills and connectors separately. Remove what nobody uses and check each removed plugin’s connector has gone too. On a Team or Enterprise plan, ask your admin this week whether Skills are on, and agree before 30 November whether plugins should keep working. Then set a simple rule: Verified plugins can be installed freely, Community ones need a second pair of eyes.

1. Build workflows powered by skills, not one skill per job, and let a scorecard send weak drafts back

John Morabito runs Winston Digital Marketing and is SVP of Search Innovation at the agency group Stellar Rising. On the 26 October episode of Search with Candour, the podcast from the UK agency Candour, John Morabito spent 56 minutes on how the agency rebuilt SEO and content work around Claude skills. It is an honest account: John Morabito has built 140 skills, about 30 are in regular use, and monthly reporting is still unsolved after a year.

Step 1. Build each skill from what you already teach people. Take your SOP documentation, gold-standard examples of past work and any training videos or guides made for new staff, and give all three to Claude to build the skill. The agency’s keyword research skill takes Ahrefs or Semrush exports plus the client’s first-party data and gets the work 80% to 90% done. “The process that used to take us 8 hours now takes us two. Max.”

Step 2. Chain skills into a workflow. The line John Morabito credits to a director at Stellar Rising: “We’re not really aiming to build skills here. We’re aiming to build workflows that are powered by skills.” One skill will not replace a technical SEO audit. Several skills, each handing over to the next, can get close.

Step 3. Turn your repeated feedback into a guardrail skill. Collect the comments you keep making in reviews and edits into a skill that runs at the end of every deliverable and checks for those problems before anything goes out. John Morabito also runs a humanizer skill on every output automatically.

Step 4. Give Claude your voice from real conversations. John Morabito has recorded calls for about four years and built a skill, “write like John”, from those transcripts and old emails. A free call recorder such as Fathom, connected to Claude, is enough to start.

Step 5. Split content production into steps with a score at the end. The pipeline runs as separate steps on a dashboard: a brief built from the top results, entities and People Also Ask questions; a draft; an internal linking step trained on the client’s site; brand voice; humanizer; and a scorecard based on Google’s quality rater guidelines. “If it doesn’t meet a certain score, it kicks it back to the writer and just keeps writing it until it meets a high enough scoring.”

Step 6. Put client rules in a client skill. For a client in a heavily regulated sector, the agency moved the knowledge from the client’s own custom GPT into a Claude skill, so every draft starts from the same rules on what can be claimed.

Step 7. Add the expert interview. The missing ingredient, in John Morabito’s view, is subject-matter expertise. A 30-minute interview with someone on the client’s team beats anything written from search results alone. “If you could find the time to do it, every article would rank amazingly.”

Step 8. Use data connections, not Claude’s own browsing. The agency connects tools such as Apify, DataForSEO and SEOGets, a multi-site Search Console dashboard John Morabito says costs $29 a month, through MCP.

What to take from it. Start with the work you already know how to teach, build small skills that each do one thing, connect them, and put the quality check at the end of the chain with a score that sends work back.

Read more:

2. Build a content agent in Claude Code, no coding, that drafts posts from a Notion inbox every evening

Grace Leung, a digital growth consultant, published a 20-minute walkthrough on 30 October of building a content agent in Claude Code without writing code. The video is sponsored by Anthropic, which Grace Leung says on camera, its description includes a Max plan referral link, and Grace Leung sells a paid programme on AI marketing systems. It uses the Code tab in the Claude desktop app, and Anthropic’s documentation says the desktop Code tab and cloud scheduled tasks are available on Pro, Max, Team and Enterprise. Grace Leung’s example turns newsletters and transcripts into Instagram carousels, but the build works the same for turning webinars, newsletters or customer calls into LinkedIn posts. It goes up in five levels.

Level 0. Set up the workspace, the step most people skip. Create one folder for all your agents (Grace Leung’s is “Business OS”) with your logo, brand images and a few posts you like. In the desktop app’s Code tab, choose Local and point it at the folder. Grace Leung recommends Auto mode, which works without interrupting you but still blocks risky actions, with Claude Opus 5.5. Ask Claude to read your website and the folder and create a CLAUDE.md file, the guide it reads at the start of every session, and a brand context file covering business, audience, voice and visual style. Answer its follow-up questions.

Level 1. Build the agent with one job. Paste Grace Leung’s free “Agent Designer” prompt from the video description with a one-line description of the job. It makes Claude limit the agent to one job, ask you at most three questions, including a yes-or-no checklist for “done”, show a one-screen plan, and save the agent as a subagent of under 50 lines. Its rules include never inventing facts, numbers, quotes or dates. Start a new session, call the agent with @ and its name, and give it a transcript. Expect the first draft to need work.

Level 2. Make it reliable with a skill. Ask Claude to interview you about what to improve in that draft, and tell it to “think hard”. The answers become a skill holding the workflow and checklist. Assign the skill to the agent, and update CLAUDE.md so requests for posts go to this agent.

Level 3. Make it learn with memory. Ask Claude to create a memory file the agent reads before every job, saving only feedback that applies to future posts, not one-off edits. Give feedback as bullet points, then test on a different source to check the lesson carried over.

Level 4. Run it in the cloud on a schedule. Connect Notion, set up a Notion “content inbox” database for ideas, and ask Claude to sync the folder to a private GitHub repository, which a free account covers. Then create a scheduled task, Grace Leung’s “Daily Content Drop”, that runs each evening, turns new inbox items into drafts and suggests improvements to its own playbook. The drafts wait for review in the morning.

What to take from it. Go one level at a time: workspace, one agent with one job, a skill, memory, then the schedule. Grace Leung is clear the output still needs feedback rounds, so treat the evening run as a first draft a person approves, not something that publishes itself.

Watch more:

3. Write a Notion rulebook so your AI assistant builds every page the way your team does

Jeff Su’s 27 October video starts with a familiar problem. Ask Claude, with Notion connected, to create the same task in two fresh chats and you get two different results. Any assistant with a Notion connector knows how to use Notion. It does not know how you use it. Jeff Su’s fix is one markdown file of rules the assistant reads before it touches anything, shown working in ChatGPT and Claude, with the prompts and a starter rulebook free on jeffsu.org. Jeff Su also sells courses and a template folder, which the method does not need.

Step 1. Make a folder with the starter rulebook. Create an empty folder (Jeff Su’s is “Notion Ops”) and put the starter “Notion conventions” markdown file inside. It already holds general rules, such as confirming the target database first and using its closest template rather than an empty page.

Step 2. Mount it with Notion connected. In ChatGPT, Jeff Su creates a project and mounts the folder. In Claude, Jeff Su attaches the same folder to a chat.

Step 3. Teach one page type by example. Pick one ideal page. Jeff Su’s ideal task has a checkmark icon, a verb-first name, a pinned one-line callout and bullet points of context. Paste the teaching prompt with a link to it, and the assistant writes your task rules into the file.

Step 4. Test in a fresh chat. With the folder mounted, ask for a new task. It should match the ideal structure without you describing it.

Step 5. Teach a whole database, and explain your tags. Paste a database link rather than a page, and the assistant studies every property, option and template. Say what each tag means: Jeff Su’s “research” tag applies only when there is a source such as a website or PDF. Jeff Su then asked for a newsletter summary and a meeting prep page, and both came back correctly tagged and in the right layout.

Step 6. Repeat for every database. Jeff Su’s general prompt has the assistant explain back how it thinks you use a database and ask about anything unclear before writing rules. In the demo, Claude asked the difference between three status options, exactly what a new colleague would get wrong.

Step 7. Write down workflows that span databases. With Jeff Su’s project workflow in the rulebook, one prompt produced a project page with its properties filled in, five standard deliverables and linked tasks in the right format.

What to take from it. This is the difference between an assistant that adds to the mess in your campaign tracker and one that keeps it tidy. Start with the database you use most, probably campaigns or the content calendar. And the idea is not specific to Notion: your CRM or project tool can have its own rulebook in the same folder.

Watch more:

4. Five prompts that turn Claude, Search Console and a crawler into an SEO diagnostic tool

Wajari Velásquez is an SEO consultant who runs a one-person consultancy, working with designers and developers across Spain, Latin America and Europe. On the WP Remote Show on 27 October, Wajari Velásquez showed five examples from current clients of what changes when Claude is connected to the usual SEO tools, estimating that an audit which took three days now takes about an hour. The show is produced by WP Remote, a WordPress maintenance tool, and nothing in the method depends on it.

The set-up. Claude is connected through MCP to Google Search Console, Google Analytics, PageSpeed Insights, the crawler Screaming Frog and Sistrix, a paid visibility tool, so it can answer questions across sources you used to check one at a time. Wajari Velásquez runs every audit to a written standard procedure.

Prompt 1. Run the full audit to your own SOP. Ask Claude to audit the site against your procedure: crawling and indexing, internal linking, canonical tags and hreflang. The deliverable is each problem with its fix, not a PDF of problems.

Prompt 2. Find out why impressions dropped. For a client whose impressions and clicks had fallen, the prompt was simply to connect to Search Console and identify the drops. In minutes Claude found a Cloudflare security setting blocking Googlebot.

Prompt 3. Hand the developers a shareable fix. Google was still indexing PDFs a client had blocked in robots.txt, a common mistake, because blocking a crawler does not remove pages already indexed. Claude produced the analysis and fix as a Claude artifact, a shareable page, and the link went straight to the developers.

Prompt 4. Find the blog posts losing ground. Ask Claude to analyse blog articles only, from the past year, that are losing impressions and clicks, and return the 20 most worth recovering using Sistrix data alongside Search Console. For one client those posts had lost 43% of their clicks in six months.

Prompt 5. Match AI Overview citations to leads. Ask how many AI Overview citations the site has and which topics generate them, then which cited content actually produces leads. One client had more than 500 citations, and the useful part was separating informational content that helps the brand but produces no enquiries from the pages that lead to a contact form. “Less visits, more conversion.”

What to take from it. The prompts are simple. The value comes from several data sources connected at once and from knowing which question to ask. Wajari Velásquez’s warning is that you have to be “the pilot” at every step, and to check your conversion tracking before trusting any analysis built on it, because that is where most projects fail.

Watch more:

1. Ahrefs: half or more of AI citations disappear the next day, and three-quarters of those come back

Ahrefs followed about 37,500 citations across five AI tools every day for 20 days, publishing on 28 October. It asked Google AI Mode, ChatGPT, AI Overviews, Gemini and Copilot the same questions daily, from 563 questions on AI Overviews to 3,173 on ChatGPT, and tracked every site cited on day one. Ahrefs did not give the study’s calendar dates.

At least half the citations had gone by the next day on every tool: about 60% on AI Mode and ChatGPT, 50% on AI Overviews and 80% on Gemini and Copilot, which often cited nobody. But about 75% of the lost citations came back within the 20 days, half within three days, and only 20% to 27% of sites cited on day one never returned. Citations that lasted eight days or more almost always came back after dropping out.

Pages that already rank in Google and get organic traffic held their citations longer, while Domain Rating and backlinks made little difference. When a question named a brand, its own site stayed cited for all 20 days in about 30% of cases, four to seven times the average. In 22 questions about Ahrefs itself, its own site was cited almost every time for facts such as pricing, but Google’s AI tools preferred third-party sources for reviews and comparisons.

What to take from it. Do not judge AI visibility from one check: count how many of the last seven days you were cited, and Ahrefs suggests brand mentions may be a steadier measure. Ordinary SEO still does the heavy lifting, since ranking pages are the ones AI keeps citing. For review and comparison questions, the sources that count are other people’s sites, which is a case for review platforms and digital PR. Ahrefs sells Brand Radar, which tracks AI citations.

2. Google’s head of search research: questions are three times longer and buyers want human voices alongside the summary

On Ads Decoded on 28 October, Ginny Marvin talked to Shivani Mohan, Google’s Vice President of Data Science and User Experience Research for Search, about how search is changing. Google gave no sample or method, so treat these as Google’s own observations.

Shivani Mohan said AI Mode queries are three times longer than traditional searches and “one in every six searches is multimodal”, using images, voice or files. On the earnings call the same day, Sundar Pichai said one in five AI Mode searches use voice, images or documents, a different measure, so the two figures should not be compared. Follow-up questions add context rather than signalling a bad answer, searches are moving earlier in the buying journey, and people want human perspectives, such as podcasts, reviews and YouTube, next to the summary. In Shivani Mohan’s words, “generic answers” are “not hitting the mark”. Ginny Marvin added that lead generation businesses should build content for the specific pain points buyers describe.

What to take from it. If buyers ask longer questions earlier, the pages that win answer a specific situation, not a keyword. Take the questions from your search terms and sales call notes and make sure each has a page that answers it directly. And put your own experts on camera and on podcasts, because that is the content Google says people want next to the summary.

3. OpenAI: people are using ChatGPT for a wider range of work, not just the same few tasks

OpenAI published data on 26 October on how people use ChatGPT for work, covering consumer plans only, without Enterprise or Codex. It classified work messages into 165 activities from O*NET, the US government’s occupations database, across nearly 100 countries, quarterly from early 2025 to the third quarter of 2026, automatically and with statistical noise added for privacy.

Across the 74 countries observed every quarter, the ten most common work activities fell from 56.0% of classified work messages to 51.1%, so use is spreading across more kinds of work. “Developing marketing content” is one of the activities named. OpenAI sells the product it is measuring.

What to take from it. If your team’s AI use is still mostly drafting copy and summarising documents, ask each person to list the three tasks they now use it for and share the list. The tasks only one person has found are the ones worth teaching everyone.

1. Canva Brush is a new AI model that edits your existing designs, from one line of copy to a whole deck

Canva launched Brush this week, a vision-language model for design from Canva Research, a week after its Craft image model. Brush reads an existing design and makes the changes you ask for, such as restyling images, changing every header in a deck or turning a slide into a social graphic, and the result stays editable. Canva says it is up to 69% faster on general edits and 46% cheaper per edit than its previous design agent, which are Canva’s own figures. It is available now to Canva Pro subscribers by selecting Brush in Canva AI, with wider availability “coming soon” and Brand Kit support still in development. Try it on repetitive edits, such as resizing a campaign set or updating a deck for a new event.

2. Microsoft Advertising builds your first Performance Max campaign from your brand guidelines, and lets you set 40 rules for AI-written copy

Microsoft Advertising’s October update, published on 28 October, lets new advertisers share their website, marketing plans, brand guidelines and product images, which Microsoft turns into an editable snapshot of the business before building a Performance Max campaign. For all advertisers, messaging guidelines are in open pilot for Performance Max and Search campaigns with AI Max: up to 40 rules on claims, tone and wording for AI-generated copy. The Conversions API, which sends conversions from your site, CRM or app server to server, is now generally available, directly, through Google Tag Manager or partners such as Segment and Tealium. Performance Max also gains age and gender targeting, search term and auction insights reports, and landing page reporting from mid-November. The post does not state regions. Write your 40 guidelines once and reuse them on any ad platform that lets AI write copy.

3. HubSpot will let Claude, ChatGPT and Copilot read your Sensitive Data, if a Super Admin switches it on

HubSpot’s October developer rollup, on 30 October, adds a public beta that lets Super Admins allow supported AI connectors to read Sensitive Data in accounts that have it turned on. ChatGPT, Claude and Copilot can then read Sensitive Data properties and engagements, while Gemini and HubSpot’s remote MCP server read engagements only. Access is read-only and users still need the CRM permissions. Before anyone switches it on, agree with whoever owns data protection which properties count as sensitive and which assistants may use them.

4. Gemini in Google Sheets now shows the web source behind each AI-generated cell

When Gemini fills cells in Google Sheets, it now shows the web sources it used inside each cell, according to a Workspace Updates post on 28 October, and one request can now cover up to 3,000 cells. It is on by default for users with Gemini in Sheets, with a gradual rollout that began on 20 October. For competitor lists, account research or market sizing, that turns a column you have to recheck by hand into one you can audit cell by cell.

1. OpenAI cancelled the Pro 200 usage cut and doubled Work and Codex usage on its other Pro plans.

On 29 October OpenAI cancelled the reduction to the Pro 200 plan’s Work and Codex allowance announced on 29 September. Subscribers on the lower allowance have had it doubled to match, Pro 100 and Pro 500 now include twice their previous allowance, and Plus is unchanged. If your team budgeted for the cut we flagged this month, you no longer need to.

2. Gemini quietly stopped tagging the links it cites, and its iPhone app sends no referrer.

Google added utm_source=gemini to cited links in early September, then stopped around 28 September, according to the SEO Natzir on 27 October and Search Engine Roundtable on 30 October. Desktop Gemini still passes the referrer, but the iPhone app sends none, so those visits land in Google Analytics 4 as Direct. As with the ChatGPT tracking bug in our 17 October issue, check Direct before concluding AI referrals have fallen.

3. Semrush says some Google results served to SEO tools are “poisoned”.

On 27 October Semrush said some Google results reaching its systems are invalid, from whole pages of junk to loosely matched results, and that it now detects and excludes them. It gave no figure or start date, and Semrush sells the rank tracking it describes. Last week Ahrefs reported the same problem on 5% to 20% of requests. If a ranking swing looks strange, check Search Console first.

4. OpenAI released GPT-6.1 Terra, a cheaper model for computer use.

GPT-6.1 Terra arrived on 26 October in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu. OpenAI says it understands what is on screen, acts across apps, substantially outperforms the previous Terra model on its agent benchmark, and costs 50% less in the API, all OpenAI’s own figures. If you use ChatGPT Work for browser tasks, compare it in Advanced.

5. Bing is now showing ads inside its Copilot-powered AI Overviews.

Search Engine Roundtable reported on 30 October that Microsoft is showing ads within Bing’s AI Overviews. With Google expanding AI Mode ads the same week, check whether your Microsoft Search campaigns are serving there and treat AI answer placements as part of the normal search mix.

6. Advertisers used Gemini to make more than 1.5 billion ad assets in three months.

On its 28 October earnings call, Google said Gemini helped advertisers create more than 1.5 billion campaign assets in the third quarter, over seven times a year earlier, without saying how many ran or how they performed. Check which automatically created assets are switched on in your Google Ads account and what they say.

1. You now decide how AI uses your content, one choice at a time. The shared robots.txt token separates training from everything else, and Bing splits training from AI answers page by page. Meanwhile Google’s AI Mode ads are on by default with no separate opt-out, and Ahrefs found AI citations come and go daily. The controls are getting more precise, but the defaults still favour the platforms. Write down what your company wants on training, AI answers and AI ad placements, so each is a decision rather than an old robots.txt rule or an unchecked setting.

2. The assistant is becoming the place the work gets done. Gemini now edits whole decks and creates finished files in the chat, and Claude has put every skill, connector and plugin on one page. Grace Leung’s agent drafts posts while nobody is at a desk, Wajari Velásquez audits a site across five connected tools in one conversation, and HubSpot is opening more of the CRM to assistants. For 2027 planning, the question is less which tools your team logs in to and more what your assistant can reach, and who checks its output.

3. Written rules are beating clever prompts. Jeff Su’s Notion rulebook, Grace Leung’s CLAUDE.md and memory file and John Morabito’s guardrail skill all write down how your team works, once, so every AI output starts from it. Microsoft’s 40 messaging guidelines apply the same idea inside an ad platform. Consistent results come from a short set of rules files kept up to date, and someone on the team should own them.

The problem wasn’t Claude. It was my folders.

In January I started handing Claude whole marketing jobs instead of single prompts. Plenty came back needing redoing. When I dug into why, my files were a mess and Claude was guessing.

So I wrote it a welcome guide, the way a good Airbnb host does. Now work that used to take two or three hours comes back in about twenty minutes, in my voice.

My Claude Operator System masterclass shows you how to set up the same system. No code, and nothing goes out without your yes.

(Now only $150 or 2 payments of $75. Individual price $953).

Show me how it works

Tell me: What’s your biggest takeaway from this week’s news from the world of AI Marketing? Love to hear your thoughts in the comments.

Leave a comment

Thanks for reading the Marketing with AI Weekly Roundup! If you enjoyed this read, the best compliment I could receive would be if you shared it.

Marketing with AI Weekly Roundup: Build a Claude SEO System, Test Ten Video Ad Angles for $3.30, and Check Meta’s AI Ad Edits

Robert Gillespie — Sun, 25 Oct 2026 11:00:58 GMT

Welcome to issue #56 of the Marketing with AI Weekly Roundup. Published every Sunday, it shares the top stories, playbooks, research, data and quick hits from the world of Marketing with AI over the last 7 days.

Subscribe to receive the week's most important AI marketing stories direct to your inbox.

Subscribe now

Three studies landed in the same week arguing that AI search sends you almost nothing. Marie Haynes looked at a year of analytics across 100 websites and found ChatGPT sent 0.62% as many visits as Google organic search. SEO Works found organic search delivered 39 times the traffic of AI assistants, and Kevin Indig could not find a reliable link between AI visibility and branded or direct traffic. Within a day, the rebuttals arrived, and their main argument is one you will recognise from last week’s lead: AI visits keep landing in the Direct bucket, where nobody counts them.

Anthropic opened a research preview of Claude Duet, a separate app where Claude works as an always-on agent across your connected tools, sending a morning brief and asking before it acts. Meta’s new terms say advertisers are “solely responsible” for whatever its AI features create or change in their ads, including features Meta switches on itself, and the date those terms take effect has moved from 30 October to 20 November. Salesforce agreed to buy Braze for about $5.66 billion and marked its Marketing Cloud Campaign Agent as generally available.

The Playbook has two things to try. Eliot Prince walks through an eight-step Claude system for SEO and AI search that ends in a live dashboard, and Creatify shows how to make ten B2B video ads, each testing a different buying reason, for about $3.30.

1. Claude Duet is an agent that watches your apps, works while you’re away and asks before it acts

On 21 October Anthropic opened a research preview of Claude Duet, a new app where Claude works as what Anthropic calls a proactive agent. It is available by waitlist, in the US only, to Pro and Max subscribers, with invitations going out in waves. It runs as a standalone app on iOS and Android and on claude.ai, it is included in Pro and Max at no extra cost, and Anthropic’s help centre says it is not yet available on Team or Enterprise plans, with “plans to bring it to those plans later” and no date.

Each morning Duet sends you a short brief on the most important items across your connected apps. Through the day it checks those apps, every few hours according to the announcement, and messages you only when something needs your attention. You can also hand it work, from scheduling meetings to building dashboards, sales forecasts, prototypes and web apps. It has its own virtual machine and browser, so your laptop and phone can be off while it works. It draws on the same connectors and plugins as your Claude account, more than 6,000 of them in the Claude Marketplace, and Anthropic’s own examples include HubSpot follow-ups, Canva designs and drafting a LinkedIn job post for your approval.

The safeguard is the approval step. Before Duet sends a message, makes a payment, signs a document or deletes anything, it shows you exactly what it is about to do and waits. If you cannot approve straight away, it finishes the rest of the task and brings the approval back later. Some things are off limits even with approval, including trading securities, moving money between accounts and accessing someone else’s account. It runs on Claude Opus 5.5 with no model choice, and it uses your plan’s normal five-hour and weekly limits, with its background checks counting towards them. Anthropic is giving Duet users one full limit reset, available until 30 November. For now it cannot connect to Microsoft 365, it cannot make phone calls or send texts, Slack works “with limits”, and you cannot watch or take over its browser.

Brock Mesarich, a YouTube creator who got in from the waitlist, tested it live on a phone. The test showed Channels, which give each topic its own space for messages, tasks and files, a Tasks list that included items Duet had added itself, such as an important inbound email it moved onto the next day’s list, and a Make library holding everything already created across Claude chat, Cowork and Claude Code. The limit Mesarich hit is a useful one to know. Duet uses the connectors already on your Claude account, and there was no way to add a new one inside Duet, so for the email platform Drip, Mesarich set up a Zapier MCP server first, then asked Duet for a dashboard of the last campaign’s opens, clicks, unsubscribes and delivery rate, which arrived in about a minute.

Why this matters. Duet follows OpenAI’s Dots and Microsoft’s Copilot agent accounts as the third launch in a month to give an assistant a standing job rather than a single task. The difference is the morning brief. Instead of you going to the assistant with a task, the assistant comes to you with a list, and the quality of that list depends entirely on what it is connected to and what it is allowed to do.

What to do. If you are outside the US, as many readers are, you cannot use Duet yet, but the part that decides how useful it will be is already in your hands: the connectors on your Claude account. Review them now, remove the ones nobody uses, and add the ones your weekly work depends on, such as your CRM, email platform, analytics and calendar. If you are on Pro or Max in the US, join the waitlist and start Duet on read-only access to your inbox, calendar and CRM before you let it send anything. Whichever applies, write down your approval rules now: what an agent may send without asking, what always needs a person, and who gets the request when you are away.

2. Three studies say AI search sends you almost nothing, and the rebuttal says you are not counting it properly

On 21 October Marie Haynes published “GEO is a distraction”, based on 12 months of Google Analytics 4 data from 100 websites of varying sizes and more than 145 million sessions. Sessions from ChatGPT came to just 0.62% of the volume driven by Google organic search, rising to 0.73% with every other AI assistant included. That is a ratio to Google organic, not a share of all traffic. Of the 55 sites tracking conversions, ChatGPT visitors converted at 13.8% against 3.02% for Google organic, more than four times the rate, yet Google organic still drove 96.5% of the conversions across the channels analysed and 27 times as many as all AI assistants combined. Haynes also reports that ChatGPT referrals stopped growing in August 2025 and have declined since. The post gives no date range or industry mix, and an AI tool was used to total the figures across the 100 properties.

Two more studies followed. SEO Works looked at 95 websites from September 2025 to August 2026 and found organic search sent about 39 times the traffic of AI assistants, 2.83 million sessions against 72,095, and that for every 100 conversions from organic search, AI referrals delivered 3.2. AI traffic converted worse than paid search and better than direct traffic. Its AI share of organic rose from between 2% and 2.8% to 3.9% by August. Kevin Indig ran tests across 12 sites in different sectors and could not find a reliable relationship between AI visibility and later branded search or direct traffic. That is an absence of evidence across a small sample rather than proof of no effect, and Indig notes that companies with outstanding AI growth may be able to see it.

The pushback came within a day. iPullRank argued that the studies measure only the visits that happened to carry a referrer. Its first point is dark traffic: assistants strip the referrer, people copy links out of answers into a browser, and iPullRank cites Indig’s own estimate that around 70% of AI traffic is attributed to Direct. Its second is that clicks from Google’s own AI Overviews and AI Mode usually arrive as Google organic, so the biggest AI search surface sits inside the bucket the studies use as the benchmark. Salience, a UK agency whose clients are mostly retailers, ran the same comparison across its portfolio and found AI assistants grew from 0.09% of Google organic sessions in January 2025 to 2.19% by September 2026, the same order of magnitude as Haynes but still growing.

Meanwhile, the place Google puts its own AI is growing fast. Semrush tracked 400 US desktop commercial keywords across banking, software and ecommerce from September 2025 to August 2026 and found AI Overviews on 88% of them in August, up from 45.75%, and on 97.08% when weighted by search volume. In software, every “alternatives” and “best X for Y” search in the panel showed an AI Overview. Semrush measured whether the overview appeared, not whether anyone clicked, and it sells the tools that track this.

Why this matters. Last week a bug moved tens of thousands of ChatGPT visits a day into Direct and nobody noticed for a week. This week the debate about whether AI search is worth funding rests on the same referral counts. Both sides can be right: assistants send few clicks you can attribute, and much of their influence happens before a buyer types your name into Google or your address into a browser. The decision you face is not whether to believe Haynes or iPullRank. It is whether you have numbers of your own.

What to do. Pull 12 months from your own GA4 and put three lines on one chart: ChatGPT and other AI assistant sessions, Google organic and Direct. If Direct rose while AI referrals stayed flat, that is worth a conversation before anyone cuts a budget. Ask new leads how they first heard of you and add “an AI assistant” as an option. And check your “alternatives”, “vs” and “best tool for” pages, because those are the searches where an AI Overview now appears every time, and the page Google reads is the one you control.

3. Meta’s new terms make you “solely responsible” for its AI ad edits, and the start date has moved to 20 November

Meta’s revised Commercial Terms add a section headed AI Features. On Meta’s own preview of the terms, which we read on 24 October, it says AI features “may make mistakes”, and that “You are solely responsible for and assume any and all risks associated with such AI features and any resulting Content.” According to Adweek’s reading, reported on 22 October, that reaches features Meta turns on itself, not only the ones an advertiser chooses. Agencies told Adweek that Meta’s tools have been adding sitelinks, translating copy and generating images inside live campaigns without being asked, and that the media agency Exverus pulled running ads to strip those changes out. Meta says advertisers can switch off individual Advantage+ features, and describes the update as consolidating its generative AI ad terms. Nobody signs anything: continuing to use Meta’s products after the terms take effect counts as accepting them.

The date has moved. On 23 October a Meta spokesperson told Reuters the company is extending the effective date of the updated Commercial Terms and Self-Serve Ad Terms to 20 November, “to allow more time for us to explain to our partners and stakeholders what’s changing and what’s not”. The move followed complaints from publishers and advertisers to EU and UK competition authorities about the crawling clause. When we read Meta’s preview page on 24 October, it still said the terms take effect on 30 October, so the only source for 20 November so far is Meta’s statement to Reuters. The crawling clause is still there, authorising Meta to crawl the websites associated with your business, but the sentence about using that content to train AI has been removed, and the stated purpose now ends with providing and improving “advertising and business products”.

Meta also gave more detail on its AI tools on its third-quarter earnings call on 21 October. Nearly 10 million advertisers used at least one of its generative AI creative tools in the quarter. The claim that advertisers who adopt its AI business assistant’s recommendations see almost 5% more conversions at the same cost, which we noted last week came with no method, is an estimate comparing adopters with similar accounts that did not adopt. Because advertisers chose whether to adopt, that is a correlation, not a test.

A correction. Last week’s Looking Ahead gave 30 October as the date Meta’s new terms take effect. Meta now says 20 November.

Why this matters. For years the deal with automated ad tools was that the platform optimised and the advertiser approved. These terms put the platform’s edits and the advertiser’s liability in the same contract, at the moment those edits are becoming more common and harder to spot. For a B2B brand, a translated headline that misstates a product claim, or an image showing a feature you do not have, is a compliance problem as well as a creative one.

What to do. Before 20 November, open every live Meta campaign and list which Advantage+ creative enhancements are switched on, including sitelinks, translations, image generation and text variations. Switch off anything nobody has approved, screenshot the settings with a date, and keep the screenshots with your campaign records. Then add a weekly check of what is actually serving, because the change you need to catch is the one the platform makes after launch.

4. Salesforce is buying Braze for $5.66 billion, and its Marketing Cloud Campaign Agent is now generally available

On 20 October Salesforce signed a definitive agreement to acquire Braze, the customer engagement platform, for about $5.66 billion in enterprise value, paying $47.50 a share in cash. The deal is expected to close in Salesforce’s fiscal year 2028, subject to approval by Braze shareholders and regulators. Braze serves about 2,770 customers, including Canva, Instacart and Venmo, and sells real-time messaging across apps, websites, email, SMS and WhatsApp, with BrazeAI choosing content, offers, channels and timing. Salesforce says Braze is expected to strengthen Marketing Cloud’s cross-channel personalisation and bring its mobile and engagement expertise to Marketing Cloud. Everything said about the combined product is forward-looking, and nothing changes for customers of either company until the deal closes.

Two days later Salesforce published what Marketing Cloud Winter ‘27 means for marketers, describing it as more than 60 releases and “our biggest wave yet of agentic capabilities”. The labels are the useful part. Campaign Agent is GA now, and that label sits under its agentic campaigns section. Account Discovery Agent and Personalized Paths in Flow go generally available in October. A section on agentic content describes Campaign Agent turning campaign intent into on-brand, multilingual content across email, SMS, RCS and MMS, grounded in your customer data and brand guidelines, but carries no availability label of its own, so do not assume it is switched on for you. The post names only Marketing Cloud Next as an edition.

Closing a loop. Last week we said that Salesforce had described marketing agents for Winter ‘27 but that their availability by edition was not stated on the pages we could read. Campaign Agent now has a GA label. The edition question is still open for everything except Marketing Cloud Next.

Why this matters. Braze’s customer list is mostly consumer brands, so most B2B readers will not feel this deal directly. What they will feel is the direction. The two biggest CRMs are both buying or building the layer that decides what message goes to which person and when, and handing more of that decision to an agent. If you are planning a 2027 martech budget, the question to ask your Salesforce account team is not what Braze adds but what it replaces, and on what timeline.

What to do. If you run Marketing Cloud, ask your admin which edition you are on and whether Campaign Agent and Account Discovery Agent are enabled, then test Campaign Agent on one low-risk nurture before anything customer-facing. If you use Braze, note that nothing changes before the deal closes, and put a review of your contract’s renewal terms in the diary. If you use neither, take the hint and ask your own platform vendor for their agent roadmap in writing.

1. An eight-step Claude system for SEO and AI search that ends in a live dashboard

Eliot Prince, who worked in SEO for ten years before teaching AI, launched a new business website a month ago and used its first month of data to film a full walkthrough of running SEO and AI search with Claude. The video, published on 22 October, runs 43 minutes, and the method works for any company with a website, Search Console and Google Analytics. Prince sells a paid community and an SEO plugin for Claude, both mentioned in the video, but every step below uses general tools and prompts, and the prompts are in a library linked from the video.

Step 1. Connect the data before you ask anything. “Most people skip this and it dooms the rest of your work because AI just ends up guessing.” Ask Claude for an SEO audit with no data connected and it has nothing to work from. Prince adds three connectors under Customize, then Connectors. First, Composio, a middleware service that gives Claude access to apps it has no native connector for, used here to connect Google Analytics and Google Search Console through Google’s normal sign-in and then added to Claude as a custom connector. Second, DataForSEO, an SEO data API added as a custom connector, which supplies search volumes, keyword difficulty, competitor keywords, backlinks, page speed and AI mentions; it starts with about $5 of free credit, and Prince expects a $50 balance to last nearly three years. Third, Firecrawl, already in Claude’s directory, which lets Claude read many more pages than its normal web search. The test once they are connected: ask Claude, through Composio, how many visitors your site had in the last seven days, and ask DataForSEO for the monthly searches on one keyword you care about. A term with a high cost per click but low ad competition is the combination worth hunting for.

Step 2. Build a Claude Project that acts as your SEO adviser. Create a project, then give it custom instructions that tell it to use real numbers from the connected tools, read the live page before commenting on it, rank every recommendation by payoff and explain any jargon in plain English. Ask it “who are you and what do you do?” to check the instructions took. Everything from here happens inside this project, so every chat inherits the context.

Step 3. Have Claude write an “about my business” document from your website, then fill the gaps. The prompt asks Claude to act as the positioning consultant April Dunford and build a facts document from the site, so every later prompt and any AI assistant reading the site works from accurate information. It reads the site page by page and comes back with questions about what the site does not say. In the walkthrough the gaps included pricing, who the product is and is not for, competitive alternatives, proof and launch date. Prince answered by dictation, then added the finished document to the project with one click.

Step 4. Run a buyer-question gap audit. This prompt reads the site as a first-time buyer and lists the questions they would ask across what Prince calls the big five: cost, problems, risk, comparisons and alternatives. It also asks what an AI assistant reading the site would conclude. It returns one table ranked by how likely each gap is to cost you. The audit found 15 gaps, and the top one was that the membership price was only visible at the payment page while every competitor showed it, which means an AI asked what the membership costs would have to guess. Save the result as a document and add it to the project.

Step 5. Run the full SEO and AI search audit. With the context, the gap audit and the connectors in place, one prompt runs quick wins, an on-page review, a technical audit and an AI search check that asks assistants who they would name instead of you. It took 10 to 15 minutes. The scorecard came back red for quick wins, on-page and AI search, and amber for technical. One finding stood out: search traffic was growing, but almost every search that found the site already knew the brand name. The audit also produced a ranked top ten fixes, an offer to write the structured data that tells search engines and assistants what the business is and who runs it, and the names of competitors being recommended instead. Add the audit to the project.

Step 6. Build the strategy in a fresh chat. Starting a new chat is deliberate, to show that the project carries the context. The strategy prompt asks for two things: changes to the money pages, the pages that target a buyer already looking for what you sell, and a topical map of blog clusters that builds authority, here at three posts a week. The output explained each issue in plain English, for example that a competitor had 93 other websites linking to it against the new site’s two, recommended new pages backed by search volume, scored each task for business value against difficulty, and laid out a 14-week sequence and a 30-day plan. Prince’s own reaction was that existing YouTube videos already covered most of the cluster topics, so the blog plan was mostly a repurposing plan.

Step 7. Execute week one, with publishing switched off. Ask Claude to walk through week one and write the finished assets: page titles, opening copy, FAQs, internal links and structured data. Because the site is built in Replit, Claude wrote the changes as instructions for that tool, and it would have worked differently for a WordPress site. Prince switches off the Replit connector at this stage because it “has a habit of autopublishing”, and reviews before anything goes live. Blog posts are never written from nothing: the blog skill interviews Prince first, or works from the transcript of something already taught, so each post carries original insight rather than a rehash of what is already online.

Step 8. Build a live dashboard you can read in 30 seconds. The final prompt asks for a one-page dashboard showing whether things are improving and what to do next, calling the connectors for the top queries and biggest movers, with “every number labeled with its date range”, one screen and no clutter. Prince changed the prompt to save it as a live artifact rather than a static HTML file, so it refreshes. It took about seven minutes and showed daily search clicks, top queries, top pages, progress against the 30-day plan and where each number came from.

What to take from it. The order matters more than any single prompt. Data first, context second, gaps third, and only then the audit, so that every recommendation is grounded in your numbers and your business rather than a generic checklist. Each output goes back into the project, so the next step starts from everything learned so far. And the two human checkpoints in the method, publishing switched off and real expertise in every post, are where the quality comes from.

Watch more:

2. Ten B2B video ads, each testing a different reason to buy, for about $3.30

This one is a vendor’s own tutorial. Creatify, which sells AI video ads, published it on its channel on 22 October to promote its new low-cost video model, Boreal. The method transfers to any avatar video tool, and the cost figures are Creatify’s own.

The problem it addresses will be familiar to anyone running paid social for a B2B product: you do not know which message will land, whether that is saving money, saving time or having more control, and every test needs a video, so most teams test one or two angles because ten videos cost too much. The presenter makes ten talking-head ads for a real B2B product, Ramp’s corporate card, aimed at startup founders and CFOs.

Step 1. Turn each buying reason on the product page into an angle. Read the product’s own website and treat every reason it gives to buy as a separate ad angle: real-time visibility of spend, expenses submitted by text, no personal guarantee for founders, no personal credit check, catching overspending before month end, and so on. One angle per ad, so the results tell you which reason resonates.

Step 2. Pick a presenter that fits the buyer. In the avatar library, filter for a presenter in a professional setting, because the buyer is a finance lead, and use a different presenter for each ad so the test also varies who is speaking.

Step 3. Write a ten-second script per angle. Creatify’s script writer will read a product URL and draft a script from a prompt that names the product, the audience, the tone and the structure, at no credit cost. The presenter had prepared ten scripts in advance, each a one-line hook, one benefit and a call to action, such as “Expense reports eating your Fridays? With Ramp, your team submits expenses by text.”

Step 4. Generate on the cheap model. Boreal costs one credit per 15 seconds of video, against about five for Creatify’s premium Aurora model. An optional behaviour prompt sets how the presenter delivers the line. On the Pro plan, where 300 credits cost about $99, the ten ads came to about $3.30, against about $16.50 on the premium model. There is also a one-prompt shortcut that builds all ten at once, which costs more, around 17 credits, because it creates presenters from scratch.

Step 5. Check every ad like a creative director before it goes live. Is the product name pronounced correctly? Does the lip sync hold? Does the voice sound like a person? And check every claim in each script against the product page one more time. Fixing a bad ad means editing the script and rendering again for one more credit.

Step 6. Test cheap, then remake the winner properly. Run the ten angles, let the results tell you which reason to buy wins, and only then spend on a premium model or a real shoot for that angle.

What to take from it. The useful idea is not the model. It is splitting your offer into one reason per ad and testing them side by side before you spend a production budget, which used to be too expensive to do with video. Two cautions for B2B: check your ad platform’s rules on disclosing AI-generated people before launch, and be wary of avatar ads for brands that sell trust, where a real customer or employee on camera may still win.

Watch more:

1. Forrester: marketers will scale back a third of their AI initiatives in 2027 as the bills arrive

Forrester’s Predictions 2027 for CMOs, published on 21 October by seven analysts led by Mike Proulx, is titled “CMOs Reckon With An AI Reality Check”. Its headline prediction is that marketers will scale back a third of their AI initiatives because of cost. As martech vendors move to usage-based pricing, every prompt, agent interaction, generated asset and automated workflow carries a charge, and Forrester expects procurement and finance teams to take a bigger role in AI decisions as those costs get harder to forecast. CMOs, it says, will focus budgets on the use cases that deliver measurable value. Its second prediction is that AI agents will supersede legacy marketing and insights tools at one in ten companies, as teams orchestrate work through agents rather than moving between separate analytics, planning, creative and campaign tools. A third is that five brands will suffer AI-driven growth setbacks while adoption carries on regardless.

The supporting figures come from Forrester’s Q3 2026 CMO Pulse Survey of US B2C marketing executives: 82% expect to give stronger justification for future AI investments, 78% say leadership demands stronger evidence of value than a year ago, and 64% believe their use of AI is likely to produce a brand-damaging incident within the next 12 months. Forrester does not publish the sample size or fieldwork dates in the post.

What to do. Treat this as the brief for your 2027 planning conversation. List every AI tool and agent your team pays for, including the ones bundled into HubSpot, Salesforce and your ad platforms, and next to each write what it costs per month on current usage and the one metric it moves. The ones you cannot fill in are the third Forrester expects to be cut, and it is better that you cut them than finance. The survey figures are B2C, so read them as a direction for B2B rather than a benchmark.

2. Demandbase: 72% of B2B marketing teams have not grown in two years, and only 21% rarely have to correct AI output

Demandbase released research at its GO conference in New York on 21 October, commissioned from G2, which conducted in-depth interviews in September 2026 with 124 US B2B marketers, from individual contributors to directors, in demand generation, growth, digital or field marketing at companies with 250 or more employees. 72% said their team is the same size or smaller than two years ago, while 70% face shorter turnaround times, with company leadership the biggest source of the pressure. 56% said strategy, creativity, judgement, communication and leadership matter more today than two years ago, against 38% who said the same of AI, prompting, tool fluency and data literacy. Only 21% said they rarely or never have to correct AI output. On AI search, 62% have changed their search strategy and 64% have tried to influence how AI describes their brand, but only 35% systematically track how their brand appears. 56% named processes and cross-functional workflows as their biggest challenge, against 21% for the tech stack, data and tools, and 74% expect their jobs to change more in the next two years than in the last two.

What to do. The gap between 64% trying to influence AI answers and 35% tracking them is the cheapest one to close. Write down 20 questions a buyer in your category would ask an assistant, run them monthly in ChatGPT, Gemini and Claude, and record whether you are named, how you are described and who is named instead. And if your team is the same size with shorter deadlines, the finding that workflows, not tools, are the bottleneck says where to spend your next improvement effort. Demandbase sells account-based marketing software, and the release describes the research as interviews while reporting the results as percentages of those surveyed, so treat it as a strong signal from a small sample.

3. Basis: 61% of marketers now use AI every day, and only 6.5% measure its return against a benchmark

Basis, which sells advertising software, published the fourth edition of its AI and the Future of Marketing report on 19 October, based on a survey of 184 marketing and advertising professionals across agencies, brands and publishers, fielded in August 2026. 61.4% use generative or agentic AI in their work every day, up from 33.6% a year ago, and 98.4% use it at all. Basis says 51.1% of organisations now run AI agents, most often for reporting and analytics, drafting campaign briefs and creating content. Coverage of the report adds that 47.8% say their organisation replaced human tasks with AI in the past 12 months, up from 27.9% a year earlier. The measurement finding is the one that matters: only 6.5% measure the return on their AI investment against defined benchmarks, and 65.1% either do not measure it or do not know whether anyone does.

What to do. Put this beside Forrester’s prediction and the conclusion writes itself. If your team cannot show a before-and-after number for the AI tools it uses most, pick the top three and set a baseline this month, such as hours per campaign brief, time to first report or cost per asset, so you have something to defend when budgets are reviewed. The sample is small and the full methodology is in a gated report.

1. Canva launched Craft, its own image model, trained on designs rather than photographs

Canva Craft, released on 22 October, is the first image model built and trained from scratch by Canva Research, and Canva says it is “available starting today to all Canva users globally” through Dream Lab, with access for developers through Canva’s public API. It works in two stages, planning placement, scale, hierarchy and text first, then generating an image at up to 2K resolution in more than 30 aspect ratios or a custom one. It writes text inside images in 55 languages. It was trained on millions of human-designed Canva templates, including posters, social posts, presentations and infographics, and Canva’s Magic Layers tool can split a Craft image into editable design elements. Canva says it debuted fourth on the DesignArena leaderboard, its own reported ranking. Access through the Canva MCP server, so assistants can call Craft directly, is labelled coming soon, as are two cheaper versions, Craft-fast and Craft-ultrafast. For marketers, the test is text-heavy assets such as event graphics and social cards, where general image models still misspell words.

2. Gemini now takes your brand kit, connects to Google Ads and Canva, and keeps your skills in one place

On 20 October Google introduced a business landing page in the Gemini app for small business owners. You can define your brand style by uploading your website, brand kit, fonts or colour palettes, and Gemini then generates visuals for launches and campaigns with Nano Banana 2.1, Google’s latest image model. It adds marketing and finance skills and makes it easier to connect Google Business Profile, Google Ads and Google Workspace, plus Adobe, Canva, Dropbox, Shopify, Wix, Zoho and Xero. Google says it begins rolling out “in the coming days globally for all Gemini app users”. Alongside it, a new Customize menu in the Gemini app, posted on 21 October, puts your skills, prompt preferences and more than 40 connected apps in one place, for every user tier on web, mobile and desktop. That rollout started on 13 October and can take up to 15 days to appear, and most connected apps are currently available only in English in the US. The launch is framed for small businesses, but a brand kit that sits inside the assistant your team already uses is the part worth testing for any marketing team.

3. Salesforce Marketing Cloud can now be run from ChatGPT, Gemini or Claude

Part of the Winter ‘27 release, Headless Marketing lets you run campaigns, loyalty programmes, personalisation and data pipelines across Marketing Cloud and Data 360 from the agent or interface you already use. It works through the Headless 360 MCP server, which connects Marketing Cloud Next to AI assistants including ChatGPT, Gemini, Claude and Slack, and Salesforce says it goes generally available in October. The same release brings flash sending to five million app users in 90 seconds and a sender rating score with alerts and recommendations, also GA in October. If your team already works in an assistant all day, this is the piece that lets a campaign change happen there rather than in a separate login, so ask your admin whether you are on Marketing Cloud Next and what permissions an assistant would get before anyone connects one.

4. Google’s first Marketing Platform Drop puts Gemini into Search Ads 360 and LinkedIn’s TV ads into DV360

Google launched a monthly update series for its Marketing Platform on 22 October. The first edition includes a new Search Ads 360 Home, a customisable view that uses Gemini models to consolidate insights, generate suggestions and streamline execution, and Gemini-curated media bundles in the Display & Video 360 Marketplace with one-click activation for live sports and tentpole events. It also brings LinkedIn’s connected TV ad inventory into DV360 for unified buying, planning and measurement. LinkedIn’s own announcement says buyers can apply its job title, industry and seniority targeting to streaming TV, for US audiences only, switched on through a LinkedIn or Google account representative rather than self-serve. LinkedIn claims 2.6 times greater B2B awareness effectiveness than other connected TV, without a published method. Google’s post carries no beta or availability labels, so check your own account before planning around the Search Ads 360 view.

5. Gemini Notebook now works inside Gemini chats, and its slide decks are easier to steer

From 23 October, Gemini Notebook is available as a connected app in the Gemini app, so you can create, edit and query your notebooks from inside a chat in plain language rather than switching tools. Google says it is on by default except in the European Economic Area. The same week, Gemini Notebook’s slide decks and video overviews gained a step-by-step flow for steering their content and style before they are generated. If your team keeps research, call notes or competitor files in notebooks, this makes them usable in the chat where you are already drafting.

1. Google is writing AI review summaries into Search ads, positive reviews only.

A new Google Ads help page, spotted on 24 October, describes AI-generated review summaries that combine positive online reviews into a snippet in Search ads, in the US, UK, Australia and India, across all positions, devices and most verticals, with sensitive categories such as pharmaceuticals excluded. Google’s AI decides when to show them and advertisers cannot customise the text. Search your own brand and top category terms this week, note what the summary says, and fix the review sources it draws on if it is wrong.

2. Ahrefs says Google is feeding SEO tools altered results.

Ahrefs reported that Google serves altered results to suspected scrapers on between 5% and 20% of requests, and that it now filters abnormal results, requests replacements and cleans its historical data. It published no method. Other vendors report similar effects at different rates: DemandSphere put wrong daily rank checks at about 11% on 13 October, Nozzle saw a peak of 27% on 29 September and AccuRanker found under 1% in mid-October. If a ranking swing this autumn looks odd, check it against Search Console before you act on it.

3. The skills format behind Claude, ChatGPT and Copilot now belongs to an independent foundation.

On 22 October Anthropic donated the Agent Skills standard, the SKILL.md format for packaging reusable instructions, to the Linux Foundation’s Agentic AI Foundation, including its name, logo, GitHub organisation and agentskills.io. Maintainers include people from Anthropic, Google and OpenAI, and the roadmap adds versioned releases, a formal proposal process and security guidance. Anthropic says more than 70 platforms support it, including ChatGPT, Cursor, GitHub Copilot and VS Code, with a million skills published and nearly 280 million installs. Keep your team’s skills as plain folders with a SKILL.md file, so they move with you when you change assistant.

1. The AI is writing more of what carries your name, and the risk is staying with you. Meta’s terms put advertisers on the hook for edits its AI makes on its own. Google is writing review summaries into Search ads that advertisers cannot change. Canva’s new model writes the text inside your images, and Creatify’s ten ads were each checked by a person for pronunciation, lip sync and claims before launch. Demandbase found only 21% of B2B marketers rarely have to correct AI output. The common thread is that the platforms are automating the creative and keeping the liability with the brand. The answer is not to switch everything off. It is to know what is switched on, check what is actually serving, and keep a record of what you approved.

2. Measure your own numbers before you fund or cut anything. Three studies said AI search sends almost nothing, and the rebuttal said the counting is broken, a week after a bug proved it. Forrester expects a third of AI initiatives to be cut as usage-based bills arrive, Basis found only 6.5% of marketers measure AI’s return against a benchmark, and Ahrefs says even the rank-tracking data many teams rely on has been partly altered. Industry studies will keep disagreeing. The teams that come through 2027 planning well will be the ones who can show their own before-and-after numbers, with the gaps annotated, for the tools and channels they want to keep.

3. Your connectors are becoming your stack. Claude Duet is only as useful as the apps on your Claude account, and Brock Mesarich had to add a Zapier server before it could reach an email platform. Eliot Prince’s whole system starts by connecting Search Console, Analytics and an SEO data source, because without them “AI just ends up guessing”. Salesforce now lets you run Marketing Cloud from an assistant, Gemini connects to Google Ads and Canva, and the skills format those assistants share now belongs to an independent foundation. The question for next year’s martech review is less which platform to log in to and more which tools your assistant can reach, with what permissions, and who decides.

The problem wasn’t Claude. It was my folders.

In January I started handing Claude whole marketing jobs instead of single prompts. Plenty came back needing redoing. When I dug into why, my files were a mess and Claude was guessing.

So I wrote it a welcome guide, the way a good Airbnb host does. Now work that used to take two or three hours comes back in about twenty minutes, in my voice.

My Claude Operator System masterclass shows you how to set up the same system. No code, and nothing goes out without your yes.

(Now only $150 or 2 payments of $75. Individual price $953).

Show me how it works

Tell me: What’s your biggest takeaway from this week’s news from the world of AI Marketing? Love to hear your thoughts in the comments.

Leave a comment

Thanks for reading the Marketing with AI Weekly Roundup! If you enjoyed this read, the best compliment I could receive would be if you shared it.

The Marketer's Guide to Claude Artifacts

Wed, 21 Oct 2026 12:04:07 GMT

Every Wednesday, one essay on what's actually working in Marketing with AI - tools, workflows, and mindset shifts you can apply immediately.

Subscribe now

The Marketer’s Guide to Claude Artifacts

(NB: Half way down is a quick video overview of Claude Artifacts to acompany this video)

You finish a really good session with Claude. The research is solid, the positioning is sharp, and you’ve argued through the messaging until you’re genuinely happy with it.

Then the real work starts.

You copy the summary into a Word doc for the brief. You open PowerPoint and rebuild the key points as slides. You jump into Canva for the one-pager, re-typing the headline and hunting for the right hex codes. You export a spreadsheet, build a chart, screenshot it and paste it into the deck. By the time it’s all done, the five assets say slightly different things, and you’ve spent longer moving the work around than you spent thinking about it.

For most marketers, that’s still how AI works. Claude does the thinking, and then you do the assembly.

That changed quite a lot over the last few weeks, and I suspect most people haven’t noticed.

I opened the Artifacts page in Claude recently and it looked completely different from the last time I’d paid it any attention. Across the top there’s now a row of tiles under the heading “Make something new”: Docs, Slides, Design and Dashboard. On Team and Enterprise plans there’s a fifth, Motion.

Claude Design, which I wrote a full guide on back in May, is no longer a separate product. Design now lives inside normal Claude conversations, and the standalone site at claude.ai/design closes on 14 December. Docs and Slides have arrived alongside it, and on 8 October Anthropic added live dashboards and animated video on top.

The document, the deck, the design and the dashboard can now all be made and finished inside Claude, from the same conversation, using the same context and the same brand.

So this week I’ll walk you through what an artifact actually is now, which plan gets you what, how to set up your design system inside Claude (or move it across from Claude Design before the deadline), each template with a marketing use case and a prompt you can adapt, and how to chain them together so one campaign brief turns into a full set of assets.

What an artifact actually is now

If you’ve used Claude for a while, you’ll remember artifacts as the panel that popped open on the right-hand side of the chat. You asked for a bit of code, a small web page or a one-pager, it appeared in the panel, and you copied it out and pasted it somewhere else.

“Artifact” is now the single container for anything substantial Claude makes for you. A document, a deck, a landing page, a dashboard, an animation. They all open beside the chat, they can all be edited directly, and they’re all saved in one place, the Artifacts tab.

There are now five ready-made templates:

Docs, for writing you’ll keep and work on together.
Slides, for presenting.
Design, for anything that needs to look right. One-pagers, landing pages, social graphics, mockups.
Dashboards (beta), for numbers that stay current.
Motion (beta), for short animated explainers.

It’s moved quickly, so here are the dates in one place:

April: Claude Design launches as a standalone product at claude.ai/design.
16 September: Docs and Slides arrive, and Claude Design moves into regular Claude conversations. The same day, Cowork and chat come together as one Claude.
8 October: Dashboards and Motion launch in beta. Docs, Slides and Design come out of beta and become available on every plan, including Free.
14 December: the standalone claude.ai/design closes and redirects into Claude.

By October, people had already made more than 45 million docs, decks and designs in Claude. Anything you made in the old style before 16 September is now called a “legacy artifact”. Those still work and you can still share them, you just can’t create new ones that way.

Why this matters more for marketers than most

Anthropic’s September recap video put it well: you do the thinking once, then ask for a doc, a deck or a design, and “There’s no more need to re-brief.”

Re-briefing costs more in marketing than most jobs. One campaign might need a brief, a leadership deck, a sales one-pager, a landing page, social assets and a results dashboard, and each used to mean a new tool and another chance for the details to drift.

When every asset is built from the same conversation, the same files and the same design system, consistency stops being something you have to police. That’s why I think Artifacts deserves a fresh look, even if you dismissed it a year ago.

Which plan gets you what

This is the bit that trips people up, so it’s worth being clear before you go looking for a button that isn’t there.

Motion is currently Team and Enterprise only, and that includes Max. Paying for a bigger personal plan doesn’t unlock it. I’m on Max myself, so I don’t have the Motion tile either.

Three other things worth knowing before you start:

Everything needs one setting switched on. If none of this shows up for you, check “Cloud code execution and file creation”. On Pro and Max it’s under Settings, then Capabilities. On Team and Enterprise, an owner turns it on in Organization settings.
It all comes out of your normal usage. Claude Design used to have its own separate weekly allowance, which was a real constraint. That’s gone. Everything now shares the same pool as the rest of Claude, including Claude Code, and big design projects and long documents use more of it.
On Team and Enterprise, your admin controls a lot. Each template can be switched on or off for the organisation. If a colleague can see something you can’t, that’s usually why.

Three ways in (plus a shortcut)

You don’t need to learn a new tool to get started. There are three ways in, and they all end up in the same place.

Just ask. In any conversation, say “Turn this into a Claude Doc” or “Make me a 10-slide deck from this report.” You can even describe what you want and let Claude choose the format.
Use the Output button in the message box and choose the template you want.
Start from the Artifacts tab and click one of the tiles. This is the best starting point for Dashboards, because it opens with the right setup.

The shortcut: if you already know what you want, type it as a slash command. /docs, /slides and /design skip Claude working out the format and go straight there.

It all works on the web, desktop and mobile apps, although sharing settings need the web or desktop.

One tip on models, straight from Anthropic’s own team. Use Opus 5.5 while you’re still working out what you want, then switch to Sonnet 5.5 once you know. Sonnet is faster, which suits the quick rounds of changes that make up most of the work on a deck or design: tightening a slide, swapping a chart, trying a different colour palette. It also makes your usage go further.

The step you still can’t skip: your design system

In May I wrote that the single biggest mistake people make with Claude Design is skipping the design system. My own first attempt produced generic typography, off-brand colours and a look that felt like someone else’s product. Once I set the design system up properly, almost everything came out on-brand first time.

That’s even more true now, because the same design system now powers both your designs and your decks. Set it up once and your landing page, one-pager, social graphic and monthly performance deck all come out in the same colours, type and fonts. Designs go a step further and can use your actual components too, like your buttons and cards. When you start a deck or a design, you’ll see a “Choose design system” option, and once you’ve set a default, new work picks it up automatically.

Anthropic’s own designer, in a short video about how he uses Slides, put it simply: “It knows what our brand looks like, so I don’t have to find the latest fonts or slide templates.”

One note on scope. Design systems apply to decks and designs. Docs and Dashboards don’t use them yet. You can still attach your brand guide and colours to those and ask Claude to follow them, it just means you’re passing the brand in yourself.

Saved design systems are in beta on Pro and above. They’re not on the Free plan.

How to set it up inside Claude, step by step

These steps are for the version inside Claude. If you built one in the standalone Claude Design, migrate it instead (next section).

Step 1: Gather your inputs first. The principle hasn’t changed. Real examples beat specs.

Brand assets. Logo files (SVG ideal, PNG fine), your brand fonts (TTF or OTF), and your colour palette as hex codes.
Real, finished examples of your brand in action. A screenshot of your website, an existing deck, a landing page, a brochure. Anthropic is very direct on this: “Include real examples, such as a finished landing page, not just specs.”
A few lines describing your brand. What it stands for, the feel you’re going for, and any visual rules you live by (”rounded corners”, “no drop shadows”, “buttons are pill-shaped”).
Optional: a Figma file, your website URL, or a GitHub repo with your front-end code.

You don’t need all of it. I built my own from a single product box image and a few paragraphs of commentary.

A common watchout: fonts. Anthropic’s own guidance is that Claude can’t use a font that’s named in your design system but has no file. Upload the font files themselves, otherwise Claude substitutes the nearest match and your output looks subtly off-brand without you spotting why.

Step 2: Start a new design system. From the message box in any conversation, open the + menu, choose Design system, then New design system. Or go to Settings, then Design systems, and click New design system.

You’ll get two choices. Using Claude opens an empty design system beside a new conversation. Using Claude Code is for a design system that already exists in a code repository. For most marketers, Using Claude is the right one.

Step 3: Pick how to fill it in. You’ll see three cards:

Upload brand files is where most marketing teams will start. Logos, fonts, a guidelines PDF and a few finished slides or pages.
Build from connectors reads your brand from a connected tool, like your drive or wiki.
Sync from GitHub is for brands already defined in code, usually one for your web team.

If you’d rather steer it with a prompt in the conversation beside the page, this works well:

“Build our design system from the attached assets. Identify our colours, typography, spacing, components and layout patterns. Separate what you can directly see in the examples from choices you’re proposing to fill gaps. Ask me about any conflicts between the references before you finalise it.”

No brand assets yet? One creator built a design system from nothing more than a handful of Pinterest screenshots and the prompt “Make a new design system based on the vibe of these reference images.”

Step 4: Review the first draft. Anthropic says the first draft is deliberately small, and Claude then asks whether that’s enough or whether to take it further. It’s designed to ask before filling in a value your files don’t show.

Read the overview and each section. You can edit the overview, colour values, notes, fonts and assets directly on the page, and ask Claude in the conversation beside it for bigger changes. Be specific. “This doesn’t look right” gives Claude nothing to work with. “The accent colour should be #F39C12” or “the button padding is too tight” does.

Step 5: Leave usage notes for Claude. This is the step almost everyone misses. A usage note tells Claude how to use a colour, a type style or a component, and when not to. Claude reads a component’s note before it builds with it.

If you ask Claude to fix something in one design, the fix only applies to that design. If you write it as a usage note, it applies to everything Claude makes from then on. “Never use the accent colour on body text.” “Pill buttons are for primary CTAs only.” Write the rule once and it sticks.

Keep the overview short, because Claude reads it every time it uses your design system. Rules about a single colour or component belong in that item’s own note.

Step 6: Test it with an ordinary request. Anthropic’s own test is a good one. Ask for an everyday deck and say nothing about your brand:

“Make me a five-slide deck introducing our company to a new partner.”

If the deck looks like your company, the design system is working. If it doesn’t, add more examples of your brand in use, then tell Claude exactly what to change.

Step 7: Share it and set it as the default. At first, only you can use it. Share it with specific people using Share at the top of its page. On Team and Enterprise, an owner can share it with everyone in the organisation and turn on Default design system, or choose Set as default in Settings, so it’s already selected whenever anyone starts a deck or a design. There’s one default per organisation, and setting a new one replaces the old.

One more thing worth knowing: decks and designs you’ve already made keep their own copy of the design system from when they were made. If you update your brand, ask Claude to update the older pieces too.

Be patient with the first build. It’s one of the heaviest things you can ask Claude to do. One designer testing it found a full Figma import took around 45 minutes and used a big chunk of his usage for the day. He also found importing from code gave a closer result. Gather everything first and try to get it right in one session.

If you use Claude Design today: do this before 14 December

If you built a design system in the standalone Claude Design, you don’t need to start again. Some things move across and some don’t, though. Anthropic’s reason for closing the standalone site is refreshingly plain: running two versions slows down fixes for both.

Migrate your design systems. Open the Artifacts page in Claude and click “Migrate team design systems” in the banner. One migration brings over every design system in your organisation and keeps the owners and sharing settings. The originals stay untouched in the standalone version until it closes.
Finish the migration. Each migrated design system has one step left. Its owner opens it, selects “Finish the migration” and sends the request to Claude, which reorganises the files so you can preview and edit each section on the page. If it still needs a tidy-up, “Let Claude clean it up” sorts the guide, tokens and components.
Check for anything that didn’t move. Empty design systems and ones based on a built-in starter theme won’t migrate. Any that fail are listed with a reason, and there’s a “Fix in Claude Design Standalone” button to sort them before you migrate again.
Download any project you want to keep. Projects don’t migrate yet. Open each one in the standalone version, click Share, choose Project HTML, keep “Project archive” selected and click Export. You get a free .zip that doesn’t count towards your usage.
Save anything useful from your chats and comments. These aren’t carried over and will be deleted when the standalone version closes.
Replace any public links. Public links to standalone projects stop working on 14 December. If you’ve shared a landing page or brochure that way, republish it from Claude and update the link wherever it lives.

One warning for teams. Published design systems that aren’t private become visible to everyone in your organisation once they’re migrated, so check what’s published before you start. Team and Enterprise admins should also check Organization settings, then Artifacts first. On Enterprise, design systems were off by default, and the migration won’t run until they’re on.

It’s a quick job if you do it now, and a frustrating one if you leave it until the week before Christmas.

(I’ve recorded a quick video overview of Claude Artifacts to accompany this post.”)

The templates, one by one

With the foundation in place, here’s each template in turn. What it is, how it works, a marketing use case with a prompt you can adapt, and the watchouts.

1. Docs: for writing you’ll keep

What it is: a proper document that lives inside Claude, with headings, tables and multiple tabs. You and your team can write in it at the same time as Claude.

The difference from the old way is that you don’t re-prompt for every small fix. You type straight into the doc and it saves as you go. When you want Claude involved, you have three options:

Highlight the text and choose “Edit with Claude”, then describe the change. It edits that section and leaves the rest alone.
@mention Claude in a comment. Claude replies in the thread, makes the change and explains what it did. Your colleagues can do the same.
Ask in the chat beside the doc for bigger changes.

Every change is attributed to whoever made it, and Anthropic’s line on this is a good one: “Nothing is locked while Claude works, and you always have the final say.”

You can ask Claude to add a chart, diagram or timeline, and it’ll pull the numbers from a connected app and build it in the doc. The catch is that charts in Docs don’t update by themselves, even when the source is Salesforce or Google Sheets. Ask Claude to pull the latest data when you need a refresh.

Export: Word, PDF, Markdown or Google Docs. You can also ask Claude to turn a doc into a Slides deck, which I’ll come back to.

There’s a short video from one of Anthropic’s engineers that shows the right way to use Docs. He writes his launch reviews himself because, in his words, “I want it to reflect what I actually think,” but asks Claude to dig up the numbers and put the chart in the doc. Then he asks Claude “is this right?” and it either confirms or corrects him. That’s 10/80/10 in practice. The thinking and judgement are his, and the legwork and checking are Claude’s.

Marketing use case: the campaign brief that stays current.

“Create a Claude Doc for our Q1 product launch brief. Use the attached positioning notes and last quarter’s results. Sections: objective, audience, key message, offer and exclusions, channel plan, budget split, timeline, success metrics and open questions. Use only the facts in the attached files and put anything unanswered in the open questions section rather than guessing. Leave budget numbers as placeholders for me to fill in. Add a second tab for the agency brief. Ask me about any gaps before you start.”

Why it matters: briefs are usually written once, emailed round, and then quietly go out of date while everyone works from their own copy. A shared doc the team comments on, with Claude handling the rewrites, stays the single version of the truth. And putting unanswered points in their own section stops Claude filling gaps with invented facts.

Watch out: on Team and Enterprise, docs can’t be shared outside your organisation yet (on Pro and Max you can share by link). There’s no comment-only access, people are either viewers or editors. And only the owner can delete a doc, with no bin, so it’s gone for everyone.

2. Slides: for presenting

What it is: a deck built from your notes, a report, a call recording, a doc, or the conversation you’ve just had. You can retype a headline or move a box yourself, add a table, present directly from Claude, or ask for changes in chat.

Start from the Artifacts tab and you’ll see the “Choose design system” button, so pick yours first. For changes to one element, use the Comment button near Present and click what you want changed. For bigger changes, refer to the slide by number in the chat: “split slide 4 into two so it doesn’t feel so crowded.”

Export: PowerPoint, PDF, or an editable Google Slides file. Anthropic says PowerPoint and PDF exports now keep the formatting from the editor, which was a weak spot in the early days of Claude Design. I’d still open the exported file and check it before it goes anywhere important.

My favourite Slides example comes from Anthropic’s own team. One of their designers used to spend ages staring at a blank slide app. Now he goes for a walk and talks his idea through on his phone’s voice memo app. “It’s usually a mess, but it’s okay because it’s just for me and Claude.” When he gets back, he pastes the transcript in with “Turn my exact words into a presentation using Claude Slides,” makes a couple of small tweaks himself and leaves comments for anything bigger.

It’s a great habit to steal. The hard part of any presentation is working out what you think, and this lets you do it away from a screen.

Marketing use case: the monthly performance deck.

“Make a 10-slide monthly marketing performance deck from the attached GA4 export and campaign spreadsheet, using our design system. Slide 1 is a one-line summary of the month. Then traffic, leads, cost per lead and pipeline, each compared with last month and each on its own slide with one chart. One slide on what worked, one on what didn’t, and finish with three recommendations for next month. Keep every number identical to the source files and flag anything you had to calculate yourself.”

Why it matters: the monthly deck is the classic job that takes half a day and adds very little thinking. Asking Claude to flag calculated numbers gives you a short list to check, rather than having to check everything.

One trick worth stealing: if a slide isn’t working, ask “Give me 3 layouts for slide 4” rather than trying to describe the layout you want.

3. Design: for anything that needs to look right

What it is: this is Claude Design, now living inside normal Claude. A chat on one side and a canvas on the other, for one-pagers, landing pages, social graphics, brochures, email mockups and prototypes.

You refine it with chat for bigger structural changes, inline comments on a single element, or direct edits on the canvas when you just want to nudge something. Creator demos also show a Wireframes and Final designs switcher at the top. Asking for wireframes first and agreeing the layout before Claude designs anything saves a lot of back and forth.

The biggest practical change from the standalone version is that Design now has access to everything else Claude has. Your files, projects, connectors and the conversation you’ve just had. You can say “build the landing page from the brief we just finished” and it knows what you mean.

Export: .zip, PDF, PowerPoint, Google Slides or standalone HTML. You can also send designs to Canva, Gamma, Lovable, Miro, Netlify, Vercel and others, or use “Handoff to Claude Code” if a developer is building the real thing.

Getting better output

Anthropic’s own tips: core layout before detail, name components in feedback, say which devices it’s for early, ask for two or three variations, and ask for an accessibility review before you finish. Creators testing it add three more:

References beat longer prompts. Show Claude a section, poster or layout you like and say what you like about it.
Tell it what you don’t want. The defaults can drift towards the patterns every AI design tool uses. One designer simply tells Claude “no cards” for launch pages.
Ask for the details it won’t add by itself. Hover effects, buttons that press, numbers that count up. Claude can do all of them, but often won’t unless you ask.

Marketing use case: the campaign one-pager and social set.

“Using our design system and the approved brief we’ve just written, create a one-page sales sheet for [product] aimed at [audience]. Headline, three benefit blocks with icons, one customer quote, a pricing summary and a single clear CTA. Keep every claim and number identical to the brief. Show me wireframes first. Once I’ve approved the layout, create a matching LinkedIn banner and a 1080x1080 social card using the same headline.”

Why it matters: that’s a campaign’s worth of visual collateral from one brief, all on-brand and all saying the same thing. The wireframe step means you agree the structure before Claude spends effort on polish.

If you read my May guide, all five use cases in it (ABM microsites, sales collateral, comparison pages, pricing tests and email mockups) still work exactly as written. They just live in the Design template now, with your whole conversation available to them. One addition I like for landing pages: use your best-performing ad’s copy as the page headline. If the ad is what got the click, the page should carry the same promise.

Know what Design can’t do

It doesn’t generate photographs or illustrations itself. It designs with your images, shapes, icons and type. For product or lifestyle imagery, supply your own or connect an image generation tool.

A design isn’t a live website. The HTML export is a working page, but a page with checkout, CRM forms, analytics and a proper domain is still a production job.

4. Dashboards: for numbers that stay current (beta, Pro and above)

What it is: this is the newest addition, and for marketers it might turn out to be the most useful. Most artifacts show what you gave Claude at the time you made them. A dashboard connects to your data and keeps itself up to date.

You ask a question in plain English. Claude writes a query for each chart, runs it against your data and builds the dashboard. It connects to data warehouses like BigQuery, Snowflake, Databricks, Amazon Redshift and ClickHouse, or to apps you’ve already connected to Claude, like Salesforce.

What I like most is the transparency. Click any number and you can see exactly how it was calculated, including the query behind it. Every chart shows when its data was last refreshed. It’s a bit like clicking a cell in Excel and seeing the formula, and it means you can check Claude’s working before that number ends up in front of your boss.

Anthropic is clear about what it’s for: quick, exploratory questions, not replacing your BI tool. When you need to go deeper, you can send a dashboard on to tools like Amplitude, Mixpanel, PostHog, Hex and Sigma, with Looker, Tableau and monday.com listed as coming soon.

The official examples lean on data warehouses, which most marketing teams don’t touch day to day. In practice, creators testing it have mostly built theirs from everyday connected apps like Gmail, Stripe, HubSpot, Ahrefs and their newsletter platform. No connector at all? Export a spreadsheet as a CSV and drop it in. Just be aware that a file is a snapshot, so that dashboard won’t update by itself.

Ask a question, not for a chart

The tip that makes the biggest difference comes straight from Anthropic: “Say what you want to know, not just what to chart.”

Weak: “A leads chart.”
Strong: “How do this month’s leads by channel compare with last month’s, and which channel’s cost per lead got worse?”

Some of the best dashboard advice comes from a creator who builds one-page business dashboards:

Decide what to leave out. If you wouldn’t do something differently because of a metric, it doesn’t go on the dashboard.
Keep it to 15 metrics or fewer.
Ask Claude to check it can actually get every number from your connected sources before it builds anything.
Wireframe it first so you agree the structure before Claude builds the real thing.

Marketing use case: the budget decision dashboard.

“Make a dashboard from [my CRM / this CSV]. I need to decide where to move budget next month. Put the three numbers that matter most at the top, each compared with last month: leads, cost per lead and pipeline value. Then break them down by channel and by campaign. A qualified lead means [your definition]. Add a filter for date range and region. Reconcile the channel totals against the overall total before you show me anything, and tell me every assumption you made. Ask me questions first if anything’s unclear.”

Why it matters: the decision comes first, so Claude builds something that answers it rather than a wall of charts. Defining “qualified lead” yourself matters more than anything else in that prompt. Seeing the query tells you how a number was counted. It doesn’t tell you whether it was counted the way your sales team counts it.

Watch out for

Viewers need their own access to the data. The person you share a dashboard with needs their own connection to the source to see the numbers, otherwise that section shows an error. Sensible, but it means you can’t just send a live dashboard to someone outside the business.
It can change your data, not just read it. An analyst testing it on BigQuery found Claude created a new cleaned-up table in his warehouse to build the dashboard. He’d sensibly connected an account that could only see the one dataset. Do the same, and give Claude access to only what it needs.
Layout changes can’t be rolled back. The queries behind each chart have a history you can restore, but the layout doesn’t. Duplicate a dashboard you like before asking for big changes.
It’s not a BI replacement. That same analyst gave it 8 out of 10 for a beta, but was clear it’s for quick exploration, not a replacement for Power BI or Tableau.

5. Motion: animated explainers (beta, Team and Enterprise)

Motion turns a prompt into a short animated video: a 30-second explainer for an all-hands, an animated chart for a board deck, a product walkthrough or a launch clip. As I said earlier, it’s Team and Enterprise only, so I don’t have it on my Max plan.

What makes it different is how it’s made. Motion doesn’t use a video generation model. Claude writes code that animates your text, charts, shapes and images, so there’s no generated footage and no AI-generated people. The upside is that it’s easy to edit. Change one word, number or scene length and the rest of the video stays exactly as it was. You export it as an MP4.

Anthropic’s one piece of prompting advice is the brief any good video producer would ask you for: “Say who will watch it, where it will play, and roughly how long it should be.”

Marketing use case: the 20-second result for LinkedIn.

“/motion Turn the attached campaign results into a 20-second explainer. Who will watch it: marketing leaders scrolling LinkedIn on their phone. Where it will play: in the feed, muted, 4:5. The one thing they must remember: [one number or one sentence]. Use only the figures in the attached file and our brand colours. One idea per scene, no more than eight words on screen at a time. End on: [your CTA]. Before you animate anything, give me the storyboard scene by scene.”

Why it matters: asking for the storyboard first means you fix the story in text, which is quick, before Claude animates it, which isn’t. And limiting it to your attached figures keeps Motion explaining your results rather than inventing a better-sounding one.

If you’re on Team or Enterprise, Anthropic’s Get started with Claude Motion guide and launch article are the places to start, and Ruben Hassid’s walkthrough is a very good practical guide.

On Pro or Max, you can still make animations, you just don’t get the Motion editor. Claude can build animations in an ordinary conversation or a Design artifact. The 60-second explainer in my Goals and Loops issue back in August was made in Claude Design, well before Motion existed. What you miss without Motion is the editor built around it, which lets you tweak and comment on individual moments and export straight to MP4.

One campaign, every artifact

Each template is useful on its own, but the real step forward is running one campaign through all of them, from the same conversation, files and design system.

Here’s how I’d use it for a product launch:

Doc: the approved campaign brief. Offer, audience, dates, pricing, exclusions, proof points and CTA. This is your single source of truth, and it’s where you and the team argue things out in comments.
Slides: “Turn this brief into a 6-slide sales briefing using our design system. Keep every claim and number identical to the brief.”
Design: the one-pager, landing page and social graphics, built from the same brief with the same design system.
Dashboard: once the campaign is live, a dashboard tracking the numbers that would change a decision.
Motion, if you have it: at the end of the month, a 20-second animation of the one result the business needs to know about.

The document is where the thinking happens, and everything else is built from it.

One caveat, though. Changing the brief doesn’t automatically update the deck or the one-pager. They’re built from the same source, but they’re separate artifacts once they exist. If the price changes, ask Claude to update each asset, then check: “Compare the deck and one-pager against the brief and list anything that doesn’t match.”

That chain used to mean days of work across a writer, a designer and an analyst. Now it’s a handful of prompts, plus the time you spend checking the work, which is still your job.

The honest caveats

A few things you should know before you build your week around this.

Two of the five are still in beta. Dashboards and Motion will change, and some of their integrations are still “coming soon.” I wouldn’t build a client-facing process on either just yet.

It all uses your normal limits. Design systems, long documents and big designs draw from the same pool as your everyday chats. Plan your sessions and use Sonnet for the quick rounds of edits.

Sharing depends on your plan. Everything starts private. On Pro and Max you can share with anyone who has the link. On Team and Enterprise, sharing stays inside your organisation unless an owner switches on external sharing. Anyone opening a shared artifact needs a Claude account, and if it uses Claude, they use their own plan rather than yours.

Duplicate before big changes. Design has no version history yet, and dashboard layouts can’t be rolled back.

Organisation is still basic. Artifacts pile up quickly. Pin the ones you come back to, so you’re not scrolling to find last month’s dashboard.

Keep a copy of anything that matters. Exports are two clicks, and it means your important work always lives somewhere you control.

You still need taste and judgement. Claude is fast and, with a design system, on-brand. It doesn’t know whether the claim on slide four will get you in trouble with legal, or whether the number on the dashboard is counting what your sales director thinks it’s counting.

The bigger picture

A year ago, Claude was something you talked to, and then you took what it gave you somewhere else to finish the job. Into Word, into PowerPoint, into Canva, into a spreadsheet. Every handover was a chance for the work to lose detail and drift off-brand.

Now the document, the deck, the design and the dashboard are finished inside Claude, using your brand, your files and your data, and exported only when they need to go somewhere else.

It fits neatly into the way I teach 10/80/10. You do the first 10%, the thinking, the brief and the decisions about what matters. Claude does the middle 80%, drafting the doc, building the deck, designing the one-pager and pulling the numbers. You do the final 10%, checking it’s right and deciding what goes out. Artifacts make that middle 80% far bigger than it’s ever been, because it now covers the assembly as well as the drafting.

The marketers who get the most out of this won’t be the ones who try every template in a week. They’ll be the ones who set up their design system properly, keep one approved brief as the source of truth, and build a repeatable chain from that brief to every asset the campaign needs.

Setting up the foundation takes an afternoon. After that, everything downstream gets easier.

This week’s challenge: open the Artifacts page and start with the foundation. If you’re on Pro or above, build your design system first, and if you’ve got anything in the standalone Claude Design, migrate it now rather than in December. Then pick the one template closest to a job you already do every week. If it’s reporting, export a spreadsheet you look at every Monday as a CSV, drop it into a Dashboard and ask the question you actually need answered. If it’s content, take your next campaign brief, make it a Doc, then ask for a deck from it and see how much of the rebuilding you skip. Reply and let me know what you built, I’d love to see it.

Thanks for reading Marketing with AI. If you found this useful, sharing it with one other marketer is the best compliment I could receive.

Need 1:1 Help to Embed Claude into Your Marketing?

I work with organisations to automate tasks, streamline workflows, and get more done with AI - building the skills, systems and workflows that turn Claude into their marketing operating system. Consulting, coaching and training, in person or on demand.

From 0.5 days overviews to 2-day workshops, I can help you automate more marketing work and streamline your workflows to get mroe done with AI. I offer both in-person and online sessions to suit your needs.

Let’s talk →

Marketing with AI Weekly Roundup: ChatGPT Traffic Went Missing, Claude Docs and Slides Go Free, and Google Adds More Links to AI Answers

Robert Gillespie — Sun, 18 Oct 2026 10:03:24 GMT

Welcome to issue #55 of the Marketing with AI Weekly Roundup. Published every Sunday, it shares the top stories, playbooks, data and quick hits from the world of Marketing with AI over the last 7 days.

Subscribe to receive the week's most important AI marketing stories direct to your inbox.

Subscribe now

CLAUDE FOR MARKETERS MASTERCLASS - NEW MODULE ADDED:

THE BEGINNER’S GUIDE TO CLAUDE CODE FOR MARKETERS

For marketers who want to start building, get access to over 2hrs of Claude Code for Marketing training. You also get access to the full Marketing Operator System, the Setup Skill that builds your working environment, and 5-Minute Slides, together for the first time.

Just $150 OR 2 payments of $75

(Individual price $953).

Check Out This Limited Time Offer

For about a week, every visit ChatGPT sent to your website arrived with no label. After GPT-6 rolled out on 7 and 8 October, links in ChatGPT answers stopped carrying the tag that tells your analytics where a click came from, so those visits were filed under Direct. One Adobe analysis of 500 pages saw ChatGPT visits fall by 31,000 a day while Direct rose by 26,000. It turned out to be a bug and the tag came back on 15 October, but if you report AI referral traffic, your October numbers have a hole in them.

Anthropic made Claude Docs, Slides and Design available on every plan including Free, and added live dashboards that connect to your CRM and short animated explainers. Google announced that AI Mode and AI Overviews will show more links, higher up, with a test that puts the one or two official sites for a task at the top of the answer. Semrush’s first UK benchmark found that fewer than half of the brands AI answers mention are also among the sites those answers cite, which is a different problem from the one most visibility scores report.

The Playbook has three things to try. An Anthropic sales rep describes the six scheduled agents that route his inbound leads, draft his emails in two languages and keep Salesforce current. Grace Leung builds eight ChatGPT Work agents for marketing jobs, from a Monday data analyst to a proposal that writes itself from a call transcript. And Blazing Zebra turns one call transcript into a branded deck, a live Gantt chart and a client information pack with Claude working inside Google Slides, Sheets and Docs.

1. ChatGPT stopped labelling its links for a week, and AI traffic showed up as Direct

OpenAI rolled GPT-6 out to Plus, Pro, Business and Enterprise plans on 7 October and to Free and Go on 8 October. From 8 or 9 October, depending on which analyst’s data you read, links in ChatGPT answers opened without the utm_source=chatgpt.com tag that OpenAI added in June, and without passing a referrer. With neither signal, Google Analytics 4 and most other tools file the visit as Direct, the same bucket as someone typing your address into a browser.

Cedric Huesler of Adobe spotted the mechanism and published the numbers. Across roughly 500 pages where ChatGPT supplied at least 5% of visits, comparing 9 to 13 October with the week before, ChatGPT visits fell by about 31,000 a day and Direct visits rose by about 26,000 a day, with Direct climbing on two of every three pages that lost ChatGPT traffic. Mehrab at SEO Mode, who has logged ChatGPT answers daily since May, found that 97% of answer links carried the tag from May to early August, and 0 of 5,185 links did between 9 and 15 October. Graphite reported ChatGPT sessions in GA4 down between 22% and 57% for the sites it watches. Links in the Sources panel and shopping links kept the tag throughout, so only the inline links were affected.

On 15 October Huesler updated his post: “Relieved. It was a bug. ChatGPT is tagging its links with utm_source=chatgpt.com again.” Glenn Gabe still saw some untagged links on the morning of 16 October before confirming all were tagged. OpenAI has not published a statement, and its help centre page for publishers continued to say the tag is included throughout. Whether the referrer has returned is not confirmed, so the tag may now be the only signal you have.

Why this matters. The GA4 channel grouping for AI assistants, which launched in May, depends on exactly these two signals. A week without them means any dashboard that tracks AI referrals, any report that compares October with September, and any argument you are making internally about the value of AI search visibility rests on a number that is wrong for those seven days. The deeper lesson is that AI referral measurement is held together by a convention one vendor can switch off by accident.

What to do. Add an annotation in GA4 covering 8 to 15 October and read Direct and the AI assistant channel together for that week. Do not report an October drop in ChatGPT traffic to anyone until you have done that. Then set up a weekly check that counts landing sessions with utm_source=chatgpt.com and alerts you if it hits zero, because a bug like this will happen again and you want to know in a day rather than a week.

2. Claude Docs, Slides and Design are now free for everyone, and Claude can build live dashboards and animated explainers

On 8 October Anthropic announced that Claude Docs, Slides and Design are out of beta and available on every plan, including Free, and introduced two new betas. Anthropic says more than 45 million documents, decks and designs have been made since the three tools arrived in conversations on 16 September. They gained real-time co-editing, admin-chosen templates, external sharing where an admin allows it, PowerPoint and PDF export that keeps formatting, export to Google Slides as an editable file, and editing on mobile.

Claude Dashboards is in beta on paid plans. You connect Claude to Amazon Redshift, BigQuery, ClickHouse, Databricks or Snowflake, or to a CRM such as Salesforce, ask a question in plain language, and Claude builds a dashboard that keeps itself current as the data changes. Every number shows the query behind it and when it was last refreshed, which is the detail that makes it usable in a meeting. Anthropic’s own example is comparing this week’s signups with last month’s. A dashboard can be handed on to Amplitude, Grafana, Hex, Mixpanel, Omni, Perplexity, PostHog or Sigma, with Looker, Tableau and monday.com listed as coming.

Claude Motion is in beta on Team and Enterprise only. You type /motion and describe what you want, and Claude produces a short animation as editable code from your own text, charts, shapes and images, exportable as MP4. There is no video-generation model behind it and Anthropic says it makes no AI-generated people. It opens in Adobe, Descript, HeyGen, Higgsfield, invideo, Luma AI and Runway, with Canva and Captions to follow. Anthropic’s examples include a 30-second all-hands explainer, an animated chart for a board deck and a pricing comparison for a sales kickoff.

One deadline. The standalone Claude Design site at claude.ai/design is being folded into Claude and stays open only until 14 December. Design systems can be migrated from the Artifacts page, but chats, comments and public links on standalone projects will not carry across. For Enterprise admins, Dashboards and Motion are off by default, while Docs, Slides and Design switched on by default on 15 October.

Why this matters. The document, the deck and the one-pager were the outputs most marketing teams still finished by hand in another tool. All three are now included in the free tier of one of the three main assistants, and the dashboard is the first of these outputs that stays alive after you make it. For a team that pays for a reporting tool mainly to turn CRM data into a weekly chart, that is worth a serious look.

What to do. Connect your CRM and ask for the weekly pipeline-by-source view you normally wait for someone to send. Check the query under each number before you trust it. If anyone on your team has a design project on the standalone site, move the design system before 14 December. And if you run a sales kickoff or a customer webinar in the next quarter and you are on Team or Enterprise, try one 30-second Motion explainer before you brief a video.

3. Google is putting more links, higher, inside AI Mode and AI Overviews

On 15 October Hema Budaraju, Google’s vice president of product management for Search, announced five changes to how AI Mode and AI Overviews send people to the web. Google says it has already made “continued and substantial increases” to the number of inline links in AI answers and is now launching updates that show them higher up in responses, more frequently.

The other four. Google will test a dedicated section at the top of some answers that surfaces the one or two official places on the web to complete a task or get the exact information, with a line of context about why each link is useful. The carousels of timely articles that appear for news-worthy topics are rolling out to more countries and languages, and they highlight sources a user has marked as a Preferred Source, a feature Google says has grown more than sevenfold in use since the start of the year. The related links at the end of some answers get thumbnails and a more clickable design. And this week Google began testing clickable creator and publisher names on video results in AI Mode that link to their Search profiles.

Google published no click-through data alongside the post, and one measurement detail is worth knowing: Search Console gives every link inside an AI Overview the overview’s single position, so a link moving higher within an answer will not change the position you see reported.

Why this matters. For a year the complaint about AI answers has been that they keep the click. This is Google saying, in writing, that it wants more clicks to leave the answer, and it tells you which clicks: the official source for a task, the timely article, the creator people recognise. For a B2B company, the “official” section is the one to watch. If a buyer asks how to do something with your product, Google is testing whether to put your documentation, your pricing page or your login page at the top of the answer, labelled as the place to go.

What to do. Pick ten tasks a buyer might ask an assistant about your product or category, from “how do I integrate X with HubSpot” to “what does X cost”. Make sure each has a single canonical page on your site that is plainly the official answer, with a clear title and no competing near-duplicates. Set up a Search profile for your company and for the people who present your videos. Then ask customers who read you often to add your site as a Preferred Source, because that setting now affects what they see in AI answers as well as in the news carousel.

1. Six scheduled agents that route inbound leads, draft bilingual emails and keep the CRM current

Shahboz Ibragimov is a business development representative on Anthropic’s Japan team. On 16 October he published a detailed account of how he runs his day with six Claude agents, which he calls the Nakama, Japanese for the crew. It is a sales role and it is Anthropic writing about its own product, but the structure transfers directly to inbound marketing operations, and the path he recommends for starting needs no developer tools.

The before. Until earlier this year he spent a couple of hours each morning routing and qualifying leads, then wrote emails in Japanese or English, and CRM updates happened in leftover time and were often a day or more behind. Now, in his words, “my agents prepare most of that before my day starts, and I review it.”

Six agents, one job each. Every agent has a single responsibility, because “a single responsibility keeps each one easy to reason about and easy to track” when something is off.

1. Furiwake, “sorting”. Runs first thing. It takes the daily inbound lead queue and routes each lead to the right rep, balancing lead count, priority and who is in that day, so every lead has an owner before the team starts. It ends each run by posting a report to a dedicated Slack channel listing every lead that came in and where it went.

2. Sabaki, “handling”. Takes the leads assigned to him, qualifies and categorises each against pre-set criteria, takes the defined next step for that category, which is either a reply pointing to self-serve or a customised first email, drafts that email from the shared knowledge base, and updates the lead record in Salesforce.

3. Henji, “reply”. Scans his inbox every hour through the working day and drafts a reply to each customer email in Japanese or English, matching the thread. He reviews, edits and sends.

4. Shoudan, “opportunity”. Pushes recent customer interactions into Salesforce opportunities so the records stay current with no manual entry.

5. Nokori, “what remains”. Runs at the end of the day, reads the meeting transcripts attached to his calendar, summarises each meeting and drafts the follow-ups.

6. The coordinator. Checks a dedicated Slack channel about every 15 minutes for plain-language change requests, updates the other agents’ instructions between runs, and writes product and pricing changes into the shared knowledge base. Only he can post change requests, and every change is saved as a numbered version.

The knowledge base matters more than the prompt. All the agents read and write one shared store. It holds three things: the difference between what each agent drafted and what he actually sent, saved as a note on how he would phrase it; whatever the agents dig up while researching an answer in internal docs and official pages; and product and pricing updates. The effect crosses agents: “I fix a phrasing once in Henji’s drafts, and Sabaki and Nokori stop making the same mistake.” He says it took “a few hundred cheap rounds of drafting, fixing and saving” before drafts sounded like him, and that what fixed the voice was those saved corrections, not a longer prompt.

Start without any developer tools. His first version, which he recommends to anyone, was three pieces per agent inside ordinary Claude: a skill holding the instructions for one job, a scheduled task that runs the skill, and a shared folder on his laptop as the knowledge base every skill writes its learnings back into. “You don’t need Managed Agents to start.” He moved to Claude Managed Agents, Anthropic’s hosted agent service, which is still in beta, only when he had more agents than he could keep track of and the same skill was being edited from too many different chats. On the hosted version the skill became the agent’s system prompt, the scheduled task became a schedule on the deployment, the folder became a shared memory store, and credentials for email and Salesforce sit in a managed vault each agent can use but never see. He describes each agent’s job in plain language and Claude wrote the short setup scripts.

What always comes back to him. A written rulebook defines each agent’s lane. Pricing questions, sensitive accounts and ambiguous requests always route to him. Every draft from Sabaki, Henji or Nokori arrives with a short summary at the top; he reads the summary, confirms, revises and sends. Nothing in the account goes to a customer without him. Follow-ups that need a further step, such as an upgrade path or an escalation to another team, are handed to Claude Tag in Slack, also in beta, which reads the thread, lays out the options, drafts what is needed, sets itself a reminder to check back on a given day and tags him when a step needs sign-off.

Cost control. Rule-following agents, Furiwake and Shoudan, run on Sonnet, the smaller model. Judgement-heavy ones, Sabaki, Nokori and the coordinator, run on Opus, the coordinator because “one bad edit would reach every agent”. Henji needs judgement but runs hourly all day, so it runs on Sonnet and asks Opus for advice only when it hits a decision it cannot make. He raises the effort level, how long the model thinks, only where the reasoning is hard. His three common mistakes: adding to instructions rather than updating them, giving agents more context than the task needs, and running every job on the biggest model at the highest effort.

The result, which is his own estimate. He counts about 60 hours of work a week saved, measured by taking the weekly log of what the agents completed and costing each task at the time it would have taken him by hand. His third lesson is the one most readers will recognise: “Give Claude more room, sooner.” He had kept the knowledge base updates and the instruction edits for himself, and “both worked fine once I let go of them.”

Do this now. Take one inbound job your team does every morning, such as sorting form fills or drafting first replies, and write its rulebook on one page: the categories, the next step for each, and what always goes to a person. Save it as a skill, give it a scheduled run, point it at a shared folder, and for two weeks save every edit you make to its drafts as a note in that folder. That folder is the asset. The agent is replaceable.

2. Eight ChatGPT Work agents for marketing jobs, from a Monday data analyst to a self-serve proposal

Grace Leung published a 23-minute walkthrough on 16 October of ten agents built with ChatGPT Work, OpenAI’s agent mode inside ChatGPT. Two of the ten, a website redesign and a holiday campaign, are consumer examples, so we have taken the eight that fit a B2B team. The video is sponsored by HubSpot, which she says on screen, and it promotes a free five-skill pack and her paid programme.

The setup that makes it an agent. Her framing is that a prompt gives you an answer and the work is still yours, while an agent takes a goal, plans the steps, uses tools and returns the finished thing. Out of the box, “it’s just like a smart new hire on day one”. What changes that is three things: a project so it knows your brand, a skill so it knows how you do the work, and a scheduled task so it runs without being asked. ChatGPT Work runs in the cloud on chatgpt.com or locally in the desktop app. She uses the cloud so agents keep working with her laptop shut, and switches to desktop only when a job needs files in a local folder.

1. Report to story. Set up a project with the brand context in its instructions. Attach the month’s reports and call notes, or point it at a connected drive, and use the Sites plugin plus a custom skill she calls “storyline first”. The rule in the skill is that every slide leads with a hook and one key message rather than a performance summary. The result is an interactive deck where you can hover any number for the detail, switch between a leadership version and a team version with one click, and edit headlines and body copy directly before downloading, so the team never has to go back to the AI for changes.

2. Audience research in your buyers’ own words. For a small-business accounting product, she asked for what owners struggle with in their current tools, in their exact wording, and told the agent to run a sub-agent for each source type: Reddit, communities and software review sites. Each sub-agent scrapes one source at the same time, which she says goes faster and deeper than one agent skimming everything. After more than 35 minutes it returned a 30-plus-page report of findings, suggested messaging directions and a proposed repositioning of the landing page, with every quote linked to its thread so she could verify it. Save the whole process as a skill and rerun it for the next product.

3. A data analyst that explains why, then runs every Monday. Demo requests had dropped about 30% in a month. She keeps demo requests, ad spend and website traffic in one Google Sheet, and recommends cleaning data into a sheet or BI tool as a shared layer whenever more than two sources are involved. Using the data plugin, she asked the agent to find out why, check each likely cause against the data and build an interactive dashboard. Fifteen minutes later the dashboard opened with its conclusion: mobile conversion breaks explained 95% of the lost demos, with a device breakdown, a trend line and a list of other hypotheses. Her point is that the value is the hypothesis building across datasets, not the chart, and you still check it yourself. Then she scheduled it: refresh every Monday and alert her if demo requests fall more than 10%.

4. Lead generation from the deals you lost. “The fastest way to find new customers is to start with the people who already said maybe.” For a B2B content platform she loaded the old lost leads, used the free LinkedIn plugin to confirm each contact’s current company and look for buying signals such as funding, hiring content marketers or a product launch, and found several contacts had moved jobs, which turns an old lead into a new account with a warm contact. The agent scored each lead on the three things she always uses, ICP fit, level of pain and timing urgency, then produced a plan with an angle and channel per account, lookalike accounts to add, a two-week outreach sequence and personalised Gmail drafts. She notes that in practice you would split this into two steps and enrich with proper ABM data.

5. A campaign brief that starts from your best campaign. Connect a project to the team drive holding brand assets and product shots. Use the template creator plugin to build a reusable brief template, and her tip is to upload the brief from a campaign that performed well and build the template from that. Ask for three campaign angles, pick one, and ask the agent to write the brief and create the assets and save everything to the drive. In about 15 minutes she had a full brief covering messaging, offer, creative direction and channels, plus six assets in feed, square and story sizes using the product shots from the drive.

6. Competitive teardowns by three sub-agents at once. Give it three competitor sites and your own, and ask it to explore each one “like a buyer”. It splits into three sub-agents, one per competitor, and after about 20 minutes saves a SWOT, a tracker comparing target buyer, positioning, core offers, pricing, lead magnet and site flow, and a folder per competitor with annotated screenshots of each funnel stage. She thinks the project plan it added was overkill but calls the tracker the most useful artefact for positioning work, especially combined with the audience research above.

7. A voice profile saved as a skill. Pasting samples and saying “write like this” only copies the surface. She pulled her last 20 long-form video scripts, asked the agent to analyse rhythm, structure and word choice, to split the work into sub-agents because it is a batch job, and to produce a voice profile and a reusable skill. The profile covered her voice identity, speaking patterns and story structure, with examples and a checklist to run before a final draft. She then tagged the skill in a ChatGPT document and had it draft a hook and intro from an outline. Her honest note: “Claude still does a little bit better job here.” The same method works for email, social, a newsletter or several client brands.

8. A proposal that writes itself from a call. The project already holds pricing, past proposals and the brand kit. Attach a discovery-call transcript and ask for a proposal that can win the deal. The agent works out what the prospect needs, researches, and builds both a PDF and a live proposal page carrying the prospect’s name and logo, opening with their situation rather than a template intro, choosing the most relevant case study, and presenting three pricing plans the prospect can select, update and accept on the page. Her principle: “put what stays the same inside a project and only attach what changes.” The same shape produces a statement of work, a budget request or an audit report.

Also in the video. An AI visibility agent that takes one high-intent buyer question at a time, checks what ChatGPT and Google AI Mode currently answer, and writes a content brief containing the answer snapshot, the fan-out questions the AI also answers, how you want to be described, the structure and evidence to use, and a writing assignment per sub-question. She warns the browser may be blocked from ChatGPT and you may need to log in.

Do this now. Build number three first. Put demo requests, spend and traffic in one sheet, ask ChatGPT Work why last month’s number moved, and read its hypotheses against your own. If it finds one you had not considered, schedule it for Monday mornings with an alert threshold.

Watch more:

3. One call transcript into a branded deck, a live Gantt chart and a client pack, with Claude inside Google Workspace

Casey of Blazing Zebra ran a marketing agency for more than a decade and now teaches business owners to put AI to work. On 11 October he published a 13-minute walkthrough of Claude’s new Google Workspace add-on, which arrived in public beta on 6 October for all paid Claude plans. His example is a community festival with sponsors, so swap in a client kickoff, a campaign plan and a partner pack as you read.

Set up. Install Claude from the Google Workspace Marketplace, at admin level on a Team account or personally, making sure it is the one published by Anthropic. Then in Sheets, Docs and Slides go to Extensions, search add-ons and install it there too, after which Claude opens in a sidebar. If you connected Google Sheets, Docs or Slides to Claude before, disconnect and reconnect each one under Customize and Connectors so they refresh. Set the approval level: he recommends starting with “needs approval” and, inside the Google extension, “ask before making any edits”, then moving to “accept all edits” once you trust it. It works in Chrome and most browsers but not in Firefox. From the sidebar you can also allow web search, add other connectors and choose the model, as you would in the Claude app.

Copy first, design second. Working from inside Claude rather than Google, he attached a call transcript and asked: “Can you please use your Google Slides connector to create a slide deck for sponsors from this call transcript? No design yet, just content and copy.” The deck that came back followed the transcript exactly. Then the second prompt: “Can you please stylize these with some beautiful renderings and various layouts and color ideas using the Unity design system that you have access to?” Every slide came back with a different layout in one consistent style, which he says is the first time he has seen Claude do that in a single prompt.

The design system shortcut. A design system in Claude Design is what makes the second prompt work, and building one from scratch takes time. His trick is designmd.app, a site of ready-made design systems. He found one called “cut and paste”, a paper-collage zine style that fitted a local event, pasted its link into Claude Design and said “please create a new Claude design system based off of this resource”, and had a usable system within minutes. For your own business, spend a few hours on it properly. For a spec project or a pitch, borrow one.

A Gantt chart that moves when you change a date. In Sheets, he first asked Claude to build a task list from the project documents, selecting the specific files and columns it should use. Then: turn this task list into a Gantt chart using conditional formatting with formulas, so that changing a start or end date in the list moves the bar. He changed a task from 7 October to the 17th, with an end date of the 19th, and the bar moved. He added a budget tab and a risk tab that flags anything risky, and styled the sheet by uploading the design system notes and asking Claude to apply them. Because it is a normal Google Sheet, the whole team can share it and keep it current.

A client pack from the knowledge base. In Docs, he opened the sidebar, loaded the knowledge base files and asked for an information pack for one sponsor, a dentist sponsoring the kids’ zone, that answered the questions in her email and added the day’s schedule, arrival instructions and what to bring. Then he applied the design system to the document. His hunch, which he tested, is that the design system applies better when you work from inside the Claude app, which has direct access to it, than when you upload the design file into Google Docs.

When to use Claude’s own Slides instead. Asked to apply a heavily stylised design, Claude recommended its own slide tool, so he ran both. Claude Slides produced the tape marks and off-kilter collage layouts the design system called for, which Google Slides cannot render, and it even rebuilt a site map from the PNG maps he uploaded. Google Slides stayed native and shareable. His rule: Google Slides when the deck has to live in Google, Claude Slides when the design is the point.

Images and batching. Claude Design cannot create images, so he asked it for short image prompts with dimensions, generated them in Magnific, which he notes is not a sponsor, and dropped them back in. And the sponsor pack is the kind of thing to turn into a skill: write it once, then batch it for every sponsor, each one branded and carrying that sponsor’s specific details.

Do this now. Take the transcript of your last client or stakeholder call, run the two-prompt deck in Google Slides, and judge the copy before you judge the design. Then build one Gantt chart from your live campaign plan with conditional formatting, share it, and see whether the team keeps it updated when the chart does the updating.

Watch more:

1. Semrush: fewer than half of the brands AI answers mention are also among the sites they cite

Semrush published the second 2026 edition of its AI Visibility Index on 13 October at its Spotlight conference in London, adding a UK benchmark for the first time. It covers seven monthly snapshots from January to July, each running a fixed set of 2.5 million UK and 25 million US prompts, built from Semrush search data, through ChatGPT, Gemini and Google AI Mode, plus a larger variable set of AI Overviews results, across 22 categories.

The central finding separates two things most visibility tools blend. Being named in an answer and being cited as a source for it are different outcomes, and averaged across categories and platforms, fewer than half of the top 100 mentioned brands also appeared among the top 100 cited domains, in every month and in both markets. Ford sat in the top 100 mentioned brands on all four platforms in July, while ford.com was the 1,055th most cited domain on ChatGPT, 1,146th on Gemini and 1,320th on AI Mode. In the UK, Tesco and Asda were both top-100 mentions everywhere, but on ChatGPT Tesco was the 25th most cited domain and Asda the 606th. The four platforms agreed with each other less in the UK than in the US on which brands to mention and which sites to cite, AI Overviews overlapped least with the other three, and in UK AI Mode the ten leading domains took between 23.2% and 35.8% of all citations, against 12.8% to 15.4% on ChatGPT. Differences between categories were larger than differences between platforms.

What to do. A brand that is mentioned but not cited is being described by other people’s pages, which is a PR and third-party content problem. A brand that is cited but rarely mentioned has the right pages and the wrong positioning, which is a messaging problem. Run a fixed set of customer questions through each platform and record mentions and citations separately before you fund any AEO work, then give the gap to the team that can fix it rather than defaulting to “create more content”. Semrush sells AI visibility tools and now owns MarTech, which wrote one of the analyses we read; the examples here are consumer brands, so expect your category to behave differently.

2. Typeface: 49% of marketing leaders plan to cut back on traditional SEO, and 62% say brands are trading quality for AI citations

Typeface’s Signal Report, The AI Visibility Gap, published on 14 October, surveyed more than 500 US and UK marketing leaders with titles from director to vice president. 49% plan to deprioritise or stop investing in traditional SEO as AI discovery grows, and 44% plan to raise spending on answer engine optimisation over the next 12 to 24 months. 64% expect AI chatbots to drive more brand discovery than search within three years, and 97% expect AI to be a top-three way customers find their brand by 2030. The finding Typeface leads with is self-critical: 62% say brands are optimising for AI citations at the expense of content quality, 61% say the same about brand authenticity, and 68% say they are making content for AI that people will probably never read. Only 30% say their teams collaborate closely on AEO.

What to do. The two halves of this report describe the same teams, which is the point. If your plan for next year moves budget from SEO to AEO, write down what quality means for the pages you expect AI to cite and who signs them off, because the respondents’ own verdict is that the race to be cited is lowering the bar. Typeface sells content and AEO software, and the report gives no fieldwork dates.

3. G2: two thirds of B2B buyers have agents in production, and nearly a third say staff spend six hours a week watching them

G2’s 2026 AI Agents Report, published on 14 October, draws on a survey of more than 1,000 B2B software buyers plus G2’s review data. Agent adoption among B2B software buyers rose 25% year on year and 66% of organisations now have agents in limited or scaled production, with nearly half of those that completed pilots saying most have gone to full production. G2’s name for what comes next is the distrust tax. 31% say employees spend six or more hours a week supervising agent work, 45% say time spent overseeing and correcting agents consumes more than a quarter of the productivity gains they expected, and 46% report higher operating costs than planned. 38% of organisations with agents in production run more than 100 of them, yet 47% manage them informally. Buyers want proof before autonomy: 52% require human quality review before production and 77% re-evaluate agent performance at least monthly. On pricing, only 9% think per-user or per-seat pricing is fairest, and 72% say an ecosystem of connectors, actions and skills is very important or a must-have.

What to do. Before you add a third or fourth agent to a marketing workflow, count the hours someone spends checking the first two and put that number next to the time saved. Then list every agent running in your team’s tools, including the ones inside HubSpot, Salesforce and your ad platforms, because the 47% managing informally are the ones who will be surprised by a bill or an email that went out wrong. G2 is a software review marketplace that sells to the vendors it surveys about, and the report gives no fieldwork dates.

1. Claude now works inside Google Docs, Sheets and Slides

Anthropic’s Google Workspace add-on, announced on 6 October, is in public beta on all paid Claude plans. It opens Claude in a sidebar beside the file you have open, reads the document, sheet or deck and your selection, and edits in place: fixes and suggestion cards in Docs, formulas, pivot tables, charts and Python-based cleaning in Sheets, and new slides that match the deck’s layout and theme in Slides. The default is “ask before edits”. Signed in, you get the same models, connectors and skills you use in Claude, so a quarterly review deck can pull Salesforce account history through a connector and the format can be saved as a skill. Separate Google Docs, Sheets and Slides connectors, also in beta, let Claude create and edit Google files from inside a Claude conversation. On Team and Enterprise plans an owner has to enable the connectors first. It does not work in Firefox and only handles native Google files, not Word or Excel files opened in Google.

2. Meta’s Ads Manager assistant now remembers your account, and is testing building campaigns by chat

At Advertising Week New York on 6 October Meta said its AI business assistant inside Ads Manager now carries a deeper memory of your business, “what you tried last quarter, your seasonality, and the goals you’ve described”, with creative insights from Ads Creative Studio brought into the chat. In test: importing your strategy, voice and brand guidelines from other AI tools, scheduled tasks and alerts for performance monitoring, weekly summaries and budget reminders, and what Meta calls agentic capabilities, where “soon” you will plan and create a campaign, generate creative, update targeting or change a budget conversationally, with previews and approvals. For larger advertisers, Meta’s ads MCP server now lets an outside assistant upload creative, preview ads and send work to Ads Manager as a draft, with every action gated by advertiser-set permissions; AdManage.ai, Databricks, Kochava, Perplexity, Snowflake and Triple Whale are among ten partners. Two things are live for everyone: Customer Lifecycle Strategy, and video generated from existing static assets in Advantage+ creative. Meta’s claim that advertisers who adopt the assistant’s recommendations see 5% more conversions at the same cost comes with no method or footnote.

3. Salesforce Winter ‘27 went live on 12 October, with Agentforce switched on by default

The release gives the Marketing Cloud Engagement MCP server 40 new tools, so an AI assistant can create and run Automation Studio activities, manage campaigns and Content Builder folders, upsert bulk data, query data extensions and pull email tracking events in plain language. The Agentforce platform is now enabled by default for every org whose edition includes it, with no change to billing, and the on-off switch is being removed from Setup later in the release; you can still turn it off under the Einstein setting or deactivate individual agents in Agentforce Builder. One date to act on: from 20 October, prompts that call Gemini 2.5 Pro, Flash or Flash-Lite are rerouted to Gemini 3.5, and Salesforce’s admin blog warns the responses can change, so retest any marketing prompt templates that run on Gemini before then. Salesforce has also described a Marketing Goals Agent and a Content Agent for this release; their availability by edition is not stated on the pages we could read, so ask your admin.

4. Gamma 5 rebuilt its editor to take the “AI smell” out of decks

Gamma, the AI presentation tool, launched Gamma 5 on 6 October as a full rebuild. The headline changes are a new editor for reworking a generated deck rather than regenerating it, a wider range of visual styles, the ability to import and edit PowerPoint, PDF and infographic files, and persistent brand identity settings so every output starts from your fonts, colours and logo. Gamma offers no data on whether audiences notice the difference, so run your next internal deck through it beside your current template and ask a colleague to pick the one that looks produced by a person.

5. Kling 4.0 makes 30 seconds of video in one pass, and HeyGen launched its own voice model

Kling AI released Kling 4.0 globally on 16 October. It generates up to 30 seconds of video in a single pass from as many as 15 reference assets and 10 keyframes, takes image, video and audio references and prompts up to 8,000 tokens, and can be driven by an agent through its MCP server and command-line tool. Kling 4.0 and a faster 4.0 Flash are coming to Adobe Firefly. Enterprise API access and pricing are “coming soon”, and two same-day releases disagree on whether 4K HDR output is available now. On 9 October HeyGen launched HeyGen Voice, an in-house voice model that is free in the platform and API, with a Professional Voice Clone at $99 a month trained on 30 minutes to three hours of consented speech. HeyGen’s survey of its own small-business users found 71.6% had recorded a business video and never posted it. If you make explainer or demo video, the 30-second single pass is the number that changes a workflow.

1. Google has delayed skills inside the Workspace apps, and the Gemini app rollout is under way. Two weeks ago we reported that skills would reach Gmail, Docs and the other Workspace apps from 5 October. Google’s 13 October update says those rollouts “are slightly delayed” and are now “delayed until further notice”, while skills in the Gemini app for Workspace accounts began on 13 October and should finish by mid-November, with sharing and a Drive integration to follow. Build in the Gemini app for now.

2. GPT-6 is now the default model in ChatGPT for every plan. Plus, Pro, Business and Enterprise moved to GPT-6 Sol on 7 October and Free and Go to GPT-6 Luna on 8 October. The release adds what OpenAI calls Intelligent UI, answers that mix text with charts, comparison tables, diagrams, forms and small working tools when the question suits it. The Pro reasoning setting stays on GPT-6 Astra and does not use Intelligent UI. If your team has prompts that assume a plain-text answer, check what they return now.

3. Claude Haiku 5.5 arrived, and Max and Team plans now include monthly API credit. Released on 7 October, Haiku 5.5 costs about 75% less to run than Haiku 4.5 on Anthropic’s figures, at $0.10 per million input tokens and $0.50 per million output for prompts under 100,000 tokens, and HubSpot reports it scored 92.8% on its simulated CRM task suite. Alongside it, Max 5x subscribers get $100 a month of credit on the Claude Platform, Max 20x gets $200 and Team plans get up to $500 pooled, for running their own agents and apps, claimed by linking a Console organisation in billing settings. It does not cover extra usage in Claude Code or Cowork and does not roll over.

4. Copilot Studio agents can now have their own Microsoft 365 account. Announced on 15 October and rolling out through Microsoft’s Frontier early-access programme, an agent account gives an agent its own identity, mailbox and presence in Teams and the company directory, so colleagues can @mention it in Teams, Outlook and Office comments and delegate work by email while it carries on between interactions. It needs an E5 licence with the Agent 365 add-on or an E7 licence, plus a Frontier for AI Teammates licence per agent. In this preview the agent cannot reply to anyone outside your tenant or read documents labelled confidential.

1. AI referral measurement is held together by convention, and this week it slipped. ChatGPT’s tag disappeared for a week by accident and 31,000 visits a day on one sample moved into Direct. Google announced more links from AI answers in the same week it reminded everyone that Search Console reports one position for a whole AI Overview. Semrush found that mentions and citations are different things that most scores add together. Every number you report about AI search this quarter rests on labels another company controls. The teams that will keep leadership’s trust are the ones that annotate the gaps, measure mentions and citations separately, and keep a fixed set of buyer questions they retest by hand.

2. The finished output moved into the free tier and into the file you already have open. Claude Docs, Slides and Design on every plan, a sidebar in Google Docs and Sheets, Gamma rebuilt around editing rather than regenerating, Meta’s assistant drafting campaigns inside Ads Manager, and Grace Leung’s deck, dashboard and proposal arriving as finished artefacts. The assistant used to be where you went to get a draft to take elsewhere. Increasingly it is where the thing is made and where it lives. The skill that matters more each month is the brief: the project context, the design system and the one-page rulebook that make the first output usable.

3. The agent became an employee, and employees need managing. An Anthropic rep with six named agents and a coordinator. Google and Microsoft giving agents an email address and a directory entry. Salesforce switching Agentforce on by default. And G2 putting a name, the distrust tax, on the six hours a week someone spends watching all of this. Nobody in this issue is arguing that agents do not save time. The argument is about what sits around them: a rulebook, a shared memory that learns from corrections, a scoped credential, and a person whose job now includes reviewing the summary at the top of every draft.

CLAUDE FOR MARKETERS MASTERCLASS - NEW MODULE ADDED:

THE BEGINNER’S GUIDE TO CLAUDE CODE FOR MARKETERS

Just $150 OR 2 payments of $75

(Individual price $953).

Check Out This Limited Time Offer

Tell me: What’s your biggest takeaway from this week’s news from the world of AI Marketing? Love to hear your thoughts in the comments.

Leave a comment

Thanks for reading the Marketing with AI Weekly Roundup! If you enjoyed this read, the best compliment I could receive would be if you shared it.

Marketing with AI Weekly Roundup: Rebuild Gemini Gems as Skills, Set Rules for Always-On Agents, and Check Google's View of Your Brand

Robert Gillespie — Sun, 04 Oct 2026 10:04:28 GMT

Welcome to issue #54 of the Marketing with AI Weekly Roundup. Published every Sunday, it shares the top stories, playbooks, data and quick hits from the world of Marketing with AI over the last 7 days.

Subscribe to receive the week's most important AI marketing stories direct to your inbox.

Subscribe now

FINAL CHANCE TO INVEST IN CLAUDE FOR MARKETERS MASTERCLASS

For marketers who want their hours back. The full Marketing Operator System, the Setup Skill that builds your working environment, and 5-Minute Slides, together for the first time.

Now just $150 OR 2 payments of $75

(Individual price $953).

Check Out This Limited Time Offer

If your team has built Gems in Gemini, they now have an end date. Google announced on 30 September that skills, reusable instructions you call with a slash, are replacing Gems across the Gemini app and Workspace, with the Workspace rollout starting on Monday. That makes Gemini the third major assistant to settle on skills as the way you save how a job should be done.

OpenAI used its developer day to launch Dots, agents with their own cloud computer that keep working between conversations, though UK and EU readers on the Pro plan cannot have one yet. Two datasets show Google’s AI Overviews now appear on roughly four in five searches for a brand’s own name. And Bain found that only 6% of marketing organisations say AI is delivering a significant performance impact, five months after nearly all of them had adopted it.

The Playbook has three things to try. Anthropic’s own sales team explains how it replaced the contact form with a buying agent. HubSpot’s Kieran Flanagan builds the three small files that stop AI drafts sounding like everyone else’s. And Jeff Su shows the one-page rulebook that lets an assistant file and name everything in a shared drive.

1. Google is replacing Gems with skills, and the Workspace rollout starts on Monday

On 30 September Google introduced skills in Gemini. A skill is a saved set of instructions for a job you repeat. You run one by typing a forward slash and its name in the prompt bar, you can use several in the same prompt, and a skill can carry reference files such as text documents, PDFs and images. Unlike a Gem, which is a separate chat you have to open, a skill works inside whatever conversation you are already in.

For personal Google accounts, skills are live in Gemini chat now for users over 18. For Workspace, Google’s update post gives dates. Skills start rolling out inside the Workspace apps on 5 October and in the Gemini app for Workspace accounts on 13 October, with both expected to finish by mid November. Every Business and Enterprise edition is included, Starter among them. One limit to plan around is that skills do not yet sync between the Gemini app and the Workspace apps, so a skill you want in both has to be created in both.

Gems are being retired. Google says support ends starting in November for personal accounts, no sooner than 1 March 2027 for Workspace business, enterprise and non-profit customers, and 1 June 2027 for education. Remaining Gems will be migrated automatically into draft skills, though Google recommends rebuilding the important ones by hand. Opal, Google’s experimental app builder, is being turned down in November as well. Google’s help page adds two details worth knowing: only 100 skills can be active at once, and you can import a skill built on another platform by uploading its SKILL.md file.

Why this matters. Claude has had skills for a year, and OpenAI began retiring custom GPTs in favour of skills in September. With Gemini following, all three now store “how we do this job” the same way, as a short written instruction file. For a marketing team that means the work of writing down your brand voice rules, your campaign brief format or your reporting steps is no longer tied to one vendor. For teams on Google Workspace it also means something they built is going away on a fixed date.

What to do. List the Gems your team relies on this week, and note who owns each one and whether any sit inside a Workspace Studio flow, because those are the ones most likely to break. When skills reach your account, rebuild the two you use most and compare the output against the Gem before switching. If you already have skills in Claude or ChatGPT, try uploading one of those files to Gemini and see how much carries across.

2. OpenAI launched Dots, agents that keep working after you close the chat

At its developer day on 29 September, OpenAI introduced Dots. A Dot is an always-on agent inside ChatGPT with its own cloud computer, powered by GPT-6 Astra, the model built to operate software. You give it an ongoing responsibility rather than a single task. It carries on between conversations, remembers the context, uses the apps you connect through ChatGPT’s plugins, of which OpenAI says there are more than 4,000, and messages you with progress, questions or a decision it needs. You can reach it in ChatGPT, Slack or Microsoft Teams. OpenAI lists email marketing among the jobs it has tested Dots on internally.

Who can use one. One Dot is included on the Pro plan and on Business Premium seats. OpenAI’s help centre says Pro access excludes the European Economic Area, Switzerland and the UK at launch, while Business Premium has it in every region ChatGPT supports. Enterprise and Edu workspaces get a beta that is switched off until an admin enables it. A Dot has to be created in the desktop app or on desktop web, not on a phone. For the first month, Dot usage does not count against your plan’s allowance, and OpenAI says it will set out usage terms after that.

The same event brought GPT-6.1 Sol, an upgrade to the Sol model we covered last week, available in the API and to Plus, Pro, Business, Enterprise and Edu users. OpenAI describes it as close to Astra’s capability at one fifth of Astra’s token price.

Why this matters. Microsoft announced Autopilot, its own persistent agent, four days earlier. The two largest workplace assistants are now selling the same idea, an agent that holds a job rather than answers a prompt. The marketing jobs that fit are the recurring ones nobody enjoys: the weekly competitor check, the pipeline digest, chasing the status of campaign assets. The limits are just as clear. A UK marketer paying for Pro cannot use a Dot today, and nobody yet knows what a Dot costs to run after the free first month.

What to do. UK and EU readers who want to test one need a Business Premium seat, so find out which plan your company is on. Whatever your access, pick one recurring job now and write two short lists for it: what the agent may do on its own, and what it must ask about first, such as sending anything to a customer. That page is the brief you will need on day one, for a Dot or for any of its rivals.

3. AI Overviews now appear on about 83% of searches for a brand’s own name

Two datasets published this week show the same jump. Ahrefs analysed 232.8 million US desktop search results from July to September, 92.5 million of them for branded keywords. On 29 September, 82.91% of branded keywords returned an AI Overview. That is a single day’s figure, and the monthly averages show the direction: 61.3% in July, 65.1% in August and 73.3% in September. In a separate panel of the 100 companies on Interbrand’s Best Global Brands list, searching the bare brand name produced an AI Overview for 63 of them by 30 September, against 5 on 1 July, and about three quarters of those overviews sat in position one.

DemandSphere, which tracks branded keywords daily across all markets and devices, saw a sharper move inside September alone, from 26.1% on the first of the month to 82.1% on the 29th, with a peak of 90.5% on the 27th. It has not said how many keywords it tracks, and its daily figures swing widely, so read it as confirmation of the trend rather than a precise measure.

Ahrefs also looked at what those overviews cite. Across the 57 overviews in its brand panel where the source list was available, Wikipedia appeared in 71.9%, YouTube in 38.6% and LinkedIn in 22.8%.

Why this matters. A search for your company name used to be the one results page you more or less controlled, with your homepage first and your own links beneath it. Increasingly the first thing a buyer sees is Google’s summary of you, assembled partly from sources you do not own. For a B2B company that is the page a prospect lands on after a colleague mentions your name, which makes it one of the most commercially important searches you have.

What to do. Search your brand name today, then your brand plus “pricing”, “reviews” and “alternatives”. Record what the overview says and which sites it cites. If it is wrong or thin, start with the sources Ahrefs found most often: your Wikipedia entry if you have one, your YouTube channel and your LinkedIn company page. Both studies come from companies that sell search tracking, and Ahrefs’ brand panel is made up of very large global brands, so a mid-sized B2B company may see different numbers.

4. Bain: nearly every marketing team has adopted AI, but only 6% say it is paying off in a big way

Bain published AI in Marketing: How Leaders Achieve Double the Revenue Impact on 30 September, based on a survey of 1,397 chief marketing officers, chief financial officers and senior marketing and finance executives carried out in April 2026. Only 6% of marketing organisations say AI delivers a significant performance impact today. Business Insider, which had the report first, puts adoption among the same respondents at 95%.

Bain then compares two groups. Leaders, 121 companies with revenue growth above 11% and market share gains of seven points, are set against laggards with flat or falling revenue. The share of laggards calling AI a core capability rose from 8% to 30% in a year, so adoption is no longer what separates the two. What does:

Leaders are 3.7 times more likely to have redesigned workflows around AI rather than adding it to existing ones, and twice as likely to have restructured teams and job descriptions.
Leaders are 1.8 times more likely to follow a centralised AI roadmap set from the top. Laggards are 10 times more likely to have no defined AI strategy for marketing.
Leaders are 2.4 times more likely to embed AI directly in their marketing tools. Only 6% of leaders rely on general-purpose assistants such as ChatGPT, Copilot or Gemini, against 16% of laggards.
Nearly 70% of leaders regularly change strategy and spending based on what AI tells them, against 31% of laggards.

Why this matters. This is the third study in three weeks with the same shape. Alexander Group found 96% adoption and 41% able to prove a return. HubSpot found 98% using AI and 10% transformed by it. Bain’s version gives the clearest account of what the small group does differently, and none of it is about which tool they bought.

What to do. Name one person who owns the marketing AI roadmap, even if it is a one-page list. Then choose one workflow, not one task, to rebuild from the first step to the last with AI in it, and attach a revenue or pipeline number to it. Three caveats: the fieldwork is five months old, the sample mixes B2B and consumer companies with no separate B2B figures, and Bain sells the transformation work it recommends.

1. Replace the contact form with an agent the buyer chooses to talk to

Anthropic’s sales team was receiving tens of thousands of inbound requests a month, and buyers were waiting days for answers to basic questions about pricing and compliance. On 30 September it published how it replaced the form with what it calls a buying agent, built on Claude Managed Agents, its hosted agent service, which is still in beta. The detail is useful whichever vendor you use.

It is opt-in. On the Contact Sales and Pricing pages, inside the product and by email, the buyer chooses at the start whether to talk to the agent or to a sales rep. Nobody is forced through a bot to reach a person.

Three jobs, three exits. The agent asks follow-up questions to understand what the buyer needs and how they use Claude today. It answers questions on pricing, seat minimums, security requirements and billing from a knowledge base, and recommends a plan and a number of seats. Every conversation then ends one of three ways: the buyer goes straight to checkout, the buyer is handed to a rep with the full conversation attached, or the buyer simply has the answer they came for.

A goal worked better than rules. The team started with detailed instructions and found that a short statement of purpose performed better. The version they quote is “Your goal is to understand customer requirements, qualify prospects, and recommend the best plan.” Their summary is that less is more, with knowledge and context doing the work that a flowchart of rules could not.

Sales and content people edit the prompt. Engineering owns the code, but sales and content leads review and change the system prompt themselves. Changes go to a staging version first, every change is saved as a version that can be rolled back, and the prompt has been updated weekly since launch.

It recommends the smaller plan when that fits. The agent will suggest the Team plan over Enterprise where that suits the buyer, and the team treats every handoff from agent to rep as feedback on a gap to close. Fixing those gaps has cut the share of conversations that need a person by about half.

Their results, which are theirs. One engineer built the first version in a few weeks. Leads the agent passes to reps convert to opportunities more than twice as often as leads from the old form, deals close about five days faster, the emails exchanged per deal fell from about ten to about six, and one inside sales rep closed 2.5 times as many deals. The agent handles thousands of conversations a day. This is a vendor describing its own product in use on its own buyers.

Do this now. Pull the last hundred contact form submissions and list the ten questions that come up most. Write the answers on one page, including the ones your team usually saves for a call, such as minimum contract size. Then define your three exits: when a buyer can buy without you, when a person takes over, and what counts as answered. That page is the knowledge base, whether an agent or a new starter ends up using it.

2. Three small files that stop AI drafts sounding like everyone else’s

Kieran Flanagan, HubSpot’s senior vice president of marketing, used the 29 September episode of Marketing Against the Grain to build what he calls a creative context system: a folder holding three short text files that any assistant can read. His starting point is the idea from Nicolas Cole that we covered last week, that AI produces commodity content because it knows everything except what makes you distinct. His argument is that when production costs nothing, “judgment and taste is the only thing with a price”.

creator.md: how you think and write. The wrong way to build it is to describe yourself in adjectives such as professional, conversational or insightful. Instead, export your real work. He used his LinkedIn posts and his Substack archive. Ask the assistant to shortlist the best performers, then go through the list yourself and remove the one-offs. His own top posts included a job change, a product launch and a piece of satire, none of which is a pattern worth repeating. The file that comes out covers how he builds an argument, the ways he tends to explain ideas, his tone and sentence style, and a line on how to adapt each for LinkedIn or Substack. The instruction at the top of his is that AI should preserve the way he sees a problem and not average him out.

If you have no archive yet. He recommends copywork, the old copywriter’s exercise of writing out an admired writer’s work line by line. Give the assistant pieces by a writer whose structure you would like to learn from, use that as a starting file, and edit it towards your own voice as you publish. He is blunt that this does not replace learning to write.

audience.md: who you write for. He builds it from three inputs. First, five dream readers, real people whose public profiles you give the assistant so that it can describe what they have in common. Second, the people already engaging with you, by analysing who comments on your posts and telling the assistant to ignore comments that look machine-written. Third, outside research into what that audience currently cares about and is tired of hearing. His file ends with a rule: do not explain why AI matters, explain what excellent marketing looks like now that AI exists.

taste.md: your quality bar. This is the one he says you cannot hand over. Collect a dozen or more pieces of writing you consider excellent for your audience and say why each one works. His list included Kevin Kelly’s 1,000 True Fans, for one simple idea that lasts, and Ben Thompson, for naming something that was already happening. Then have the assistant interview you. It asked him whether the idea or the prose matters more, what makes a contrarian argument good, and when a framework is worth using. The answers become a file with his standards, a list of things never to do, and a publish test: could the reader explain the core idea to someone else in ten minutes or less? If not, simplify.

How he uses it. He uses it for ideas and as a sparring partner, never for final drafts. In the episode he asks for a first draft, which is fine, then asks the assistant to run it through taste.md alone, and the opening changes to a much sharper line about AI unbundling the marketer. He then feeds in a LinkedIn post of his that had underperformed. The file’s verdict was that the opening hedged and the post carried too many ideas, which he agreed with. His tip is to have the assistant read one file at a time, because loading all three at once weakens each of them.

This is HubSpot’s channel, the prompts are offered in exchange for an email address, and the tool he uses to analyse his commenters, YouSpot, comes from HubSpot’s co-founder Dharmesh Shah.

Do this now. Start with taste.md, because it needs no data. Pick five pieces of writing you wish your team had produced, write two sentences on why each works, and ask your assistant to interview you about them and turn your answers into a one-page quality bar. Run your next draft against it before anyone else reads it.

Read more:

3. A one-page rulebook that lets AI file and name everything in your shared drive

Jeff Su, who makes productivity tutorials, published a 14-minute walkthrough on 29 September of a system he says has filed his documents every day for five months. It has passed 90,000 views. The value for a marketing team is that agents are only as useful as the files they can find, and most shared drives are organised by habit rather than by any rule a machine could follow.

The rulebook has three parts. It is a plain text Markdown file, which the assistant can both read and edit. The first part is an organisation framework, which says where a file belongs and why. His version starts with areas of ongoing work, and inside each area sits a resources folder for material you keep returning to, a projects folder for work with an end date, and an archive. His worked example is a product marketing area, with the brand kit, logos and guidelines under resources, live campaigns under projects and finished ones in the archive. The second part is naming conventions built around what you will search for later: short descriptive names for standalone documents, the year and quarter first for project folders, and the full date for invoices and receipts. The third part is a filing workflow that tells the assistant when to stop and ask.

Setting it up. Create a root folder containing an inbox folder and a reference folder, put the rulebook in reference, and give a desktop assistant access to the root. He uses ChatGPT Work and says the same approach works in Claude and in Gemini’s desktop agent, but not in an ordinary browser chat, which cannot reach your files. Drop a few files in the inbox and ask the assistant to read the rulebook and propose a plan. It returns each file’s current name, proposed name and proposed location, and moves nothing until you approve.

The rulebook learns. In his demo the assistant could not tell whether two receipts were personal or business, so it asked. He told it which card was which and added “remember the card rule”, and the assistant wrote that rule into the rulebook itself. The next receipt was filed without a question.

For a drive that already exists. Install the desktop sync app for Google Drive, OneDrive or Dropbox so the assistant can see the folders. Ask for two folder trees with nothing changed: the current state, and a proposed future state under your rulebook. His own audit found his finance records living in two places. Answer the clarifying questions it asks, then move the folders yourself on the first pass, with the assistant guiding you, so that you learn where things now live. Leave renaming for later so it does not turn into a renaming project. Where you want an exception, such as a shared folder kept at the top level for convenience, write it into the rulebook.

Then automate it. Create a daily scheduled task that checks the inbox folder and proposes changes for you to approve. He says to expect more questions in the first few days and about 95% accuracy after two weeks, which is his own estimate.

The video is sponsored by Dropbox, which he says on screen, and it links to a free starter rulebook and to his paid course.

Do this now. Write the first version for one team folder, not the whole drive. One page covering where campaign assets, briefs, reports and brand files live, how each is named, and when the assistant must ask. Run it against last month’s downloads and see what it gets wrong.

Read more:

1. The world’s biggest advertisers are measuring AI visibility, and 8% have a strategy for it

The World Federation of Advertisers published its State of AI in Marketing 2026 on 30 September, based on 54 senior respondents at 46 of the world’s largest brand owners, with combined advertising spend of $93 billion. 96% now use generative or agentic AI in marketing, up from 45% in 2023. Content creation leads at 80%, followed by research and insights at 72% and media buying at 54%, with 46% using AI for workflow automation or autonomous agents.

Two findings travel well to B2B. On AI search, 73% are actively monitoring and measuring their visibility in AI answers and 67% are adapting their SEO, but only 8% say they have a clear strategy in place. On agencies, only 13% say they have a great deal of visibility into how their agencies use AI, even though 69% require agencies to disclose it and 59% have added AI clauses to contracts covering intellectual property, confidentiality and training data.

If you work with an agency or freelancers, the second finding is the one to act on. Ask where AI is used in the work you buy and what your contract says about it. This is a small sample of very large consumer-facing advertisers, and the release gives no fieldwork dates.

2. Forrester: more than half of B2B marketing leaders say they have already cut headcount because of AI

Forrester analyst Ian Bruce summarised its State of AI in B2B Marketing, 2026 report on 29 September. AI is in production across every marketing use case Forrester tracks, with most between 43% and 48% of organisations and advertising and media buying highest at 55%. Among marketing decision-makers, 76% say AI will augment people rather than replace them, yet 55% say they have already reduced headcount because of AI use, and a majority are reassessing hiring plans. Half say AI often presents incorrect information with unwarranted confidence, which creates checking work that can cancel out the time saved, and 30% name data privacy and security as their biggest obstacle.

The figures sit awkwardly together, which is the point. Teams are being resized on the expectation of productivity that half of the same leaders say is undermined by errors. If you lead a team, measure the checking time as well as the drafting time before agreeing to a number. Forrester’s post gives no sample size or fieldwork dates, and the full report is for clients only.

1. Microsoft Advertising now connects to HubSpot, and LinkedIn company lists grew tenfold

Microsoft Advertising’s September product update, published on 30 September, announced that it is now live inside HubSpot. You can build audiences from HubSpot contacts and reach them across Bing, Copilot, Outlook and the rest of Microsoft’s network, a new lead from a Microsoft campaign can trigger a HubSpot workflow such as assigning a rep or starting a nurture email, and ad clicks are tied to the contacts and deals in your CRM so you can see which ads produced pipeline. HubSpot’s knowledge base sets out which subscriptions qualify. In the same update, the company lists used with LinkedIn profile targeting rose from 1,000 to 10,000 company names per list, with bulk upload, for Search and Audience campaigns in supported markets. For anyone running account-based campaigns, that turns a sample of a target account list into the whole list.

2. Google Sheets can turn a spreadsheet into a working dashboard, and Workspace gained an image editor

Google’s Workspace update on 1 October introduced Sheets canvas. You describe what you want, such as a Kanban board, a scenario planner or an executive dashboard, and Gemini builds it as an interactive view of your data. It lives as a tab in the same file, stays in sync with the underlying sheet in both directions, and can be shared like any other sheet. The same update added Google Pics, an image tool inside Docs and Slides and as its own app, which lets you select and edit individual objects in an image and upscale to 2K or 4K, and Google Vids can now turn a document into a short explainer video with a script and voiceover. Google says these are rolling out to Workspace customers now but its post does not say which editions get which feature, so check before you plan around them. The campaign tracker most teams already keep in Sheets is the obvious first test for canvas.

3. Perplexity’s agent can now start work when something happens, not only on a schedule

On 29 September Perplexity announced Automations for Computer, its agent for multi-step work. An automation runs on a schedule or when an event happens in Slack, Gmail, Outlook or Linear, and uses Computer’s memory, skills and connected apps when it does. Perplexity’s announcement does not say which plans include it or how credits are charged, so check your account before relying on it. A sensible first use is a competitor pricing or messaging check that runs weekly and posts only what changed.

Read more:

4. Adobe rebuilt Marketo around agents and opened it to Claude, ChatGPT and Copilot

On 28 September Adobe announced what it calls the new Marketo. Coworker for Marketo Engage builds and maintains workflows and analyses data from plain-language requests inside Marketo. A Marketo Engage MCP Server connects Marketo to outside assistants including Claude, ChatGPT and Copilot. Two further products were announced alongside: Marketo Optimizer, which brings lead, account and buying-group journeys together and recommends next actions across marketing, paid media and sales development, and Marketo Qualifier, which combines inbound lead qualification and outbound prospecting in one flow for sales development teams. Adobe’s post describes what each does without stating availability or price for any of them. If you run Marketo, ask your account manager which of the four you can switch on today.

5. ChatGPT can take meeting notes without a bot, and joins Slack and Teams threads

Two more releases from OpenAI’s developer day. A Meetings plugin, in beta for Pro and Business plans with Enterprise to follow, records meeting audio and produces notes and a personal summary with action items, saved to ChatGPT Space, the new home for your files and notes. Notes can stay private or be shared with a team. And @ChatGPT in Slack and Microsoft Teams, for Business and Enterprise plans, lets you bring ChatGPT into a thread, where colleagues can add context and carry on the same conversation without a ChatGPT licence of their own. Before using the Meetings plugin on customer calls, read OpenAI’s help article on how recordings are handled and agree how you will tell the other side they are being recorded.

6. ZoomInfo’s agents now run a whole go-to-market play from a plain-English description

On 1 October ZoomInfo launched Agent Teams, a day after announcing it had bought the orchestration company DoubleO.ai. You describe a play in plain English and a set of specialised agents carries it out across your systems and ZoomInfo’s data, sharing context as they go and resuming when a new signal arrives, such as a prospect visiting your pricing page. ZoomInfo’s examples include tracking champions when they change jobs, and analysing a closed-won deal’s calls and emails to draft a case study, find similar companies and write outreach to them. It is available now in GTM Studio with no new contract or product to buy.

1. Google now says manual review of AI content is “critical”. Its guidance on using generative AI content, updated on 1 October, states that “it is critical to manually factcheck and review all AI-generated content for accuracy and trustworthiness before publishing”. The review extends to title tags, meta descriptions, structured data and image alt text. Most review steps cover the article and skip the metadata, so add it to yours.

2. ChatGPT custom instructions can now run to 5,000 characters. The limit rose from 1,500 on 29 September for Plus, Pro, Business, Enterprise and Edu users. That is enough room for your role, your audience, your formatting rules and a short list of phrases to avoid, so it is worth rewriting yours rather than adding to the old one.

3. Claude Sonnet 5.5 arrived. Released on 28 September, it is priced at $2 per million input tokens and $10 per million output tokens. Anthropic says it runs more than 30% faster than Sonnet 5 and costs up to 30% less for most work, on its own testing. Haiku 5.5 is due in the coming weeks. If your skills or automations run on Sonnet, rerun one and compare.

4. OpenAI is halving what the $200 Pro plan includes. From 30 October, Pro 200’s usage in ChatGPT Work and Codex falls from 20 times the Plus allowance to 10 times, and GPT-6 Pro messages in chat fall from 200 to 100 a week, according to The Next Web’s report of OpenAI’s Pro help page. Existing subscribers keep current limits until 29 October and receive a one-off usage credit that expires at the end of the year. A new Pro 500 plan costs $500 a month.

5. Microsoft put a price on Copilot Credits and moved the start date. Last week we reported that agent work in Copilot was moving to usage-based billing with no prices published. CRN, citing Microsoft’s partner FAQ and credits guide, now reports 1 cent per credit, with a light task using up to 300 credits, a medium one up to 700 and a heavy one more. For Copilot Business customers with up to 300 seats, usage billing becomes the default on 1 December rather than 2 November, with a default cap of 4,000 credits per user per month that an admin can set to zero. We have not found these figures on a public Microsoft page.

1. The assistants agreed on a format, so your instructions now outlast the tool. Gemini adopted skills, joining Claude and ChatGPT, and will accept skill files written elsewhere. ChatGPT tripled the space for custom instructions. Flanagan’s three files and Su’s rulebook are the same idea built by hand: short written documents that tell any assistant how you work. For two years the advice was to learn a tool. This week’s releases reward the teams that wrote down their standards instead, because that writing now moves with them.

2. Agents moved from doing a task to holding a job. Dots, Microsoft’s Autopilot, Perplexity’s Automations, ZoomInfo’s Agent Teams and Google’s monitoring in AI Mode all continue after you close the window. Anthropic’s buying agent shows what that looks like when it works: a narrow job, three defined exits and people editing the brief every week. The practical question has changed from what to ask an assistant to what it may do alone and when it must come back to you.

3. Adoption stopped being the difference. Bain found 6% reporting significant impact, the WFA found 8% with a clear AI search strategy, and Forrester found headcount being cut alongside doubts about accuracy. Nearly everyone has the tools. Bain’s leaders rebuilt workflows, named an owner and ran far more experiments, and that is slower and less visible than buying a licence. The search data makes the same point from outside: Google now summarises your brand on four in five searches for your name, whether or not anyone on your team owns that page.

FINAL CHANCE TO INVEST IN CLAUDE FOR MARKETERS MASTERCLASS

For marketers who want their hours back. The full Marketing Operator System, the Setup Skill that builds your working environment, and 5-Minute Slides, together for the first time.

Now just $150 OR 2 payments of $75

(Individual price $953).

Check Out This Limited Time Offer

Tell me: What’s your biggest takeaway from this week’s news from the world of AI Marketing? Love to hear your thoughts in the comments.

Leave a comment

Thanks for reading the Marketing with AI Weekly Roundup! If you enjoyed this read, the best compliment I could receive would be if you shared it.

Marketing with AI Weekly Roundup: Check Your Brand Across Three AI Engines, Build Ads That Learn, and Make HD Video for Free

Robert Gillespie — Sun, 27 Sep 2026 10:02:10 GMT

Welcome to issue #53 of the Marketing with AI Weekly Roundup. Published every Sunday, it shares the top stories, playbooks, data and quick hits from the world of Marketing with AI over the last 7 days.

Subscribe to receive the week's most important AI marketing stories direct to your inbox.

Subscribe now

FINAL CHANCE TO INVEST IN CLAUDE FOR MARKETERS MASTERCLASS

For marketers who want their hours back. The full Marketing Operator System, the Setup Skill that builds your working environment, and 5-Minute Slides, together for the first time.

Now just $150 OR 2 payments of $75

(Individual price $953).

Check Out This Limited Time Offer

In January, ChatGPT handled 70% of the prompts that Comscore’s US panel typed into AI assistants. By June it handled 50%, with Gemini on 30% and Claude on 11%. If your checks on how AI describes your company still start and finish in ChatGPT, you are now testing half the market, and a separate study published two days later shows the other half cites different sources.

Microsoft rebuilt Copilot around an agent that keeps working when you are not there, and moved agent work onto a usage meter. Google made 1080p AI video free for anyone with a Google account. And Anthropic’s Opus 5.5 brought top-tier performance down in price on the same day OpenAI put two new GPT-6 models into ChatGPT Work.

The Playbook has three things to try. A writer with fifteen years of ghostwriting behind him explains why AI drafts sound generic and what to feed a model instead. A practitioner shows an ads agent that logs every creative like a product. And a video creator runs voiceover, music and a translated version from one Claude conversation.

1. Microsoft rebuilt Copilot around an agent that keeps working, and put agent work on a meter

On 25 September Microsoft redesigned the Copilot app around three areas. Home is the new starting point, where Chat and Cowork, Microsoft’s mode for handing off longer tasks, come together, with Word, Excel and PowerPoint built in so documents are created and edited without leaving Copilot. A daily command centre called Today, pulling together mail, calendar, Teams threads, meetings and tasks, is described as coming soon. Code lets people who do not write code build small dashboards, apps and automations from a plain-language request, running in a sandbox inside the organisation’s own Microsoft tenant. Autopilot is a persistent agent with its own identity, memory, computer and workspace, which carries on with recurring work, follow-ups and scheduling when you are away.

Almost none of this is broadly available yet. Home and Code start rolling out through the Frontier programme, Microsoft’s early-access scheme, in the coming weeks, with Code reaching Frontier at the end of September. Code will be in preview for Microsoft 365 Premium and Pro subscribers later this year. Autopilot enters private preview at the end of the month.

The change that reaches every Microsoft 365 customer is how it is paid for. In a post two days earlier, Microsoft said chat, Copilot inside the Microsoft 365 apps and ready-made agents such as Researcher and Analyst stay on the per-user licence. Cowork and other agent work move to usage-based billing in Copilot Credits, one shared meter across the platform, on the reasoning that long-running jobs vary too much for a flat fee to absorb. Admins can see where the credits go and set limits before the invoice arrives. Microsoft has not published prices.

Why this matters. Two of the biggest assistants merged their chat and agent modes nine days apart, Claude on 16 September and Microsoft now. The recurring jobs used to sell the new Copilot are marketing jobs: pulling campaign data into Excel to work out return on spend, turning results into a PowerPoint recap, building a tracking dashboard nobody in engineering has time for. For the many B2B marketing teams inside Microsoft 365 companies, those are also the jobs that will now draw on a metered budget rather than a licence someone else already paid for.

What to do. Ask IT two questions this week: whether your organisation is in the Frontier programme, and who will own the Copilot Credits budget and its limits. Then pick one report you rebuild by hand every month as the first job to try in Code or Autopilot when it reaches you, and time how long it takes today so you have something to compare against.

2. ChatGPT’s share of AI prompts fell from 70% to 50% in six months, and each assistant cites different sources

Comscore published its Q2 2026 AI Intelligence Report on 22 September, drawn from US opt-in panels of between 500,000 and one million people. Between January and June 2026, ChatGPT’s share of prompt volume across the assistants Comscore tracks fell from 70% to 50%, while Gemini’s rose from 17% to 30% and Claude’s from 2% to 11%. Copilot held at 5%. Total usage kept growing, so ChatGPT did not shrink. Its desktop conversations reached 168 million in June, and its monthly visitors across devices rose to 99 million from 59 million a year earlier. Its rivals grew faster: Gemini’s desktop conversations went from 8 million to 89 million in a year, and Claude’s from 1.1 million in October 2025 to a peak of nearly 40 million in April, before settling at 22.3 million in June.

Two days later, the B2B agency Foundation and the AI search platform AirOps published The State of B2B AI Discovery 2026, which shows why that split matters to anyone selling software or services. They logged 380,000 AI answers and about three million citations across 81 B2B software categories on six surfaces, ChatGPT, Claude, Gemini, Perplexity, Google’s AI Overviews and AI Mode, between 22 June and 24 July. Google’s AI features draw on Google’s live search index, while the standalone assistants use their own retrieval, so what each one cites differs. Social platforms made up about 7% of citations on Google’s surfaces and 0.3% on ChatGPT. Gemini sent 71.6% of its citations to vendors’ own sites, the highest of any engine. Across all six, brand and product pages were the single biggest source, at 68.4% of citations, and the share going to review sites ranged from 7.6% in managed databases to 25.1% in recruiting software.

Why this matters. Most checks on AI visibility, whether done with a paid tool or by someone typing buyer questions into a chat window, start with ChatGPT and often stop there. The Comscore figures say that covers half the prompts, and the AirOps figures say the other half draws on different sources, so a good result in one assistant does not predict the others. Your own product and comparison pages matter more than most content plans assume, and in some categories your review profiles carry a quarter of the weight.

What to do. Take the ten questions a buyer asks when shortlisting in your category. Run each one in ChatGPT, Gemini and Claude, and record for each engine separately whether you appear, what it says about you and which sources it cites. Fix the pages those answers draw on, starting with product, pricing and comparison pages, and treat your G2 or equivalent review profile as a page you own. Two caveats. Comscore measures a US consumer panel’s share of prompts, not business users. And Foundation has a commercial relationship with AirOps, and both sell AI visibility services.

3. Google made 1080p AI video free for anyone with a Google account

On 23 September Google opened AI video generation in Google Vids to anyone with a personal Google or Google Workspace account, at no cost, using the Gemini Omni 1.1 Flash model we covered when it launched at the end of August. On a desktop, go to vids.new and choose “Create AI videos”.

The controls are built for people without editing experience. You can start from a template, upload reference images or adjust the prompt. You can set the length of each clip to the second, which matters when scenes have to match a voiceover. You can extend a scene while keeping characters, style and lighting consistent, and generate new scenes in 1080p or upscale existing AI clips to match. Finished videos download or go straight to YouTube or Google Drive. Every clip generated this way carries Google’s SynthID watermark, embedded invisibly in the frames. Google says advanced features and higher usage limits come with paid Workspace and Google AI plans, and it has not said how much free use you get. Workspace domains may take a few days to see the change.

Why this matters. Short explainer video is one of the formats B2B teams find hardest to resource, because it has usually needed an agency or someone with editing skills. A free tool inside the Google account everyone already has, with scene lengths you can set to a script, removes both of those obstacles for simple explainers.

What to do. Pick one landing page or feature announcement that has no video and make a 60-second explainer for it this week. Write the script first, generate a scene for each line, and use your own screen recordings wherever the product has to be shown accurately. Decide before you publish how you will label AI-generated footage, because viewers cannot see the watermark even though software can detect it.

4. Opus 5.5 brought top-tier performance down in price, and OpenAI moved two GPT-6 models into ChatGPT Work

Anthropic released Claude Opus 5.5 on 22 September, the first model in a Claude 5.5 family, with Sonnet 5.5 and Haiku 5.5 due in the coming weeks. Anthropic says it performs at the level of Fable 5.1, its top generally available model, on most work, and costs about 40% less to run than Opus 5 on typical workloads, through lower prices and fewer tokens used per task. Those are Anthropic’s own measurements. API pricing is $4 per million input tokens and $20 per million output tokens, 20% below Opus 5. For subscribers, Anthropic raised the five-hour usage limits on Pro, Max, Team and seat-based Enterprise plans, and gave subscription users a one-off rate limit reset they can save and use when they choose. It also says the model communicates more naturally, and one customer in the announcement reported answers 40% less verbose without losing accuracy.

The same day, OpenAI released GPT-6 Sol and GPT-6 Luna, faster and cheaper models built on the GPT-6 Astra work we covered on 12 September. They are in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users, and not in the ordinary Chat window. Free and Go users can use Luna in the desktop app, and Enterprise administrators have to switch the new models on. API prices were halved against the previous generation’s promotional rates, to $2 per million input tokens and $10 per million output tokens for Sol.

Why this matters. For most marketers the headline is not a benchmark score. It is that a model Anthropic says matches its best one became cheaper to run while the subscription limits went up, so work that used to hit a usage ceiling part way through the afternoon, such as long documents, research pulls and skills that run several steps, now has more room. On the ChatGPT side, the newest models sit in Work rather than the chat box most people type into, so a team that never opens Work will not see them.

What to do. Rerun the three prompts or skills you used most last month on Opus 5.5 and compare the outputs side by side before relying on anything you concluded about older models. Pay attention to length, since Anthropic is claiming tighter writing, and check whether that suits your brand voice rules. If your team uses ChatGPT, open Work once and check which models your plan and workspace settings give you there.

1. Stop asking AI to write for you, and give it your positions to repeat

Nicolas Cole has spent fifteen years writing and ghostwriting non-fiction online and runs the writing platform Typeshare. He spent seventy minutes on Greg Isenberg’s podcast on 21 September explaining why AI writing reads as generic and what to change, and the episode has passed 128,000 views.

Three tiers of content. Cole sorts everything into commodity, personality and original. Commodity content is ideas nobody owns, repeated by thousands of people in slightly different words. Personality content carries details that are yours: what happened to you, where you work, what you got wrong. Original content is new thinking, which he argues only comes after years of writing about a narrow subject. What people call AI slop is mostly commodity content, which is why it feels anonymous. His quick test is the pronoun. Writing that tells the reader what “you should” do leans towards commodity, while “I did” and “I think” lean towards personality.

Why AI gets your commodity content wrong. Commodity ideas come in opposing pairs. His example is personal finance, where one camp says skip the daily coffee and invest the difference and the other says the coffee is irrelevant. Neither position is original, but you hold one of them and a model does not know which. B2B has the same pairs: gate the content or give it away, keep MQLs or drop them, publish pricing or hide it. Every time a draft lands on the side you disagree with, it reads as not you.

Robots repeat, people write. As a ghostwriter, Cole built clients’ posts from sentences they had already said in books, talks and recorded meetings, which he calls approved language. He wants AI to do the same job at scale, recombining and republishing what you have already written and approved rather than inventing new sentences on your behalf. In his words, “What I want is the robots repeating me.” The time that saves goes on writing the new material only you can produce.

The unmade data set. Everyone using the same models is drawing on the same data. What no model has is your experience, which he calls the unmade data set: formative moments, mistakes, customer conversations, numbers. Writing those down once gives the model something to repeat that nobody else can. He used one sentence about being a top-ranked World of Warcraft player as a teenager in hundreds of pieces over a decade, because people associate ideas with whoever repeats them, and personal detail makes the association stick.

For company accounts. Brand channels, he says, should repeat the company’s chosen commodity positions, while personality content belongs to named people such as the founder and the leadership team.

Voice is smaller than people think. He breaks voice down into word choice and sentence length and structure. When a ghostwriting client said a draft did not sound like him, Cole added two or three sentences citing published research, because that client always argued from evidence. He changed about 1% of the piece and the client was delighted. When an AI draft sounds wrong, look for the one variable that is off before rewriting the whole thing.

He sells Typeshare, which is built around this method, and a writing course, and the episode carries a sponsor.

Do this now. Write down ten positions that are common in your category and mark which side your company takes on each, with one line of reasoning. Add five short stories only your team could tell, each attributed to the person who lived it. Put both wherever your AI tools read from, whether that is a Claude project, a skill or HubSpot’s Context Home, and instruct the model to reuse that material and flag anything it had to make up.

Read more:

2. An ads agent that ships ten creatives a day, logs each one like a product, and asks before it publishes

Cody Schneider, co-founder of Graphed, joined HubSpot’s chief marketing officer Kipp Bodnar on Marketing Against the Grain on 24 September and built several marketing agents live in Claude Code, speaking his instructions rather than typing them.

What he means by an agent. Not, as he puts it, a god in a box running your ad account. An agent is software with a thinking loop, working on a live stream of data and making decisions from it. Most of each process should run as ordinary code, with a model called only where judgement is needed, because that keeps it reliable and cheap.

Why ads need far more creative now. Bodnar’s argument, which Schneider agrees with, is that platform targeting has become good enough that the scarce input is creative. A campaign that once ran five ads now needs fifty or more, and the teams doing best are producing at that volume.

Every creative gets an ID and a record of where it came from. He treats each ad image like a product SKU. The instruction that generated it is stored in a database under its own ID, along with the customer pain point it addresses, the outcome it promises and the source it was based on, such as a competitor’s ad. Once ad performance flows into the same data warehouse, the agent can see which pain points and sources produced results, not just which images won, and uses that to decide what to make next.

Ten a day, not three hundred at once. His team ships about ten new creatives a day and runs two or three learning cycles a week, which comes to around 300 a month. Launching all 300 at once, he says, leaves most accounts stuck in the platform’s learning phase unless the budget is large.

A negative-keyword agent for Google Ads. A model scores every search term that triggered an ad from 1 to 10 for how closely its intent matches the product. Terms below the line become negative keywords and are never paid for again. The model makes that one judgement and ordinary code does the rest.

Read from a warehouse, write to the API. Ad accounts get banned when agents hammer a platform’s API with read requests, he says, not because an agent is involved. His fix is to pull all the data into a warehouse through a pipeline, run every analysis there, and use the platform’s API only to upload ads, keywords and budget changes.

A human gate that trains the system. For teams not ready to let an agent publish, every new creative goes to a person in Slack to accept or reject, and the accepted ones become the reference set for future creative. In regulated industries he writes the words the company can and cannot use, along with the phrasing Meta tends to flag, into the system’s rules.

An SEO agent works the same way. It researches bottom-of-funnel keywords, writes each article from what already ranks on page one, publishes it to the CMS through its API, and then refreshes published articles against live performance data, which he says is the step most teams skip.

His results, which are his. He cited a business with 250 physical locations whose Google Ads cost per lead fell from $70 to $35, and Meta cost per lead from $82 to $15, in four weeks on the same budget. He also cited a start-up whose Google Ads cost per lead fell from $1,100 to $250 over four months with an agent that checks the data daily and rewrites the code running the account. These are client figures with no independent check, and the first business is not a B2B company.

Where he says to start. Give your coding agent, whether Claude Code or Codex, one API key for a tool you already use, stored in a local environment file, and ask it to do a job you would otherwise do by hand, such as cleaning up records in HubSpot. Our addition: start with a read-only key or a sandbox account, because a key with write access lets the agent change live records.

This is HubSpot’s channel, it links to a free HubSpot guide that asks for your details, and Graphed sells the data stack he demonstrates.

Do this now. Start the log even without an agent. For every ad you launch next week, record the pain point, the promised outcome and where the idea came from in one spreadsheet row next to the ad’s ID. After a month, sort by results and you will know which ideas worked, not just which images did.

Read more:

3. Voiceover, music and translated versions from one Claude conversation

On 14 September ElevenLabs expanded its connector for AI assistants so it can create as well as manage. It is an MCP connector, the standard that lets an assistant use another company’s tools. One installation in Claude, ChatGPT or Cursor now gives the assistant ElevenLabs’ voices, transcription in 99 languages, dubbing that keeps the original speaker’s tone, music, sound effects and more than 50 image and video models. You sign in once, there are no API keys to manage, and everything generated lands in your ElevenLabs workspace for editing.

Taylor Jackson, a photographer and video creator who studied marketing, published an eight-minute walkthrough on 24 September. It is sponsored by ElevenLabs, which he says at the start, and his example is a consumer hardware launch, but the workflow transfers directly to B2B video.

The sequence. He plans the campaign in Claude, then has Claude turn the draft script into a shot list before filming. Once the footage exists, he asks for the voiceover in the same conversation, briefed as he would brief a voice actor with a tone and a length, in his case warm, confident and about 15 seconds. Because Claude already holds the campaign context, he finds the first take is often the final one. If nobody can agree what the brand should sound like, he describes a voice, generates samples and sends them round rather than debating it in a meeting. Music comes from a written description instead of a stock library search, and sound effects can be matched to where the audience lives. For a Canadian campaign he produces English and French versions that keep the timing of the original, so the edit does not have to be recut for each language.

Two practical notes from him. Ad variants tend to perform better when the voice resembles the audience, so he generates different voices for different target groups. And commercial rights to what you generate depend on your ElevenLabs plan, so check yours before anything goes into a paid campaign.

To set it up, find ElevenLabs in Claude’s connectors directory, select connect and sign in with an ElevenLabs account. The free plan comes with 10,000 credits, which is enough to test the workflow.

Do this now. Take one product demo or webinar clip you already have and, in a single Claude conversation, produce a new voiceover, a music bed and a version in the language of your biggest non-English market. Decide before publishing how you will tell viewers that a voice is synthetic.

1. B2B buying on LinkedIn starts 124 days before the deal reaches your CRM

Factors.ai, which sells account-based marketing and attribution software, published From Benchmarks to Blueprints: The LinkedIn Revenue Engine on 23 September, and LinkedIn’s marketing blog featured it the next day. It draws on two sets of its customers’ data: more than 850 B2B companies with over $150 million of LinkedIn ad spend, and more than 50,000 closed deals worth over $5 billion from the 100-plus companies that connect their CRM.

Sustained engagement on LinkedIn starts 124 days before a deal is created. Accounts with six or more people engaged before the deal existed won about 17 percentage points more often than accounts with one. Engaging three or more additional people during the active cycle added 16 points, and deals that involved an end user, an influencer and a technical evaluator showed a 22-point lift, the strongest role-based result. Accounts reached by both paid and organic LinkedIn activity had win rates 10.8 points above the baseline. In this data LinkedIn returned 1.6 times its ad spend against 1.18 for Google, and LinkedIn-sourced deals were 34% larger. LinkedIn’s own summary adds that 76% of its ad spend targets director-level roles.

The practical read is that most LinkedIn targeting aims at one senior person while wins go with reaching the wider group, early. Widen job-title targeting to include evaluators and end users, pair paid campaigns with employees posting organically, and judge campaigns over four months rather than one quarter. These are a vendor’s customer figures showing things that occur together, not proof that one causes the other, and the release gives no period for the data.

2. HubSpot: nearly every marketing team uses AI, and one in ten has changed how it works

HubSpot’s State of AI report, previewed on its Marketing Against the Grain site on 23 September, surveyed 6,000 marketing, sales and service professionals. 98% of marketing teams use AI in some form, but only 10% have reached what HubSpot calls a transformational level, meaning AI has materially reshaped their workflows, decisions and results. The uses with the most marketers reporting measurable outcomes are campaign performance insights, at 83%, and website lead capture, at 80%, and 72% of marketers say AI has made their MQL targets easier to hit. ChatGPT is the tool of choice for 57% of go-to-market teams, Canva AI for 35% and Claude for 20%. Trust, privacy and security are the main barriers.

The number to take away is the gap between 98 and 10. Using AI is now standard and is no longer an advantage in itself. HubSpot has not published fieldwork dates or countries, the sample mixes marketing with sales and service roles, and it is HubSpot’s own research.

3. YouTube is now the most-cited platform in AI answers

Digiday brought the data together on 23 September. Meltwater found YouTube was the most-cited platform across eight AI products in August, Claude, ChatGPT, Copilot, AI Mode, AI Overviews, Gemini, Perplexity and Grok, overtaking Reddit. Tinuiti’s Q2 citation report found YouTube’s share of citations in Google’s AI Mode more than quadrupled between January and April 2026, and more than doubled in AI Overviews. Scrunch, tracking sponsored YouTube videos from May to July, found citations fade quickly: the citation rate fell about 43% within a month, from 18.4% to 10.5%.

For B2B teams that treat YouTube as a brand channel, this makes it a search channel as well. Long-form explainers with chapters, descriptive titles and a clear answer early on are what these systems can cite, and the Scrunch figure argues for refreshing videos rather than publishing once. Meltwater’s sample size is not given.

1. Adobe’s tools now work inside Gemini, and Acrobat has joined Adobe’s Claude plugin

On 24 September Adobe brought tools from Photoshop, Lightroom, Adobe Express and Firefly into Google’s Gemini, rolling out globally on every Gemini plan. You connect it under Personal Intelligence, connected apps, and sign in with an Adobe account, which can be a free one. Adobe’s examples are marketing ones: consistent product photos, turning one design into versions for each social channel while keeping the branding, and customising templated campaign assets.

At the same time, Adobe’s plugin for Claude gained Acrobat for the first time, taking it to more than 80 tools across Acrobat, Express, Photoshop, Illustrator, Premiere, Lightroom, InDesign and Adobe Stock, in Claude and Claude Code on desktop, mobile and web. Two new editors let you work by hand inside Claude: one for PDFs, where you can reorder pages, edit text and add comments, and a layer-based one for Express designs, where you can change a single image, line of copy, colour or font without regenerating the whole design. The old Adobe connector updates automatically. The Acrobat tools cover the dull weekly jobs: combining case study PDFs, redacting a customer’s name before a deck goes out, and converting a PDF into an editable format.

2. Draft in Google Docs from a Gemini Notebook

Since 23 September you can type @ in the Gemini side panel or bottom bar in Google Docs, pick one of your Gemini Notebooks, the tool that used to be called NotebookLM, and have Gemini draft from that notebook’s sources with inline citations. It had been in beta since June and is now rolling out generally, on Workspace Business Standard and Plus, Enterprise Standard and Plus, Education Plus, and Google AI Pro and Ultra, but not Business Starter. Nothing is needed from admins if Gemini is on in Drive, and users need Workspace smart features switched on. For a B2B team the obvious use is one notebook per product or campaign holding the research, customer interviews and analyst reports, used as the base for white papers, briefs and proposals, with the citations as a built-in check that the numbers match their sources.

3. G2 put its buyer intent data into ad platforms and its review data into your AI assistant

On 23 September G2 released a set of tools for software vendors. G2 Audiences lets Buyer Intent customers build account-level audiences from companies researching their category, competitors or product on G2, and send them through Bombora into ad platforms including The Trade Desk, DV360 and StackAdapt without exporting spreadsheets. Verified Leads is a self-serve pay-per-lead product through G2’s Software Advice, with each lead qualified on budget, authority, need and timing and delivered with headcount, budget, required features, decision-makers and timeline. Review requests can now go out through Gmail and LinkedIn and run as HubSpot campaigns. And G2’s MCP connector, already available in Claude and ChatGPT, now also runs inside my.G2 for teams that cannot use outside AI tools. Read this alongside the second Top Story: review sites carried up to a quarter of AI citations in some software categories, so your G2 profile is now one of the pages AI assistants read about you.

4. Claude in Slack can now use your own connectors

Claude Tag, Anthropic’s Claude for Slack channels, could previously use only the connectors an admin attached to a channel. Since 24 September it can also use your personal ones, such as your Google Drive, calendar or CRM, when you ask it something in a channel. Claude asks the first time it needs one. You choose review mode, where you see the reply before it posts, or auto mode, where it posts unless it detects something sensitive, and Enterprise admins can require review for everyone. Activity using your connectors is logged under your account. It is rolling out on Team plans now, with Enterprise coming soon. For a marketing team that already works in Slack, it means asking Claude in the campaign channel to compare the plan in your own Drive with the numbers in a shared dashboard, without pasting in either.

5. Perplexity lets you choose how hard its agent works, and opened a skills marketplace

Perplexity’s September changelog, dated 21 September, added Effort Mode to Computer, its agent for multi-step work. A slider running from Light through Standard and High to Ultra picks the model and reasoning depth and spends credits accordingly, available on the web for personal accounts first. Perplexity has since made GPT-6 Sol its Light option and Claude Opus 5.5 its Standard option. A public Skills Marketplace lets anyone browse reusable skills for Computer without signing in, and Enterprise teams can share skills internally with admin approval. There is also a side chat for asking a quick question about a running task without interrupting it. The slider is the practical part, because routine research does not need the most expensive setting and choosing deliberately makes the credits last.

6. ChatGPT Voice can now use your connected apps, and works inside ChatGPT Work

On 23 September OpenAI added plugins to ChatGPT’s Live voice mode on web, iOS and Android, so during a spoken conversation it can use the apps you have connected, such as Gmail, Google Calendar or Slack, under the permissions you already have. Voice also now works inside ChatGPT Work on web and mobile, where you can ask out loud for a document, presentation or spreadsheet, and a task that is unfinished when you hang up carries on in text. Voice in Work needs both Voice and Work access on your plan, and in company workspaces it follows admin settings. The default plugin setting lets ChatGPT go ahead with low-risk actions without asking and checks with you before sensitive ones such as emailing someone, so review that setting before you start dictating tasks on the move.

1. Google Ads added two budget modes to its investment recommendations. The Recommended Investment Strategy tab now offers Holistic mode, which moves budget from weaker campaigns to top performers and adds extra weekly spend, and Growth mode, which adds budget to constrained top performers without cutting anything. Target CPA and target ROAS are adjusted to match, and the forecasts rest on seven-day projections. Both modes end with you spending more, so treat them as proposals to check rather than advice.

2. Search Console now reports searches that start with an image. A new filter splits the Web search type into Text-based and Multimodal, covering Google Lens, Circle to Search, image uploads and Chrome’s “Search this image”. The data was not previously counted, so a rise in totals from 24 September is reporting catching up rather than a ranking gain. There is no query data for image-based searches, only pages. The same split appears in the Generative AI features report.

3. AI Overviews are linking out far more, and Google is testing links that keep people inside Google. Data from Peec AI, an AI visibility vendor, shows the share of AI Overviews containing links to outside websites rising from close to zero to more than 25%, with no sample or period published. In the same week, Search Engine Roundtable reported normal-looking links at the bottom of AI Overviews that open a follow-up question in AI Mode instead of a website. Being cited is still not the same as being visited.

4. Anthropic opened the Claude Marketplace. Launched on 23 September, it brings plugins, connectors, agents, products and consulting partners into one place, and lets organisations spend part of a committed Anthropic budget on partner software. Named partners include Atlassian, Google, Microsoft, Notion, Salesforce and Snowflake, and the marketing agency Power Digital says its Claude use grew from about 12 engineers in January to more than 800 people. On 25 September Anthropic also opened a portal for developers on paid plans to submit plugins to its directory. For marketers, the takeaway is that more of your tools will arrive as a Claude plugin.

5. Gemini added 14 more connected apps. From 23 September you can @-mention apps including Airtable, Linear, monday.com, PandaDoc, Wispr AI, Zoho, Adobe, Picsart, Squarespace and Webflow inside a Gemini chat. According to Google’s availability table, most are limited to US users, English and personal accounts. Adobe is available worldwide, and monday.com works with work accounts.

6. LinkedIn now lets colleagues vouch for each other’s work history. Members can confirm that someone worked or studied with them, provided the person vouching is verified, has been connected for at least a year and uses two-factor authentication. Admins of verified company Pages can remove people falsely claiming to work there, and LinkedIn is testing a setting that makes anyone linking themselves to a Page verify through a work email. Worth checking your own Page’s employee list, since fake employees are a common scam tactic, though Page verification is tied to a Premium Company Page or an advertising relationship.

1. There is no longer one AI answer to optimise for. ChatGPT’s share of prompts fell to half in six months. The engines cite different sources, with social content almost absent from ChatGPT’s citations and prominent in Google’s. YouTube overtook Reddit as the most-cited platform. AI Overviews began linking out more while Google tested links that keep people inside AI Mode. Any one of these would be a footnote. Together they mean visibility in AI answers has become several separate measurements, and a single score or a single engine tells you less than it did in the spring.

2. The tools moved into the assistant, and the bill moved to the task. Adobe’s apps now run inside Gemini and Claude, ElevenLabs’ audio and video tools run inside Claude, G2’s data runs inside Claude and ChatGPT, and Gemini Notebooks run inside Google Docs. At the same time Microsoft put agent work on a meter, and Perplexity lets you set how much each task spends. The assistant is becoming the place where marketing work gets finished rather than drafted, and its cost is shifting from a fixed licence to what you use. That makes it cheaper to try things and harder to forecast what they will cost.

3. When production is nearly free, what you feed the machine is what counts. Google made HD video free, and Opus 5.5 made a top model cheaper to run. Against that, all three practitioners in this week’s Playbook spent their effort on the inputs: Cole on positions and stories no model has, Schneider on a record of which pain points and sources produced results, Jackson on a brief detailed enough that the first take is usable. Factors’ data points the same way from the buyer’s side, with wins going to the teams that reached more of the buying group, earlier. The part left for people is deciding what to say and to whom.

FINAL CHANCE TO INVEST IN CLAUDE FOR MARKETERS MASTERCLASS

For marketers who want their hours back. The full Marketing Operator System, the Setup Skill that builds your working environment, and 5-Minute Slides, together for the first time.

Now just $150 OR 2 payments of $75

(Individual price $953).

Check Out This Limited Time Offer

Tell me: What’s your biggest takeaway from this week’s news from the world of AI Marketing? Love to hear your thoughts in the comments.

Leave a comment

Thanks for reading the Marketing with AI Weekly Roundup! If you enjoyed this read, the best compliment I could receive would be if you shared it.

Marketing with AI Weekly Roundup: Turn Sales Calls into Campaigns, Build Decks in Claude, and Prioritise Google Ads Leads

Robert Gillespie — Sun, 20 Sep 2026 10:03:29 GMT

Welcome to issue #52 of the Marketing with AI Weekly Roundup. Published every Sunday, it shares the top stories, playbooks, data and quick hits from the world of Marketing with AI over the last 7 days.

Subscribe to receive the week's most important AI marketing stories direct to your inbox.

Subscribe now

FINAL CHANCE TO INVEST IN CLAUDE FOR MARKETERS MASTERCLASS

For marketers who want their hours back. The full Marketing Operator System, the Setup Skill that builds your working environment, and 5-Minute Slides, together for the first time.

Now just $150 OR 2 payments of $75

(Individual price $953).

Check Out This Limited Time Offer

The two CRMs most B2B marketing teams run on each announced a campaign agent this week, within a day of each other. HubSpot’s takes a sales call transcript and returns a brief, the assets and the automation to run it. Salesforce’s does much the same and arrives in October.

Read the availability labels before you read the keynotes. Nearly everything HubSpot showed marketers is public beta on Professional and Enterprise and spends credits. Most of what Salesforce showed marketers is dated next month. The part of each launch you can use on Monday is smaller than the announcement, and it is listed below.

Two other changes will reach more desks than either conference. Claude now writes documents and slide decks inside an ordinary conversation and exports them to Word, Google Docs and PowerPoint. And Google is piloting a score for every lead that comes through a Google Ads form, along with a bidding change built for sales cycles that take months to close.

The Playbook is three jobs you can finish in an afternoon: the five Claude skills worth building first, four changes that turn LinkedIn posts into qualified pipeline, and one prompt that audits your website for the trust signals Google and AI search both look for.

1. HubSpot rebuilt Marketing Hub around agents, and most of it is public beta

HubSpot used its UNBOUND conference in Boston on 16 September, the event that was called INBOUND for fifteen years, to relaunch the marketing side of its platform. The organising idea is that you start from a problem HubSpot has already found, then hand the fix to an agent.

Marketing Studio has been rebuilt to open on those problems: your visibility score in AI answers, a campaign segment that is underperforming, leads that came in and were never followed up. You tell Breeze Assistant which one to fix and it calls in one of three new agents. The Campaign Agent plans a connected campaign from a goal, an AI search recommendation or an uploaded document such as a sales call transcript. The Content Agent researches a topic and writes pages and posts designed to be cited in AI answers. The Nurture Agent rewrites the subject line, preview text, body and call to action for each individual recipient at the moment of send, sitting on top of the workflows you already have. You choose what percentage of sends it touches.

Availability is where the detail matters. According to HubSpot’s own knowledge base, Marketing Studio and all three agents are public beta, on Marketing Hub Professional and Enterprise, and they consume HubSpot Credits. The Campaign Agent also needs a Super Admin to switch on generative AI access. HubSpot has not published what each agent action costs in credits, so watch consumption before you scale anything.

What is generally available, on every plan and at no credit cost, is the unglamorous part. Context Home is a single screen showing the business context every HubSpot agent reads from, covering brand voice, products, positioning, ideal customer profile and personas, with a completeness score and a list of the gaps. The CRM also now updates itself from calls, emails and meetings, with reps approving the auto-filled fields before they save.

One addition matters to anyone who works in Claude. HubSpot’s Claude connector can now create campaigns, draft, schedule and publish marketing emails, build landing pages, and read revenue attribution back, all from inside a Claude conversation. It is free on the HubSpot side and needs a paid Claude plan. You have to reauthenticate the connector to get the new permissions.

Why this matters. For a B2B team on HubSpot, the distance between “sales heard an objection on a call” and “there is a campaign addressing it” has been a meeting, a brief and a fortnight. The Campaign Agent is a direct attempt to collapse that. But every one of these agents writes from whatever is in Context Home, so a half-empty context page produces generic campaigns faster.

What to do. Open Context Home, read the score, and fill in brand voice and ideal customer profile first. That hour is worth more than any agent in the release. Then give the Campaign Agent one real sales call transcript and compare its brief with the one your team would have written. If you use the Claude connector, reauthenticate it and set its write tools to “Requires approval” before anything else, so nothing publishes without you seeing it.

A caveat on HubSpot’s headline numbers. The claim that customers generate 3.6 times more marketing qualified leads and win 3.2 times more deals compares Professional and Enterprise customers “using AI with high quality context” against customers not using AI at all. That is two different groups of companies, measured by HubSpot, and it is not a before and after. A more useful number came from a HubSpot partner. Perry Nalevka, chief executive of the B2B consultancy Penguin Strategies, published his own firm’s results from HubSpot’s AI search tracking tool on 16 September. Across 25 tracked prompts over 30 days, the AI answers cited 9,288 sources, and 21 of them were on his firm’s own domain, which is about 0.2%. The tool put the firm’s share of voice at 0.06%. That is an honest picture of where most B2B brands are starting from.

2. Salesforce gave marketers dates at Dreamforce, and the Claude plugin went to open beta

On 29 August we reported that Claudeforce, the Salesforce plugin for Claude, shipped with 37 skills that were all sales skills, and that Salesforce’s own product page had marketing stamped “coming soon” with no date. Dreamforce ran from 15 to 17 September in San Francisco and answered part of that.

Salesforce in Claude is now in open beta. Salesforce says it is available to all customers, requested through AppExchange with your org ID. Coverage of the keynote described it being installed into a Claude Enterprise account, so check which Claude plan it needs before you plan around it. It is still 37 sales skills covering prospecting through to pipeline hygiene. Skills for marketing, service and commerce are described as planned for the near future, again without a date. Salesforce says pricing and packaging may change and availability varies by region.

What marketers did get is on Salesforce’s own release page, with labels. Available now: Palmata, a tool that reports how your brand appears in AI-generated answers, compares you with competitors on common prompts, and suggests content edits with an estimate of how much each would help. Also available now: Piper, an agent that works your website and inbox to engage, qualify and convert inbound leads and book meetings. Arriving in October 2026: a Campaign Agent that creates, runs and optimises campaigns inside the goals and guardrails you set, an Account Discovery Agent that finds high-potential accounts and buying groups, a Budget Optimization Agent that moves spend across channels towards a goal, and the ability for any AI assistant to work on Marketing Cloud journeys and messaging in plain language.

Why this matters. Put this next to the HubSpot story and the pattern is plain. Both vendors now sell the same four things: a campaign agent, an AI search visibility tool, an inbound agent and a layer that manages the data underneath. If you own either CRM, the question has moved from whether to try marketing agents to which of the ones you already pay for is switched on.

What to do. If you are on Salesforce, run a Palmata report on your category now so you have a baseline before anyone changes content, and ask your account team whether it is included in your edition. If inbound is a meaningful share of your pipeline, measure your current response time to a web enquiry this week. That number is the whole business case for Piper, and Salesforce’s own proof points are customer stories it chose: Asana reporting four times the conversation volume, and Siemens saying it now engages every inbound lead. Do not plan a marketing workflow around Claudeforce until marketing skills have a date.

3. Claude now writes the document and the deck inside the chat

On 16 September Anthropic launched Claude Docs and Claude Slides, both in beta on paid plans, and made Claude Design available inside any conversation.

Ask for a document and Claude asks a few clarifying questions, drafts it on screen, and leaves comments explaining the choices it made. You edit it directly or ask for another pass. Colleagues can be invited to view or edit, and the document exports to Word, Google Docs, PDF or Markdown. Ask for a presentation and Claude drafts the slides. You can edit them directly, present straight from Claude, or download as PowerPoint or PDF. A finished document can be turned into a deck. Everything made in Docs, Slides or Design lives at a single shareable link that opens on a phone.

Docs is switched on by default for Pro, Max and Team plans. On Enterprise it is off until an owner enables it under Organization settings, Artifacts, and it can be switched on for specific groups only. It is not yet available to organisations using customer-managed encryption keys, zero data retention or HIPAA-ready configurations, and it is not on the Free plan. You can start a document by asking in any conversation, by typing /docs, or from a template in the Artifacts tab.

Why this matters. A large share of a B2B marketer’s week is documents that accompany the actual work: the campaign brief, the monthly report, the one-page summary for sales, the deck for the quarterly review. Until now Claude produced those as files you downloaded and finished elsewhere. A shared, commentable, living document changes where the first draft and the review both happen, and the PowerPoint and Google Docs exports mean it still ends up in the format your company expects.

What to do. Take the one deck you rebuild every month, give Claude last month’s version and this month’s numbers, and ask for it as Slides. Export to PowerPoint and check it against your brand template, because that is where a beta will show its limits. If you are on Enterprise and cannot see Docs, it has not been enabled for you, so ask your admin rather than assuming it has not rolled out.

4. Google is piloting a score for every lead, and a bidding change for long B2B sales cycles

Last week’s second story was Google rebuilding how first-party data flows into its ad platforms. This week Google’s own product managers explained what that plumbing is for, in an episode of the Ads Decoded podcast recorded live in New York with Ads Liaison Ginny Marvin, and named two things B2B advertisers have not had before.

Lead intent scoring is in pilot. It rates each lead that comes through Google Ads as high, medium or low intent at the moment the form is filled in, using what the product manager described as Google-side signals from signed-in account data. The important detail, which Ginny Marvin stopped the conversation to make clear, is that the score is not used in bidding. She called it a metric to use inside your organisation “to prioritize where you spend your time”.

Journey-aware bidding is in beta, in Search campaigns using Target CPA. It changes how Smart Bidding learns when closed deals are too few, or arrive too late, to train on. That describes most B2B accounts, where a deal closes months after the click and there are only a handful of them. There is nothing to switch on, but there is a condition, and Google’s blog post about the episode leaves it out. You only benefit if you are already importing a later-stage conversion, such as a qualified lead or a closed deal, through Data Manager and bidding towards that, while measuring the form fill without bidding to it. The system then treats the form fill as the start of a journey towards the conversion you care about. In the product manager’s words, “you have to be bidding to something lower funnel”. An account that bids to the form fill gets nothing from this.

The rest of the episode is a sensible order of work, which the presenters call crawl, walk, run. Upgrade your tag first, including hosting it on your own domain through Google tag gateway to recover conversions that browsers block. Then switch on enhanced conversions. Then connect your CRM once through Data Manager so that qualified and closed-won stages flow back to Google.

Why this matters. Every B2B paid search account has the same structural problem. The platform optimises towards form fills because that is the only signal it receives quickly, and form fills are a poor proxy for revenue. Both of these features are attempts to fix that from Google’s side, one by ranking the leads and the other by teaching the bidding system to be patient.

What to do. Google did not say who can join the scoring pilot, even when asked directly on the episode, so ask your Google representative. For journey-aware bidding, check how your Search campaigns are set up today. If they bid to the form fill, the first job is to import a qualified or closed stage from your CRM and move the bidding to that. And if you do get a lead score, test it before you trust it. Compare Google’s high scores against the leads your sales team accepted over the same period, and decide in advance what happens to the leads scored low. It is a number produced by the company that sells you the clicks, from data you cannot see, and your CRM is the only place it can be checked.

Read more:

1. The five Claude skills to build first, and the six mistakes that stop them working

Last week’s first Playbook item covered how to build a Claude skill properly. This one, from Carl on the HubSpot Marketing channel on 14 September, answers the next question, which is which skills a marketing team should build first. It is a 14 minute beginner walkthrough and needs no code. It is also the third skills item we have run in five issues, which reflects where practitioners are spending their time.

His definition is the most useful one we have seen for explaining skills to a colleague: “Think of a project like a briefing document. It holds context. A skill is the actual job description. It holds the process.”

The five, each with one job. A brand voice writer, holding your voice rules, banned words and phrases, audience profile, spelling conventions and three to five examples of your best writing. A social content producer, holding hook formulas, script structure and platform-specific rules, which turns one idea into a hook, a spoken script with line breaks, B-roll directions and on-screen text. An SEO blog writer, holding heading structure, your internal link map and formatting rules for AI search, which returns search intent analysis, meta title and description, a direct answer near the top, question-led subheadings and an FAQ. An email writer, holding subject line formulas, email structure, segmentation notes and link placement rules. And a reporting analyst, holding your KPIs, report structure and how to flag wins and risks.

Build one in three steps. First, go to Settings, Capabilities and switch on code execution and file creation, which he calls the one setting most people miss. Skills then live under Settings, Skills, on Pro, Max, Team and Enterprise plans. Second, open a new chat and say: “I want to build a skill that applies my brand voice to everything I write. Ask me questions that will help you create this for me.” Answer the questions, review the file it writes, push back where it is wrong, and save it. Third, test the trigger by asking for the task naturally, “the way you would on a busy Tuesday”. If the skill does not load, the description is too vague. Add directive language such as “whenever I ask for social content” and test again.

The detail worth stealing from the email skill. It returns three subject lines, and labels each one with the approach it takes: problem recognition, specific outcome, practical question. You are choosing a strategy, not picking whichever line sounds best.

The one that saves the most time. Paste raw numbers into the reporting analyst and it calculates every period-over-period change, names the three biggest wins, the three main risks and three recommended actions. Because it pairs with Claude’s built-in Excel and Word skills, the output is an Excel workbook with a dashboard and a branded Word summary ready for leadership. His words: “The report you used to build manually every Monday morning now takes one prompt.”

The six mistakes. A vague description, which he says kills more skills than anything else. One mega-skill instead of one skill per job. Treating a skill like a prompt, when a prompt is a request and a skill is a standard. Never updating it, when you can select a skill, choose edit with Claude, paste recent work and say “Update the skill to match this writing more closely.” Installing third-party skills without reading them, which matters because skills can contain code. And expecting skills to replace taste: “If your standards are low, skills will just help you hit low standards faster.”

Skills stack without being told to. Ask for a newsletter about a new offer and the brand voice writer and the email writer both load.

This is a HubSpot channel, the templates sit behind a free guide that asks for your details, and the statistic he opens with, that 52% of marketers say AI is making content less effective because it is too easy to produce generic output, is from HubSpot’s own State of Marketing report.

Do this now. Start with the task you re-explain to Claude most often. For most marketers that is brand voice. Build that one skill, test the trigger, and use it for a week before building a second.

Read more:

2. Four changes that turn LinkedIn posts into qualified pipeline, now that anyone can write a passable post

Tommy Clark runs Compound, an agency that writes LinkedIn content for B2B technology founders, and this is his first appearance here. His 13 September video starts from a problem he had himself: the audience grew, and too few of the people arriving could ever buy.

His explanation for why the old approach stopped working is the part that belongs in this newsletter. Two or three years ago, he says, “consistency was the moat”. Simply showing up three times a week was rare enough to win. Then executive content became standard for funded B2B companies, and AI removed the cost of producing it: “anyone with a Claude subscription can create a passable post in about 20 seconds.”

Change one: one post, one person. Every post is written for a single named individual, not a persona. He borrows the BuzzFeed rule of writing for “Debbie at her desk”, a bored office worker between meetings. If you sell to finance directors, decide exactly which finance director, what worries them and what they read. Then decide what you want that person to do. An infographic of a process is built to be saved. A post that says what your buyer believes but would not say under their own name is built to be publicly agreed with. He knows which one he is writing before he posts. The common failure he sees is founders posting office moves, new hires and company milestones to an audience of other founders who will never buy.

Change two: build a moat AI cannot copy. He names three. A narrative moat means pairing every opinion with a story from your own experience, ideally one with conflict in it. AI can copy the format and the style and cannot copy what happened to you. A data moat means publishing numbers only you have, as specific as you can make them. His example is a founder breaking down his company’s Claude spend by team. A physical moat means real photos and video from real places. Across Compound’s client accounts last month, posts with real-life photos got 27% more impressions. That is his own analysis with no sample size given, so treat it as a practitioner’s observation. His rule is that you never leave a conference, an offsite or a new hire’s first day without a photo.

Change three: repeat yourself deliberately. Founders fear being seen to say the same thing twice. He argues recall comes from repetition, that most posts reach a fraction of your audience anyway, and that anyone who sees every post is probably a buyer. When a format earns a disproportionate number of saves or demos, run it weekly until it stops working.

Change four: do not wait for the algorithm. Inbound from content is, in his word, “spiky”: a good week, a quiet week, then nothing. So he puts the content in front of the right people on purpose. He builds a lead list in Apollo, sends 20 connection requests every weekday through a tool called HeyReach, which also withdraws requests nobody accepts, and relies on each accepted request becoming a mutual follow who then sees the posts. He comments on posts by his buyers, by the people his buyers follow, and by his competitors, on the reasoning that people choosing an agency read other agencies’ comment sections.

Two cautions of ours on that last step. Automating connection requests sits against LinkedIn’s user agreement, and 20 a day is his recommendation, not a limit LinkedIn publishes, so the risk to the account is yours. And if you are in the UK or EU, building lists for outreach brings data protection and electronic marketing rules with it that a US agency founder does not discuss. The manual version, a considered handful of requests a day to named buyers, carries neither problem.

He sells an agency service, offers a free guide in exchange for your details, and discloses an affiliate link for one of the tools he mentions.

Do this now. Take your last ten posts and write next to each one the single person it was for and what you wanted them to do. If you cannot answer for most of them, that is the fix, and it costs nothing.

Read more:

3. One prompt that audits your website for trust signals, and how far to let AI fix what it finds

Abhishek Sharma runs a small WordPress and SEO channel called Nerd Up, and says he has audited several hundred sites live on it. On 14 September he showed a workflow in two halves. The first half works on any website and is worth doing this week. The second half needs care.

The subject is what Google calls E-E-A-T: experience, expertise, authoritativeness and trustworthiness, the qualities its guidelines look for in a page. His argument is that these matter as much for being cited in AI answers as for ranking in ordinary search.

The audit prompt, in full. “Act as an experienced SEO auditor using Google’s official guidelines on helpful, reliable, people-first content and E-E-A-T. Audit the publicly accessible pages of [your site URL] for experience, expertise, authoritativeness and trustworthiness. Do not give a generic checklist or invent problems. Show the most important findings with the affected page, the evidence you found, why it matters in beginner-friendly language, and the exact improvement you recommend.”

He ran it in Google’s Antigravity agent tool using Gemini 3.8 Flash, which took 17 minutes. He chose Gemini because the task concerns Google’s own guidelines and he preferred its answers in a side by side test, while conceding it is not the most capable model available. Any assistant that can browse your site will run the prompt.

What it found on his own site is a useful checklist for a B2B site. Demo template pages with placeholder text that Google had indexed. A $1,500 service sold with a money-back guarantee and no refund policy or terms of service anywhere. No registered business details, with a recommendation to add organisation schema covering legal name, address and contact point. Affiliate links not labelled as sponsored. No author byline and incomplete author schema, which he calls the most common trust problem on the sites he reviews. High-priced service pages with no case studies or verifiable proof. An About page promising free help on a site that now sells services. Two near-identical posts competing with each other, with a recommendation to merge them.

The part to copy from the second half is the control he kept. Before letting the model change anything, he had it write a plan for all eight findings, then edited the plan, removing one point and telling it to leave his email address alone. He backed the site up twice, once with a plugin and once at hosting level, and says twice that nobody should attempt this without a backup. The model then made the changes through WordPress in about a minute using an application password, a separate credential you create under Users, Profile and can delete afterwards.

The part not to copy. The model initially refused to touch a live site, and he talked it past that refusal. He gave an AI tool write access to a production website. And he has no before-and-after traffic or ranking data, saying only that he is “definitely sure this makes more trust”. If you want the automated half, run it on a staging copy first.

One line in his favour. When the audit told him his service pages lacked proof, he said he would never let AI write reviews or case studies and would ask real clients instead. He also says AI should not write whole articles for the site, only short positioning lines.

Do this now. Run the audit prompt on your own site and sort the findings into those marketing can fix alone, such as author bylines, the About page and case studies, and those that need a developer. Most B2B sites will find at least three in the first group.

Read more:

1. Nearly every marketing organisation uses AI, and four in ten can prove it pays

Alexander Group, a go-to-market consultancy, published its 2026 Revenue-Ready Marketing Organization research on 15 September, drawn from executives at more than 300 companies across 11 industries including technology, manufacturing, business services and financial services.

96% of marketing organisations have adopted AI. 41% can prove a return on it. The gap between the two is the finding. The supporting numbers are about the machinery underneath: 24% describe their demand generation engine as mature, 26% generate more than 40% of sales pipeline from marketing, and while 56% use advanced attribution models, only 30% trust the results across both marketing and sales.

The practical read is that the constraint on proving AI’s value is the measurement you already had. If sales does not trust your attribution, no AI result reported through it will be believed either. Alexander Group sells consulting on exactly this, and the release does not give fieldwork dates.

2. Fast-growing B2B companies check a lead before it reaches the CRM

Integrate and Demand Metric surveyed 245 marketing, revenue operations and commercial operations leaders at B2B companies, all CRM users, and split them by revenue growth. The fieldwork was in the first quarter of 2026, so the data is about six months old.

79% of high-growth organisations automatically validate a lead before it enters the CRM, against 44% of lower-growth ones. They are more than twice as likely to say at least three quarters of their marketing data is clean, governed and ready for AI to use, at 24% against 10%. And they are more than three times as likely to see sales accept 80% or more of the leads marketing passes over, at 31% against 9%.

Read that alongside this week’s first story. HubSpot’s agents, Salesforce’s agents and Google’s bidding all act on whatever is in the record. Validation at the point of entry is cheap, unexciting, and it is the step the faster-growing companies have in common. Integrate sells lead validation software, and the survey shows these things occur together, not that one causes the other.

3. What B2B buyers do with AI: compare you side by side with your competitors

Carbon, a B2B agency, published the full tables from a survey of 189 B2B decision-makers on 16 September. It is a small sample, but it asks a more useful question than most: what buyers use AI for when choosing a provider.

57% use it to compare providers side by side. 48% use it to summarise vendor information, 45% to get background on a potential partner, 38% to generate shortlist criteria, 36% to identify risks or gaps, and 35% to draft questions to put to vendors. Only 14% do not use AI in the process at all.

That list is a content brief. If the most common job is a side by side comparison, the question is whether an AI assistant can find a clear, current, factual statement of what you do, who for, what it costs and how you differ, on your own site and on the third-party sites it trusts. If 35% are asking AI to write their questions for you, your sales team should ask it the same thing and prepare the answers.

1. Claude’s chat and Cowork modes are merging into one

Alongside Docs and Slides, Anthropic is removing the choice between a quick chat and a longer task handed to Cowork. You describe what you need and Claude decides how to handle it, including multi-step work that keeps running after you close your laptop and can be checked from your phone. Existing chats, projects, connectors and skills carry over. Claude Code is unaffected.

The rollout is staged, starting with Pro and Max over the coming weeks, even for accounts on the same plan. Team and Free follow, and Enterprise admins get at least 30 days’ notice. The way to tell whether you have it: if your message box still offers “Chat” and “Cowork”, you do not yet. For marketers the practical gain is that skills and connectors built for one mode now work everywhere, so this week’s first Playbook item applies wherever you type.

2. HubSpot agents can now be started by things that happen outside HubSpot

Until this week, an agent built in HubSpot’s Agent Builder could only be triggered by an event inside HubSpot. It can now be started by a Slack message, an update to a Google Sheet, a webhook from another tool, or a schedule. HubSpot said plainly that this is aimed at customers who run that kind of automation today in tools such as n8n or Make. There is also a one-click conversion of an existing workflow into an agent-driven one, with a simulation mode that shows what the agent would do and estimates the credit cost without writing anything to the CRM.

It is on Professional and Enterprise, and the “run an agent” workflow step is capped at 500 executions a day. Sources disagree on whether Agent Builder has left public beta, so treat it as rolling out. The useful exercise is to list every automation you run elsewhere only because HubSpot could not start the job. That list is now a consolidation plan.

Read more: https://www.hubspot.com/spotlight

3. iOS 27 changed how Apple Mail finds and handles your emails

Apple released iOS 27 on 14 September. Three changes touch email marketing. Mail search has a new ranking system that orders results by relevance, where it used to lean on recency. Suggestions in Mail can now take actions based on what an email contains, including through other apps. And a feature called Call Context recognises when someone is on the phone with a business and brings up details from that business’s emails, such as a confirmation code or booking reference. The Apple Intelligence features need an iPhone 15 Pro or later.

Apple’s examples are consumer ones, but the B2B reading is direct. Webinar confirmations, event tickets, renewal notices and invoices are all emails people search for later and act on. If the date, the joining link or the reference number exists only inside an image, the software reading the email on your subscriber’s behalf has less to work with. This is the same point as last week’s Playbook item on writing the first 150 characters as real text, with a second large mailbox provider now behaving the same way.

1. OpenAI is retiring custom GPTs.

Announced on 11 September, the plan moves them to plugins, where a GPT’s instructions become a skill and its knowledge files carry across, but custom actions that call outside tools do not transfer. On the timeline OpenAI has published for Enterprise workspaces, custom GPTs stop running on 11 December 2026, and it says other plans are expected to follow the same timeline.

2. Google is testing text-link ads inside AI Mode answers.

They look like the ordinary links in an AI Mode response, with a “Sponsored” label above the answer. It is a test, and it follows last week’s confirmation that exact and phrase match keywords can now serve ads in AI Mode.

3. Google is piloting payments to publishers whose content shapes its AI answers.

The “AI contribution pilot” runs through Search Console, covers the Gemini app, AI Overviews and AI Mode, and pays only when content “contributes significantly” while an answer is being generated. Participants see a monthly earnings figure and nothing about which pages earned it. Google confirmed it to Digiday as an early-stage pilot with at least dozens of mostly small and mid-sized publishers.

4. Salesforce has renamed its clouds back.

A year after rebranding its applications with the Agentforce prefix, Salesforce’s website once again calls them Sales Cloud, Service Cloud, Marketing Cloud and Commerce Cloud, with no formal announcement. Worth knowing before you search the help documentation for a product name that has changed twice.

5. HubSpot retired “Breeze” as the umbrella name for its AI.

Breeze Assistant keeps the name and everything else drops the prefix. The assistant itself is now on every HubSpot subscription including Free, though the agents it calls spend credits.

6. Meta launched its Creator Marketing Hub globally.

It adds an AI assistant for finding creators, and Partnership Ads can now be created and managed through Meta’s ads connector from whichever AI assistant you use. Instagram Live video ads become generally available on 29 September.

1. Your CRM moved into the assistant, and most of the agents arrived as betas. HubSpot’s connector can now build and publish a campaign from inside Claude. Salesforce’s plugin for Claude opened to every Enterprise customer. Claude itself started producing the documents and decks that surround the work. At the same time, almost everything announced specifically for marketers this week carried a label: public beta, pilot, or October. The direction is settled and the dates are not. The sensible response is to do the preparation that every one of these tools shares, which is a written, current statement of your brand voice, customers and positioning, and to stop short of redesigning a process around a beta.

2. Adoption is universal and proof is not. Alexander Group found 96% of marketing organisations using AI and 41% able to show a return. HubSpot’s headline result compares customers who use AI with customers who do not. Google’s new lead score is, by Google’s own account, not connected to bidding. Integrate’s data says the companies growing fastest are the ones validating a lead before it is stored. Every one of those points to the same place. The teams that will be able to prove AI worked are the ones whose measurement was trusted before AI arrived.

3. AI made a passable version of everything free, so the value moved to what it cannot copy. Tommy Clark’s three moats are a story that happened to you, numbers only you hold, and a photograph of somewhere you were. The sixth mistake in the Claude skills walkthrough is expecting a skill to supply taste. The E-E-A-T audit’s author refused to let AI write his case studies. And Carbon’s buyers are using AI to line you up against competitors, which rewards whoever has specific, checkable facts on the page. Four different sources in one week arrived at the same instruction: automate the production, and spend the time you save on the proof.

QUICK REMINDER:

CELEBRATING 50+ EDITIONS OF THE MARKETING WITH AI WEEKLY ROUNDUP

For marketers who want their hours back. The full Marketing Operator System, the Setup Skill that builds your working environment, and 5-Minute Slides, together for the first time.

And to celebrate 50+ editions of this newsletter, it’s now just $150 or 2 x $75 (Individual price $953).

Check Out This Limited Time Offer

Tell me: What’s your biggest takeaway from this week’s news from the world of AI Marketing? Love to hear your thoughts in the comments.

Leave a comment

Thanks for reading the Marketing with AI Weekly Roundup! If you enjoyed this read, the best compliment I could receive would be if you shared it.

Marketing with AI Weekly Roundup: Save Time with Claude Skills, Turn LinkedIn Content into Leads, and Know Where You Can Use GPT-6

Robert Gillespie — Sun, 13 Sep 2026 10:02:47 GMT

Welcome to issue #51 of the Marketing with AI Weekly Roundup. Published every Sunday, it shares the top stories, playbooks, data and quick hits from the world of Marketing with AI over the last 7 days.

Subscribe to receive the week's most important AI marketing stories direct to your inbox.

Subscribe now

CELEBRATING OVER 50 EDITIONS OF THE MARKETING WITH AI WEEKLY ROUNDUP

For marketers who want their hours back. The full Marketing Operator System, the Setup Skill that builds your working environment, and 5-Minute Slides, together for the first time.

And to celebrate, it’s now just $150 OR 2 payments of $75

(Individual price $953).

Check Out This Limited Time Offer

Three of this week’s four biggest stories have the same shape. Something shipped, and then whether you can actually use it turned out to be the story.

GPT-6 Astra arrived in ChatGPT, but not in the window most marketers work in. Google released a metric that scores the value of your first-party data, calculated by the company that benefits when you hand over more of it. And People Also Ask, one of the last units on the results page still sending clicks to websites, is now almost entirely an answer box.

Also inside: a two and a half hour course on building Claude skills that earns the time, a B2B marketer who pulled more than a thousand leads off LinkedIn this year and explains exactly how, and why the first 150 characters of your next email matter more than its design.

1. GPT-6 Astra landed, but not where the announcement said it would

When OpenAI launched Astra on 3 September, the post said it would reach “all ChatGPT Plus, Pro, Business, and Enterprise users”. OpenAI’s own Help Centre tables say something narrower.

GPT-6 Pro, the version powered by Astra, appears in the ordinary chat window only for Pro subscribers and the upper Business and Enterprise tiers. Plus plans get Astra inside ChatGPT Work and Codex, not in the normal chat box. Usage is limited, with paid credits after that. Business Standard gets 15 Pro messages a month and Business Premium gets 50 a week, and those allowances are shared with GPT-5.6 Sol Pro. Codex needs CLI version 0.153.0 or newer. ChatGPT Work itself is still rolling out, and Enterprise and Education accounts get a two week preview with Work switched off by default.

Alongside the availability change, the model is now pitched at work rather than at benchmarks. The capability that matters for marketing teams is that it operates software through the interface, so it can drive tools that have no API at all, and it produces finished documents, spreadsheets and presentations across multi-step jobs.

Why this matters. Most individual marketers are on Plus. That means the model you have read about all week is not in the window you actually use, and the surface it does appear in, ChatGPT Work, may not be switched on at your company. The capability is real. The access is the catch.

What to do. Before you build any workflow on Astra, check which surface it is in on your own plan, and whether an admin has enabled ChatGPT Work. Then pick the job that has been blocked because a tool you rely on has no integration, because operating the interface directly is the thing that is new here.

One caveat on the numbers. OpenAI says Astra produces unintended computer use outcomes 89% less often than GPT-5.6 Sol and 74.7% less often than Claude Fable 5.1. Both figures come from OpenAI’s own internal benchmark.

2. Google rebuilt its first-party data plumbing, and scored your data with a number nobody can audit

On 10 September Google shipped a block of measurement changes, all generally available now.

Data Manager now sits inside Google Analytics and Display & Video 360, not just Google Ads. Enhanced conversions reach both. The Data Manager API has been standardised on the IAB Tech Lab’s ECAPI specification, so the same integration work now travels between platforms. A new Data Strength Uplift metric in Google Ads claims to calculate the additional conversions your first-party data setup recovered, shown over a rolling 14 day window.

The more useful part is Meridian, Google’s free open-source marketing mix model. It gains agentic data-quality auditing, it can now model brand signals including Branded Google Query Volume, and GeoX came out of beta on 9 September and is available globally for geo-based incrementality experiments.

Why this matters for B2B specifically. Offline conversion import is how B2B gets closed-won revenue back into bidding, and this is the plumbing that carries it. If your CRM data is not flowing through Data Manager, the bidding is optimising against form fills rather than deals.

Now the part to read carefully. Every uplift figure Google published is its own internal data, with no third-party verification, and each one rests on a different sample:

26% higher incremental return on ad spend, for advertisers connecting offline and app data. The footnote covers Search campaigns bidding to conversion value, April 2025 to April 2026.

11% more Search conversions from enhanced conversions. That figure rests on a 14 day window, 1 to 14 January 2026.

14% uplift, which belongs to Google tag gateway rather than to the new metric, and which the footnote limits to the finance sector, comparing the second half of 2024 with the first half of 2025.

Over 20% uplift for Demand Gen campaigns, resting on another 14 day window, 3 to 17 June 2026.

A metric that scores how many conversions your data recovered, published by the platform that gains when you share more data, is one you cannot check independently. Two weeks is also not a season.

What to do. Use the plumbing, because Data Manager in Analytics and DV360 is real capability and Meridian GeoX is a free incrementality test. Then run a GeoX experiment, because a result you produced yourself is the only measurement number you fully own.

3. People Also Ask is now 97% AI-generated answers

AlsoAsked analysed 19.2 million English-language queries across 2026. AI Overviews appeared inside 86% of People Also Ask boxes in August, and 97% in the first week of September. Fourteen months ago the figure was 12%.

A second tool, Allintitle, puts it at 100%, but published no method, so 97% is the defensible number to quote.

Why this matters. People Also Ask was one of the last large organic units on the results page that still sent clicks to websites. Winning it used to mean traffic. Now it mostly means your answer is being read inside Google.

What to do. Pull the pages that earn their impressions from People Also Ask, and compare impressions against clicks across July, August and September. Then remove PAA visibility from any traffic forecast you are presenting this quarter, because it has stopped behaving like a traffic source. If the content is good, the work is to become the cited source rather than the ranked link.

Both AlsoAsked and Allintitle sell People Also Ask tools, which is worth knowing when reading their numbers.

4. Two HubSpot executives on why the B2B CMO job is being broken up

On 8 September, HubSpot CMO Kipp Bodnar and SVP of Marketing Kieran Flanagan used their podcast to describe something they are watching happen across B2B. Flanagan’s claim is that the CMO role in B2B “has dropped by about 10 percentage points”. He cites no source for that figure, so treat it as their observation rather than as data.

The pattern they describe is specific. Brand and product marketing move into the product organisation. Growth and demand generation move into sales. The single role that held both disappears.

Their diagnosis is blunt: “Marketers aren’t good enough. The last 20 years of marketing was all about I need to be an extreme specialist.” The paradox they land on is that the role is being dismantled because marketing got more important, not less, and the specialists the last two decades produced are not the generalists this moment asks for.

The part you can act on. Most companies rank a marketing leadership hire in this order: experience, then skills, then hiring ability, then mindset. They argue the order should be reversed, to mindset, hiring, skills, experience. Their reasoning is that mindset and the ability to attract good people compound, while specific platform experience depreciates faster than it used to.

The rest of their advice is practical. Doing nothing beats doing the wrong thing. Never make a hiring decision under artificial time pressure. Keep a warm list of people you would hire one day and check in with them quarterly. And when you cannot get the person who has done it all before, hire a promising marketer and pay an experienced CMO to mentor them across 18 to 24 months, which costs less than the hire you could not make.

Flanagan’s closing line is the fair summary: “I don’t think the role is any less important. I just think there’s a mismatch in the talent pipeline right now.”

Both speakers are HubSpot executives, and they promote HubSpot’s UNBOUND conference and their own book in the episode.

Read more:

1. How to build a Claude skill properly, from someone who has built 450 of them

Eliot Prince published a two and a half hour course on Claude Skills on 10 September. He says he has built more than 450 skills for his own business, for 3,000 plus students and for consulting clients. It is long, and it is the most complete treatment of this anyone has published.

The anatomy, three parts. The description, which is what the skill does plus the trigger phrases that fire it. The skill.md file, which holds the step by step instructions. And reference files, which Claude fetches only when it needs them, holding templates, checklists, style guides and worked examples. His rule for what goes where: decisions, order of operations and what good looks like belong in skill.md, while long worked examples, the full style guide and any schemas belong in reference files.

The mistake almost everyone makes first. He points at one of his own trigger phrases, “write up this transcript”, and says he would delete it. He does many different things with transcripts and does not want a skill firing by accident. If you work across more than one brand, put the brand name in the trigger phrase.

Two kinds of skill, and they need different care. Deterministic skills produce the same templated output every time, like an audit, an invoice or a monthly report. They fail loudly, so they are the easiest to build and trust. Non-deterministic skills carry judgement, taste and voice, like a brand voice skill or a thinking framework. They have no single correct output, and they fail silently, which is why they need reviewing long after you think they are finished.

Five levels, which are not a ranking. A wrapped prompt that needs nothing from you. An ask, which needs your input, such as a transcript. A connect, which reaches into Gmail, Drive, a CMS or a SaaS tool to pull or push. A chain, where skills hand off to each other like a production line. And an orchestrator sitting on top, which knows what each skill does and routes a plain-language goal to the right sequence.

His own warning about those levels is the best line in the course: “The levels aren’t a scoreboard. A level one skill that frees up two hours a week of work for me is a massive win versus a level five skill that kind of needs a lot of troubleshooting.”

Three ways to build one, all demonstrated.

*Wrap an existing prompt.* Turn on Anthropic’s skill-creator, then say: “I’m going to give you a prompt for an SEO audit skill. I want you to take this prompt and wrap it in a Claude skill so that I can fire it and invoke it at any time.” Paste the prompt. He notes Claude checked his existing skills for trigger collisions but did not test the result, so test it yourself.

*Start cold, which he considers the best way.* Do not ask for a skill at all. Give Claude the problem and the outcome you want, ask it to research best practice and write the standard operating procedure first, then feed in your own expertise. He pastes in transcripts of his own past teaching. He then refuses Claude’s offer to turn it into a skill too early, in these words: “I just want markdown SOP and template docs so that we’re not locking anything in before we’ve defined it.” Only once the output is right does he capture it. If you have no transcripts to feed it, ask Claude to interview you with five questions about your process.

*Record your screen.* In Cowork, the plus button, then Record a skill. It captures your screen, clicks, typing and voice and turns them into a repeatable skill. He uses it for a reporting job he could not describe in words, narrating his way through Search Console and Analytics. The in-product warning is worth repeating: no passwords or sensitive information on screen.

The detail that trips people up. Skills built in Claude chat or Cowork work in both, and also in Claude Code. Skills built in Claude Code do not travel back to chat or Cowork. So if you want a skill everywhere, build it in Cowork.

And the rule for orchestrators, quoted from the skill Claude wrote him: “If you run the same skills in the same order every time, you failed. You are then a chain with extra steps. A chain does not need a manager.”

His own marketing orchestrator, which he describes as “a fractional CMO in a box”, routes vague requests across brand voice, keyword research, positioning angles, direct response copy, a newsletter writer and a content atomiser, and triggers on phrases as loose as “help me with my marketing”.

His full prompt guide is behind an email signup, and he runs a paid community and consults.

Read more:

2. One marketer, a seven-person workload, and 1,000 B2B leads from LinkedIn this year

Jeff Sauer founded MeasureU, ran an Inc 5000 agency for a decade before selling it, and has trained more than 50,000 marketers across 100 countries. On 10 September he went through his entire setup on a podcast almost nobody has watched, and it is the most useful B2B walkthrough of the week.

His stack, with his own labels. Claude Code in the command line through Visual Studio Code is the daily driver, “call it the marketing manager, that’s like the person who’s in charge of everything”. Then specialists: ChatGPT and Gemini exclusively for image generation, Veo 3 and Seedance for video, Descript for editing, and Copycoders’ Genesis bots for professional-grade writing, which began as a browser tool and now reaches him as a skill server over an API from inside Claude. Codex and Z.ai’s GLM models review code.

How he stops it sounding like AI. He runs an anti-AI tuning layer across every project, between 50 and 100 trained rules. Em dashes are deleted automatically and cannot be published at all. “Stop doing this, start doing that” constructions are removed. So is “it’s not X, it’s Y”. So are triplicates, the this-this-this rhythm. So are runs of uniformly short sentences.

He is honest that things still get through, particularly in 2,000 word video scripts, where he deliberately relaxes the rules “because people are watching the video more for the B-roll and the person than they are for the script”.

The structural change he made 30 days ago. Every skill he runs now has six parts: a trigger, the non-negotiable rules that can never be skipped, the procedure which is your standard operating procedure, checkpoints that must pass before the work can continue, a named place where the facts live, and a verification step at the end. Before this he had only the steps, and patched them whenever something broke. On the change: “it’s saved my butt a lot.”

The rule underneath it is the one to take away. Claude is never the judge. He has Codex review Claude’s code “because Claude thinks that its code is perfect”, switches from Opus to Fable to ask why something failed, and runs a third model over anything he publishes. “You can never take one source of truth because it’s not that good.”

The LinkedIn engine. He publishes once a day, having tested three times a day, twice a day, once a day and once a week. Claude drafts the week’s posts. For ideas, he feeds it the transcripts of every community call and every internal call and has it mine them for angles, because “we have a community of like hundreds of people who are meeting with us and they’re asking questions”. His idea file now runs to hundreds of entries.

His best performing format is a carousel showing the first three pages of a deeper PDF guide, with a comment-this-word call to action. When one lands he sees 30,000 impressions and thousands of comments.

Then the part almost nobody has automated. Claude wakes itself every three hours, opens a browser, reads his LinkedIn, and replies to every comment as him with the resource link. He built it himself because no tool did it reliably: “I don’t think there’s a software that does auto comments for LinkedIn and that’s part of why I had to do it myself.” He has also dropped his scheduling tool and now has Claude schedule natively in the LinkedIn interface.

A year ago this ran in n8n: draft the post, put it in a spreadsheet, have n8n ingest it, post it, then verify it went out. His verdict on the change: “this is way less work to do it this way.”

Landing pages. Claude built 40 landing pages for his free resources over a weekend, now around 45. A new resource page takes him about five minutes, including the imagery, the copy, the form and the opt-in. Claude wires the form into his email tool through its API, so he no longer runs form plugins, landing page builders or a separate funnel tool. Every download triggers the asset, then five onboarding emails, then the weekly.

His own caveats, which deserve printing. On the carousel: “I could say that carousel works great. It might not work for you. I’ve seen a lot of reports they don’t work for other people.” On the idea mining: “it always biases towards the most recent input that it has”, which he puts down to the model taking the fastest route. And on his YouTube calls to action, he volunteers that those bring in about five leads each, nothing close to the LinkedIn numbers.

MeasureU is his own training business, and the episode offers a free guide in exchange for an email.

Read more:

3. Write the first 150 characters of your email for the machine that reads it before your subscriber does

This one takes fifteen minutes and applies to your next send.

The concept is live text, meaning the actual HTML text inside your email. That is what Gmail’s summaries and Microsoft Copilot scan to build inbox previews and, increasingly, to rewrite your preheader.

The rule. The first 150 characters of your email must carry your most important and most specific information. The webinar date and time. The offer terms. The actual ask. That is the window AI summaries pull from, so anything vital sitting below it may never reach the reader.

What does not count. “Images and JPEGs don’t count, and alt text doesn’t count as live text either, no matter what you’ve been told.”

The preheader fix. Write your opening body copy so it echoes what you would want your preheader to say. Then if an AI layer overwrites your preheader, the message still lands correctly. The prediction offered on the episode is that within six months half of all preheaders will be machine-generated rather than written.

The line that sums it up: “The prettiest, most designed email templates are losing right now. If AI can’t read the words inside your design, it doesn’t matter how good it looks.”

While you are in there, check your Gmail verdict. Google Postmaster Tools now exposes a Deliverability Analysis that replaced the old domain reputation score, which disappeared in August. Instead of a number, it returns a plain-language verdict and remediation steps. The seven you can get are positive user feedback, volume too low to assess, sender not compliant (failing SPF, DKIM, DMARC or TLS), SMTP errors high, negative user feedback, user feedback low meaning recipients show no engagement at all, and spam rate high, meaning above the 0.1% standard. It is more useful than the score was, because it names the fault and the fix.

The guest on this episode is credited only by first name in the show notes, so we have quoted rather than attributed.

1. AI now handles 12% of US ad spend, against 2% in 2023

Digiday reported on 10 September that automated AI tools now account for 12% of US ad spend, up from 2% in 2023, and projected to reach 27% by 2030. Meta’s Advantage+ sits at $75 billion, and Google’s Performance Max accounts for roughly 30% of search spend. Madison & Wall, whose forecast this is, attributes the growth to convenience for advertisers of every size, and describes it as market share moving rather than new money arriving.

In the same week the IAB raised its 2026 US ad spend growth forecast to 12.3%, up 2.8 points from the 9.5% it projected in January, based on more than 200 brand and agency investment decision makers. Two figures from that survey are worth carrying: customer acquisition has risen to 63% as a stated media goal, and 86% of buyers expect AI to change how they measure media within 12 months.

What to take from it. If you are still debating whether to hand a campaign to an automated type, the market settled that at scale. The more useful reading is the other direction: the share is 12%, not 60%, so “everyone is doing this” is not yet true, and there is still room to win on the parts that are manually controlled.

2. AI use in marketing is wide but shallow, and brand strategy is almost untouched

Fratzke surveyed more than 350 marketing leaders across B2B, B2C and hybrid organisations, most at companies of 501 to 5,000 people, 59% at companies between $11 million and $500 million in revenue, split evenly between director level and above and marketing management. The study does not disclose when fieldwork ran, which is worth knowing.

61% of teams actively use AI. Where they use it: content writing 53%, visual generation 48%, reporting and analytics 44%, campaign personalisation 37%. Where they do not: media planning 16%, decision support 15%, brand strategy 13%.

The barriers named are data privacy and security at 47%, loss of human creativity at 45%, and unreliable outputs at 43%.

What to take from it. This is the clearest answer available to “are we behind”, and the answer is reassuring in a useful way. If you are producing content with AI, you are at the median. If you are using it for media planning or decision support, you are in the top sixth of the market. The strategic work is almost entirely unclaimed.

Fratzke is an agency that sells strategy work, which is worth noting alongside the finding.

3. 131 SEOs rank intent match first and meta descriptions last

Cyrus and Dawn Shepard surveyed 131 search practitioners on 9 September, asking them to score 103 possible ranking factors on a seven point scale, producing 13,665 data points.

The top signals were relevance at 57.1%, backlinks at 54.8% and content quality at 47.6%. Search intent match rated the single strongest factor of all 103 tested. Meta description rated last, with most respondents assigning it no impact at all. A recurring comment in the open responses: “a hundred links from real audiences beat a thousand links nobody ever clicks”, with several respondents connecting trusted links to being cited in tools like ChatGPT.

This is expert opinion, not ranking data, and should be read that way.

What to take from it. It settles a recurring argument cheaply. If your team still hand-writes bespoke meta descriptions for ranking reasons, 131 practitioners think that time belongs on intent match instead. Write them for click-through, not for rankings.

4. LinkedIn on reaching Gen Z buyers: short-form video and expert voices

We covered the marketer half of this research a fortnight ago, the survey of 6,196 B2B marketers. The larger half we did not cover, and it is the more actionable one. Censuswide surveyed 18,050 professionals aged 18 to 80 for LinkedIn across twelve markets, with fieldwork running 20 July to 11 August.

89% of B2B marketers say future growth depends on influencing Gen Z and Millennial decision makers. The top two ways to reach Gen Z buyers are short-form immersive video at 49% and content from expert voices at 42%. Separately, 72% of Gen Z professionals say they stay on the sidelines of their professional community, which is the gap those two formats are closing.

What to take from it. If your B2B content plan is long-form written assets and a webinar, the two formats named here are the ones missing. The expert voices finding also pairs with the ranking factors study above, where practitioners connected trusted sources to AI citations.

1. HubSpot put third-party buying intent data inside the CRM

Bombora Company Surge signals are now native to HubSpot, classifying companies as High, Medium or Low research activity. You select up to 12 Bombora topics matching your ideal customer profile, then use the resulting signals as workflow triggers, segment filters, personalisation tokens and scoring criteria.

For anyone running B2B demand generation in HubSpot, this is the most immediately usable change of the week. Buying intent stops being a separate subscription somebody checks on Mondays and becomes something a workflow acts on automatically.

2. Google Ads AI Dashboards are appearing in real accounts

Describe what you want to analyse in a text prompt, and Gemini builds the visual performance report, each one carrying a real-time summary explaining what moved the numbers.

This was announced back in August. The news is that it is now actually showing up in accounts. Worth looking for before your next reporting cycle rather than rebuilding the same dashboard by hand again.

3. Catchr connected more than 100 marketing platforms to Claude and ChatGPT

Live data from Google Ads, Meta Ads, Microsoft Ads, LinkedIn Ads, GA4, Search Console, HubSpot and over a hundred other sources, now available as an official connector in the Claude directory and as an app in the ChatGPT store.

The point is asking “why did our cost per acquisition rise” in plain language against current data, with no export and no spreadsheet. Three step install, 14 day trial, no card required.

4. Gemini became a cross-app orchestrator in Google Workspace

Rolling out since 2 September. From Gmail, Chat, Drive, Docs, Slides or Sheets you can now generate formatted documents, structured spreadsheets or styled slides, research across files and threads, draft and send email, schedule meetings and manage tasks, without switching app.

English first, with rollout taking up to 15 days. Available on Business Standard and Plus, Enterprise Standard and Plus, and Google AI Pro and Ultra. The version that matters to a marketer is building the deck straight from the brief, in the conversation where the brief was agreed.

5. ChatGPT Images 2.5 reached every tier, including free

Sharper detail, more natural lighting, and better preservation of subjects from reference photos. It adds Sketch, which lets you draw directly in ChatGPT as a reference for the final image, and Templates covering formats including flyers and product photos. Generation is up to 50% faster than the previous model.

The improvement that actually matters is reliability across multiple editing turns, because most people abandon AI image tools on the second or third revision, when the subject starts drifting away from the original.

Take one existing brand asset, load it as a reference, and run three sequential edits on it. If the subject holds through all three, it is usable in production. If it does not, you have saved yourself a week of trying.

6. OpenAI put full-duplex voice in the API as GPT-Live-1

It listens and speaks at the same time, handing deeper reasoning to whatever model you pair it with, and scores 30 percentage points better on a full-duplex benchmark than its predecessor. For B2B the obvious uses are inbound lead qualification by phone, demo and webinar booking, and first-touch outbound qualification.

One pricing detail to get right before you scope anything. The five cents per minute covers the front-end voice layer only. The backend model and any tools it calls bill separately, so there are two invoices, not one.

1. Google is testing ordinary Search campaigns inside AI Mode.

Ginny Marvin confirmed on 4 September that exact and phrase match keywords are now eligible to serve text ads there, limited to queries carrying explicit and direct intent. Until now that surface required AI Max or Performance Max.

2. Claude Code added a /skill-doctor command that finds the skills you never use.

It arrived in the releases between 5 and 9 September, and it pairs neatly with the argument running through this week’s Playbook that fewer, better-built skills beat a long list of them.

3. John Mueller says it is now difficult to map positions one through ten in search.

Useful the next time somebody asks for a rank tracking report as the primary measure of SEO health.

4. Google put a character counter widget directly inside AI Overviews.

Sites whose traffic comes from offering one simple free utility are the clearest early casualty of Google absorbing the tool rather than linking to the site that built it.

5. LinkedIn’s EU transparency disclosure reports a 46% rise in detected inauthentic activity

in the first half of 2026 against the second half of 2025, covering engagement pods, third-party automated posting services, fake profiles and AI-generated comments. It also disclosed 56.5 million active EU users, up 1.4 million.

6. Anthropic published what it learned from more than 1,000 small business owners

across six weeks of workshops in ten US cities. Many attendees were using the chatbot every single day but had never tried Cowork, skills or connectors, which are the features that actually automate multi-step work.

7. Anthropic’s September threat report documents a commercial influence-for-hire operation

that ran roughly 70 fabricated news websites, more than 250 fake commenting accounts and at least 8,913 articles in around 20 languages, with invented bylines built to look like independent local newsrooms. Its three named tactics, rewriting one source story in opposite ideological directions for different audiences, adding political angles to stories that had none, and moving stories across borders stripped of their context, are a legitimate content repurposing workflow with the intent reversed.

1. Measurement got louder and harder to check in the same week.

Google released a metric that scores the value of your first-party data. The IAB found 86% of buyers expect AI to change how they measure media within a year. Automated tools now run 12% of US ad spend. What did not arrive is any way to verify the vendor numbers underneath all of it, which is why Meridian GeoX becoming free and global matters more than the uplift metric sitting next to it. An experiment you ran yourself is the only measurement you fully own.

2. The constraint moved from what the models can do to who holds the permissions.

Astra is in ChatGPT Work and Codex for Plus subscribers but not in the ordinary chat window. The new data agent needs IT to connect a warehouse. Brand memory tools ship as governed layers in early access. Usage reporting is Enterprise only. Nothing held marketers back this week because a model could not do the task. The blockers were plan tiers, admin toggles and access requests, which is a different problem and one that is solved by a conversation rather than a subscription.

3. The answer box finished eating the results page, and the org chart started moving in response.

People Also Ask went from 12% AI-generated to 97% in fourteen months. Google absorbed a utility widget rather than linking to one. The citations panel got thinner. In the same week, two HubSpot executives described the B2B CMO role being split between product and sales because the specialists of the last twenty years are not the generalists this moment needs. The channel change and the organisational change are the same story, arriving about a year apart.

QUICK REMINDER:

CELEBRATING 50+ EDITIONS OF THE MARKETING WITH AI WEEKLY ROUNDUP

For marketers who want their hours back. The full Marketing Operator System, the Setup Skill that builds your working environment, and 5-Minute Slides, together for the first time.

And to celebrate 50+ editions of this newsletter, it’s now just $150 or 2 x $75 (Individual price $953).

Check Out This Limited Time Offer

Tell me: What’s your biggest takeaway from this week’s news from the world of AI Marketing? Love to hear your thoughts in the comments.

Leave a comment

Thanks for reading the Marketing with AI Weekly Roundup! If you enjoyed this read, the best compliment I could receive would be if you shared it.

Marketing with AI Weekly Roundup: Two Frontier Models Land on Exactly the Same Price, Your AI Citation Numbers Do Not Mean What You Think, and Google's Newest Search Model Stopped Linking to Anyone

Robert Gillespie — Sun, 06 Sep 2026 10:01:09 GMT

Welcome to issue #50 of the Marketing with AI Weekly Roundup. Published every Sunday, it shares the top stories, playbooks, data and quick hits from the world of Marketing with AI over the last 7 days.

Subscribe to receive the week's most important AI marketing stories direct to your inbox.

Subscribe now

CELEBRATING THE 50TH EDITION OF THE MARKETING WITH AI WEEKLY ROUNDUP

For marketers who want their hours back. The full Marketing Operator System, the Setup Skill that builds your working environment, and 5-Minute Slides, together for the first time.

And to celebrate the 50th edition of this newsletter, for the next 48 hours only, it is just $150. (Individual price $953).

Check Out This Limited Time Offer

Anthropic shipped Fable 5.1 on Tuesday and OpenAI shipped GPT-6 Astra on Thursday. Both now cost exactly ten dollars per million input tokens and fifty per million output. That is not a coincidence worth shrugging at. When two frontier labs land on an identical sticker price within 48 hours, the competition has moved somewhere else, and this week it moved to what it costs to re-read the same context over and over.

That matters more than it sounds, because re-reading the same context is exactly what a marketing agent does. Anthropic cut the price of it by 75%. There is a caveat that most of the coverage skipped, and it applies to most of you: if you use Claude on a Pro or Max subscription, none of this touches your bill.

The bigger story for anyone reporting on AI search is that three separate datasets, published in the same fortnight and using three completely different methods, all found that being cited is not the same as being recommended. Only about one cited brand in five becomes the answer’s actual pick. Two thirds of B2B vendor mentions carry no link back to the vendor at all. And most of the citations that do appear point at pages you do not own. If AI visibility is on your reporting line, the number on it needs a footnote it currently does not have.

Then Google put its newest model into AI Mode on Wednesday and it promptly stopped linking to anyone. Google has confirmed it is a bug. The useful part is what you should not do about it, because the instinct is to go and rewrite content that was never the problem.

Also this week: Claude can now work your desktop in the background on Windows as well as Mac, LinkedIn quietly broke campaign creation for anything touching its API, and a survey of formally trained marketers asked what skills they are missing, and AI came near the bottom.

Let’s get into it…

1. Two frontier models, one price, and the fight moved to what it costs to remember

Anthropic released Claude Fable 5.1 on 1 September. Headline token prices did not move. It is still ten dollars per million input tokens and fifty per million output. What changed is the price of a cache read, which fell from one dollar to twenty-five cents per million, a 75% cut. Anthropic puts the practical effect at roughly 25% cheaper for typical workloads and up to about 45% for heavily agentic ones. On agentic benchmarks the jump is large, 55.8% on Terminal-Bench 4.0 against 42.0% for Fable 5, and 52.6% on Terminal-Bench-Science against 24.7%. It runs as claude-fable-5-1 on the Claude API, AWS, Google Cloud and Microsoft Foundry. A sibling model, Mythos 5.1, is the same underlying system with some safeguards lifted, restricted to vetted cybersecurity and life-sciences organisations.

Read this before you get excited. A cache read is what you pay to send the model context it has already seen. If you are on a Pro or Max subscription, you do not pay per token, and this changes nothing for you at all. It matters if you or your agency run anything on API billing, and it matters most for the specific shape of job where a long-running agent re-reads the same brand guidelines, ICP definition, approved claims and past campaign data on every single run. That is the workload getting the 45% figure.

OpenAI released GPT-6 Astra on 3 September, and pitched it as “the world’s best computer use model”, built to work across browsers, spreadsheets, websites and desktop applications and produce finished documents rather than instructions for making them. It scores 72.6% on OSWorld 2.0 at roughly 47% less time per task than its predecessor. Its context window is 1,050,000 tokens, of which up to 922,000 can be input. And its price, on OpenAI’s own developer documentation, is ten dollars per million input and fifty per million output, with a Fast mode at double both.

Two things about Astra need saying plainly. It launched to a limited set of organisations in OpenAI’s Trusted Access Program, with ChatGPT Plus, Pro, Business and Enterprise plus the API, Azure and Bedrock following “in the coming days”. So you probably cannot use it yet. And every benchmark quoted for it is OpenAI’s own, exactly as every benchmark quoted for Fable 5.1 is Anthropic’s own. The “AGI era” headlines you saw came from Greg Brockman closing a press briefing with that phrase, not from any independent finding.

What to actually do. If you pay per token, go and look at how your prompts are ordered. Caching only pays off when the stable material comes first and the variable input comes last, so a prompt that opens with this week’s campaign data and buries the brand context at the bottom will miss the discount entirely. That structure just became four times more valuable than it was last Monday. If you are on a subscription, the honest answer is that you got a better model at the same price and there is nothing to do.

2. Three separate datasets found that being cited is not being recommended

If you report AI search visibility to a client or a board, the number you are reporting probably does not mean what you have been told it means. Three datasets published inside the same window, using three different methods, all landed on the same finding.

Parse analysed 2,814,094 brand-attributed citations, forming 1,189,703 cited brand-answer pairs across 387,695 AI answers and 108,246 brands, on ChatGPT Search and Google AI Mode between 6 July and 23 August. Only 21.9% of cited brands became the answer’s top recommendation. Being cited does help, it roughly doubles your odds of being the pick, 22.0% against 10.4% for brands mentioned without a citation, and 99.3% of cited brands at least made the shortlist. But one in eight heavily cited brands converted under 5% of the time. Their own summary is the line worth carrying: a citation is not a recommendation about four times in five.

Cleanlist approached it from the opposite end and published its full dataset openly. Across 205 buyer-intent queries in B2B data and prospecting tooling, it found that in 66.8% of cases where a vendor was named in an AI Overview, that vendor’s own domain was not cited in the answer at all. Apollo was named in 64.4% of answers, ZoomInfo in 40%, Clay in 32%.

Cognizo, published 3 September, found that roughly 88% of AI Overview citations go to pages the brand does not control.

Three things follow from this, and they change what you do on Monday.

First, being named and being cited are different events, and most tooling only sees the second one. If your AI visibility dashboard works by tracking links back to your domain, Cleanlist’s number says it is blind to about two thirds of the times you actually get mentioned. Start tracking mentions separately from citations.

Second, stop reporting citation counts as the outcome. Four times in five, the citation went to a brand that did not get the recommendation. A rising citation count next to flat pipeline is not a contradiction, it is the normal case.

Third, most of what gets you cited is not yours. If 88% of citations point at pages you do not control, then review sites, comparison pages, forum threads and analyst write-ups are doing the work, and your own site is a minority shareholder in your own visibility.

Declare the interests, because they are real. Parse sells citation-to-recommendation tracking. Cleanlist is a B2B data vendor that chose its own queries and, to its credit, finishes last in its own leaderboard while stating eight separate limitations. Any one of these alone would be a vendor number. The reason to trust the direction is that three of them, with different commercial incentives and different methods, converged on it in the same fortnight.

3. Google’s newest AI Mode model stopped citing sources, and Google has confirmed it

And in the same week the research above questioned what citations are worth, a large share of them briefly stopped appearing at all.

On 2 September Google made Gemini 3.8 Flash available as a model inside AI Mode for AI Pro and Ultra subscribers. Within a day, SEO practitioners were posting examples of top-of-funnel queries coming back with no links and no citations at all. Gagan Ghotra published a run of them.

Google confirmed it. Robby Stein, VP of Product for Google Search, said: “This isn’t working as intended, and we’ll roll out a fix soon.”

The important part is what not to do. If your AI Mode citations or AI referral traffic dropped in the first week of September, the cause is a Google bug and your content did not become worse overnight. Do not rewrite anything in response to it. Do not let a panicked colleague commission a content refresh off the back of a dip that will reverse when Google ships the fix.

Do annotate it. Whatever you use to track AI visibility, put a note against this week so the dip does not quietly become your September baseline and make October look like a recovery you engineered. If you report AI visibility to a client or a board monthly, this is the week to explain in advance why the line moved, rather than defending it afterwards.

And take the wider lesson, which is the genuinely useful one. Google swapped the model generating a large share of AI answers, gave nobody any notice, and the citation behaviour changed materially within a day. Your AI search visibility can halve because of a decision made inside Google that has nothing to do with you. That is the argument for checking it weekly rather than quarterly, and for never building a channel strategy on a number you do not control.

1. The four-element brief, and why the second prompt is the one that matters

A 48-minute walkthrough published on 1 September goes through seven marketing workflows in Claude, and the structure underneath them is more valuable than any individual workflow.

Every prompt in it contains the same four things: the target audience, the business objective, the tone of voice, and how the output should be structured. His argument for why: “When those four pieces are present, Claude spends less time filling in missing context and more time producing useful marketing work.” His diagnosis of why most marketers get generic output is blunter. “Isolated prompts produce isolated results.”

The structural point nobody else made this week. Every one of the seven workflows is a first prompt followed by a specific, named follow-up prompt, and the follow-up is where the value sits. The first prompt produces output. The second turns that output into a decision. Even if you never run one of these seven, that pattern is the transferable thing.

Workflow one, ad variations that are genuinely different. The brief was an AI time-tracking tool, aimed at operations managers and agency directors losing unbilled hours to scope creep, on LinkedIn, with high-intent trial sign-ups as the objective. The prompt demanded a unique headline, a different hook, a specific emotional angle, a call to action and a recommended format for each variation. It returned five built on five genuinely different emotional triggers: a “revenue leak” angle using loss aversion and a before-and-after timesheet carousel, a “burnout signal” angle about a project manager close to quitting, a “scope creep trap” angle about unpaid client work, a “sceptic’s proof” angle leaning on social proof, and a “direct ask” angle built around removing friction. All held to 400 to 450 characters.

Then the follow-up: ask Claude to break down the psychology behind each one, identify which buyer awareness stage each targets, unaware, problem-aware or solution-aware, and explain the conditions under which each would perform best. As he puts it, the marketer “can then choose creative based on strategy rather than personal preference.” That is the difference between five ads and a testing plan.

Workflow two, the competitor audit that finds unclaimed territory. Paste the raw homepage copy of three competitors into one prompt and force the comparison across five fixed sections: core value propositions, audience cues and language signals, messaging angles and emotional territory, the balance between features and outcomes, and unclaimed gaps. The fixed structure is the whole trick, because without it you get three separate homepage summaries instead of a comparison.

Run on Asana, monday.com and ClickUp, it produced a territory map. Asana owns clarity and order, using governance language like “cross-functional teams, programs, initiatives” that appeals to operations and PMO leaders. monday.com owns speed and ambition, with verbs like “outpaces” and “squashes”, aimed at growth-driven companies. ClickUp owns completeness and relief, with consolidation language aimed at IT buyers tired of managing several systems. All three mention AI as a credibility signal. All three position horizontally, and not one names a specific industry, company size or user type.

It then named six gaps nobody in the category is claiming: simplicity for small and non-technical teams, transparent pricing, trust and security and governance, a specific and explained AI difference rather than AI as a badge, team culture and burnout, and vertical specialisation. The follow-up prompt asks for the three most overcrowded tropes in the category, two defensible positioning angles, and a hero headline and sub-headline for each. That is a positioning decision, not a research document.

Workflow three, a content calendar that is an arc rather than a schedule. The brief was B2B SaaS compliance monitoring for healthcare, aimed at compliance officers and operations directors, three LinkedIn posts and one article a week for four weeks, with the output demanded as a table. What came back progressed deliberately: week one urgency, week two practical problems, week three proof, week four conversion, with calls to action escalating from low commitment, follow and comment, through medium, read and watch, to high, book a demo. The mix held at roughly half educational, a quarter engagement, a quarter conversion.

Where to start, in his own order: build the brand voice blueprint first, then move to ad copy testing. One note on the video itself, workflow four uses an e-commerce email example, so skip that one and take the other six.

Watch more:

2. The five folders that stop your AI work disappearing every week (Greg Isenberg)

Published 31 August and already past 100,000 views, this is the clearest argument yet for treating marketing AI as a system you build rather than a tool you open.

The problem it names will be familiar. “Most people use AI in these random chats. They’ll ask for 10 posts and maybe they’ll copy one into a doc that they like, and then the work just disappears. Next week the AI is starting from scratch again.”

The fix is a folder called growth-os, either a repository or, in his words, “honestly just a structured folder”. Five folders inside it:

● Customer truth, holding sales calls, support tickets, churn notes and interviews

● Content engine, holding the founder voice guide, hooks that worked, and what performed

● Outbound engine, holding the ICP, trigger events, approved angles, and banned language

● Creative testing, holding ad angles, page tests and results

● Agents, holding the job specs

The prompt upgrade is the clearest illustration of what changes. Instead of “write me 10 LinkedIn posts”, it becomes “read the customer truth file, read the founder voice file, read the last five posts that drove qualified replies, and draft five new posts around pains buyers actually mentioned this week.” Same tool, entirely different output, because the context is on disk rather than in a chat you will lose.

Write agent specs like job descriptions. His template covers the data source, when it runs, what it filters out, what good output looks like, what needs human approval, the metric that matters, and where it writes the result so the system gets better next time. His worked example runs each weekday morning across twenty LinkedIn accounts, pulls commenters, enriches them, drops poor fits, and drafts messages tied to the specific post each person engaged with. The line worth stealing is about measurement: “Messages sent is activity. Qualified replies is going to be your signal.”

If you only do one thing, he gives a minimum version: five files, twenty real customer notes pasted in, and one question. “Tell me what’s changed, show me the receipts, suggest one marketing test that could create pipeline this week.”

One honest caveat. This is a framework episode. There are no client results or verified numbers in it, and the tooling he mentions is his own preference, which he says himself. The closing line is the part that survives: “The agents are going to be a commodity at some point. Your judgment about what to point them to is the moat.”

Watch more:

3. Seven ChatGPT Work skills, and the delegation trick underneath them (Paul J Lipsky)

Published 31 August, the highest-viewed practitioner video of the week at over 110,000 views, and the prompts are published free with no email gate.

Where skills live, because this trips people up. In the ChatGPT desktop app, go to Plugins in the left sidebar, then the Skills tab at the top. You can click Add, but his recommended method is faster: open a new chat in ChatGPT Work and simply describe the skill you want, and it saves it as a reusable one. Invoke any of them by typing a forward slash. This is paid-plan only, inside ChatGPT Work.

The reason to bother, in his words: “Without skills, you have to re-explain your workflows to ChatGPT every time you use it. And that costs you time and gives you inconsistent results.”

The four with the clearest marketing read-across:

● /visualize, which ships pre-installed and almost nobody has found. It turns a text answer into an interactive dashboard. He pointed it at a local folder of budget spreadsheets and got a monthly and annual toggle with a costs table. It is the fastest route from a spreadsheet to a client-ready chart without opening a BI tool.

● /handoff, which packages a stalled project for a different AI. It writes out the goal, what has been done, the decisions made, the files and links, and exactly what happens next. He moved a stuck build from ChatGPT to Claude, Claude finished it, and handed back.

● /ask, his own build, which uses ChatGPT’s browser to go and ask a rival AI. He has it open grok.com for real-time X sentiment that ChatGPT cannot see, and use Gemini when he prefers its images. His framing is the useful bit: “It takes advantage of the subscriptions I’m already paying for. You can just have ChatGPT as the central delegator who’s then sending the work off to the other tools and then bring it all back together in one place.”

● /grillme, which forces an interview before it starts building anything, so it asks you questions in rounds until the brief is actually clear.

Disclosure: the video carries a mid-roll sponsorship for design.com. The skills content itself is not sponsored.

Watch more:

1. LinkedIn surveyed 6,196 B2B marketers, and the finding is about people, not tools

Fieldwork by Censuswide for LinkedIn ran from 21 July to 6 August across 6,196 B2B marketers at junior manager level and above in twelve markets, with minimum quotas per market for both under-30s and director level and above. It is the largest B2B sample published in the window by a distance.

There is no single global release, only regional cuts, so the numbers depend on which market you quote. In the Singapore cut, 47.4% expect to increase investment in AI search optimisation over the next twelve months, 71.6% say buyers prefer human expertise over AI-generated content, and 50.8% say expert-voice content outperforms traditional campaigns at moving buyers from awareness to consideration. In the India cut, 69% expect increased AI search spend and 82% say credibility now matters more than visibility.

Disclose the obvious interest. LinkedIn sells Thought Leader Ads, and publishes these findings alongside a claim that adoption is up nearly 40% year on year. The direction of travel is still worth acting on: put named humans on your content, and start treating AI search optimisation as its own budget line rather than something absorbed into SEO.

2. Two independent samples landed on roughly the same fifth

Gartner surveyed 1,303 functional leaders at organisations above $50m revenue, with fieldwork from January to April 2026, and found 22% have successfully scaled AI across multiple business units or gone AI-first. 85% plan to increase AI spend this year, having averaged 12% of functional budget on it in 2025. The detail worth carrying: 11% do not know what their function spent on AI last year.

Salesloft surveyed 500 US sales and revenue decision-makers and found 20.6% describe their AI strategy as production-ready with measurable outcomes, with 28.2% still experimenting.

Two different samples, two different methodologies, both landing near one in five. The next time a vendor tells you everyone else has already scaled this, that is the number to hold against it. And before you ask for more AI budget, be able to say what you spent last year, because one leader in nine cannot.

3. Trained marketers were asked what skills they lack, and AI came near the bottom

The Drum surveyed 216 marketers who have completed the MiniMBA in Marketing since 2016, published 3 September. Asked what capability will matter most over the next five years, 34.3% said strategic thinking. Asked where their organisations are weakest now, 31% said strategic planning.

Only 10.2% named keeping pace with AI and new technology as their biggest current challenge, and just 4.2% chose AI and technology as the area most needing training. The biggest named training need was strategic marketing skills at 37.5%.

Declare the interest, because it is significant. MiniMBA sells strategy training, and its course opened on 7 September. A survey of its own alumni finding that marketers need strategy training is also the product pitch.

Even allowing for that, it is the only dataset published this week arguing the opposite of everything else in this issue, and it is worth sitting with. The people who have been formally trained in marketing, asked openly what they are missing, mostly did not say AI.

1. Claude can now work your desktop in the background, on Windows as well as Mac

Anthropic shipped background computer use on 2 September, in beta on Pro and Max plans, inside Cowork and Claude Code in the desktop app, on macOS and Windows. You turn it on at Settings, then General, then Computer use.

The news is not that Claude can click and type on your machine. It is that it now does so in a background window instead of taking over your screen. It will not grab your mouse mid-task, it waits if you are typing, and it asks once per session if it genuinely needs the full display.

The order it tries things in matters. It reaches for a connector first, Gmail, Drive, Microsoft 365 or Slack, then a browser, and only then your actual screen. So it is a fallback, not a first resort.

Why this is worth a marketer’s attention. The tasks that never get automated are the ones living in tools with no API and no connector, which in marketing means the ad platform with the terrible interface, the internal dashboard, the legacy CMS, the client reporting portal you log into once a month. This is aimed squarely at those. Pick one weekly task that lives in a tool nothing integrates with and hand it over as a background job.

Note this is different from the browser autonomy we covered on 29 August. That was Claude acting inside a browser without asking each time. This is control of the desktop itself, on two operating systems, a week later.

2. Google’s AI Search opt-out switch is now live on every site in the world

Alongside completing the worldwide rollout of the generative AI performance report on 31 August, Google made the accompanying control global. It decides whether your site appears in, and helps ground, Google’s generative AI answers.

We have covered the report itself before, at launch in June and again in August. The switch is new, and it is the part that needs a decision rather than a read. Google’s wording is unambiguous: sites that opt out “will not receive traffic or impressions from our generative AI features”, although it says this is not a ranking signal in standard Search.

Treat it as a one-way trade and do not let anyone flip it casually. The report still shows impressions only, with no clicks and no queries, so you can see that you were shown and still not know whether anyone came.

3. Adobe bought an agentic workflow startup less than a year old

Adobe acquired Rilo on 2 September for undisclosed terms. Founded in September 2025, it passed 10,000 users within months. It lets non-technical people describe a workflow in plain English and have agents build and run it across hundreds of tools.

The four use cases Adobe names are all marketing operations jobs: competitor intelligence, content repurposing, sales call analysis and lead generation. Adobe also owns Semrush, which shipped its own Claude connector days earlier, which makes this the second agentic marketing move by one owner inside a fortnight. If your stack is Adobe, plain-English workflow building is coming to you whether you asked for it or not.

4. Dreamdata is trying to stop your pipeline numbers changing depending on who asks

Dreamdata launched Dreamdata AI on 1 September, three products around the same idea: an Analytics Agent inside the platform, an MCP server exposing the same data to Claude and ChatGPT, and a data warehouse export for teams connecting their own agents.

The problem it names is the interesting part. Point a language model at go-to-market data and you get a fast answer that is difficult to verify, and a different answer depending on which model asked and how the person phrased it. Dreamdata’s response is a fixed set of metric definitions it calls a governed semantic layer, sitting under all three products, so the definition of a qualified lead or an influenced opportunity does not get reinvented per conversation.

That is a real problem for any B2B team letting people query pipeline data conversationally, and it is the same problem underneath this week’s CRM data research. Worth a look if you are being asked to put reporting in front of an assistant.

5. Google Ads is removing language targeting from Search campaigns this month

From late September the campaign-level language setting disappears from Search and AI Max for Search campaigns. Google will instead use the language of your creative and landing page, combined with its own understanding of which languages a user knows. Performance Max is affected on the Search channel only, and keeps the setting elsewhere.

If you run multi-language accounts, your ad copy and landing page language have just become your targeting mechanism. Audit for campaigns where the creative language and the intended audience language do not line up, because that mismatch used to be caught by the setting you are about to lose. This was originally scheduled for the end of 2025, so it is nine months late and easy to have stopped watching for.

6. Two LinkedIn Marketing API changes took effect on 31 August

Worth knowing about if anything automated touches your LinkedIn advertising, because neither change was covered by the trade press and both fail quietly.

Legacy geography values are gone. From 31 August, campaign creation using legacy geo returns a 400 error with an INVALID_VALUE_FOR_FIELD code, and Bing geography URNs are required instead. LinkedIn’s wording addresses “all partners and developers”, and the change is documented against Marketing API version 202608 onward. If you build campaigns by hand in Campaign Manager, this almost certainly does not affect you. If your agency platform, CDP sync or an in-house script creates campaigns through the API, it might already be failing.

And every sponsored account is now capped at 1,000 DMP segments, covering Matched Audiences and Predictive Audiences together. Over the cap, the /dmpSegments endpoint returns a 429 with SEGMENT_LIMIT_EXCEEDED. This applies across all API versions. LinkedIn’s advice is to delete unused segments before creating new ones. Third-party write-ups report that expired and archived segments count toward the limit as well, though LinkedIn’s own documentation does not confirm that, so treat it as likely rather than certain.

Two quick checks if you run LinkedIn at any scale. Create one test campaign through whatever platform you use and confirm it completes rather than assuming silence means success. Then count what is sitting in your segment library, including anything long expired, because a mature B2B account can be near a thousand without anyone having deliberately built that many.

1. Semrush shipped an official Claude connector, and we missed it.

It went live on 26 August, which fell between two of our issues. It puts 28.8 billion keywords, backlink data and competitor gap analysis one prompt away inside Claude, for Semrush One and SEO Classic subscribers. It needs a paid Claude plan, and queries are metered against your Semrush API units rather than being free.

2. Google’s John Mueller says serving markdown versions of your pages to AI crawlers does nothing.

In his own testing, the only crawlers that requested the markdown files were SEO tools. Not one major AI crawler touched them. He called building them specifically for bots “a stupid idea”. Check your logs before anyone on your team spends a sprint on this.

3. Gemini 3.8 Flash pricing doubles on 1 January 2027.

Introductory pricing runs at $0.75 per million input and $3.75 output through 31 December, then goes to $1.50 and $7.50. If you are modelling the cost of high-volume agent work into a 2027 budget, model the higher number.

4. Google avoided a breakup of its ad exchange, and got its auction rewritten instead.

The remedies ruling on 2 September lets Google keep AdX and DFP, but requires real-time bid data to be shared with rival ad servers, deprecates unified pricing rules, lets publishers set per-bidder floors, and bans “first look” and “last look” on open-web display. Both sides have until 2 October to file a proposed final judgment. If you buy open-web display, treat your 2026 CPM benchmarks as unreliable for 2027 planning.

5. The FTC’s orders over a fabricated AI advertising product now run for twenty years.

CMG Media, trading as Cox Media Group, along with MindSift and 1010 Digital Works, sold an “Active Listening” service they claimed used smart-device audio to target ads geographically. It did nothing of the sort. The actual product was resold email lists bought from data brokers. The three pay $930,000 between them, and are barred from misrepresenting what advertising services do, how voice data is collected, and whether consumers consented.

6. UK marketing departments are estimated to be spending GBP 63.7m a year on AI subscriptions nobody uses.

The figure comes from research into GBP 8.9bn of unused UK software generally, and singles out individually bought copywriting, image generation and transcription tools as the worst offenders. The researchers call the model conservative.

7. The AI Max auto-upgrade we flagged on 15 August began on 1 September…

…and completes by the end of the month. Campaigns using automatically created assets are upgraded with search term matching and text customisation switched on. Campaigns using campaign-level broad match get search term matching only, with text customisation and final URL expansion left off. Two accounts can both be “upgraded” and end up with very different amounts of Google rewriting their ads. Dynamic Search Ads are still not in this wave, and do not migrate until February 2027.

1. Headline price converged, so the competition moved underneath it.

Fable 5.1 and GPT-6 Astra now cost exactly the same per token. That is a genuine first, and it means the differentiation has shifted to things most buyers never look at: what it costs to re-read context, whether you can pay double for speed, and who is allowed access to the top tier at all. For marketers the practical consequence is that “which model” is becoming a less interesting question than “how is your work structured”, because the cheapest available saving this week came from reusing context rather than re-explaining it.

2. Every model this week was sold as a worker, not a writer.

Astra’s pitch is operating software and producing finished documents. Fable 5.1’s headline gains are agentic. Claude shipped background control of your actual desktop. Not one of these launches led on writing quality, which is still what most marketing teams use these tools for. The gap between what is being built and what is being used is widening, and the Playbook items above are all, in their different ways, about closing it.

3. Measurement got less trustworthy in the same week it got more important.

Google swapped the model behind a large share of AI answers and citations collapsed overnight. Search Console still reports impressions with no clicks. Three separate datasets showed that the citation metric everyone is reporting does not mean what people think it means. If you are building AI visibility into a board report this quarter, build in the caveats at the same time, because they are not going to get smaller.

If this was useful, forward it to one marketer who read about Claudeforce this week and has already told their team the CRM is about to write their campaigns. It ships with thirty-seven sales skills and none for marketing, and there is no date for the ones they are waiting for. See you next Sunday.

QUICK REMINDER:

CELEBRATING THE 50TH EDITION OF THE MARKETING WITH AI WEEKLY ROUNDUP

For marketers who want their hours back. The full Marketing Operator System, the Setup Skill that builds your working environment, and 5-Minute Slides, together for the first time.

And to celebrate the 50th edition of this newsletter, for the next 48 hours only, it is just $150. (Individual price $953).

Check Out This Limited Time Offer

Tell me: What’s your biggest takeaway from this weeks news from the world of AI Marketing? Love to hear your thougths in the comments.

Leave a comment

Thanks for reading the Marketing with AI Weekly Roundup! If you enjoyed this read, the best compliment I could receive would be if you shared it.

Marketing with AI Weekly Roundup: Salesforce Puts Claude Inside the CRM and Forgets Marketing, Claude Stops Asking Permission, and LinkedIn's Slop Number Is Being Misread

Robert Gillespie — Sun, 30 Aug 2026 10:02:41 GMT

Welcome to issue #49 of the Marketing with AI Weekly Roundup. Published every Sunday, it shares the top stories, playbooks, data and quick hits from the world of Marketing with AI over the last 7 days.

Subscribe to receive the week's most important AI marketing stories direct to your inbox.

Subscribe now

MY BEST BUNDLE YET FOR MARKETERS (LIMITED TIME OFFER):

For marketers who want their hours back. The full Operator System, the Setup Skill that builds your working environment, and 5-Minute Slides, together for the first time. $250, everything included. (Individual price $953).

Check Out This Limited Time Offer

Salesforce and Anthropic announced Claudeforce on Wednesday. It puts your CRM inside Claude, it shipped with thirty-seven prebuilt skills, and every one of them is a sales skill. The marketing section on Salesforce’s own product page reads “coming soon” and carries no date.

So if you saw the coverage and told your team that Claude is about to start reasoning over campaign data, walk that back before anyone builds a plan on it. The pilot is real. The marketing half of it does not exist, and nobody at Salesforce will say when it will.

Anthropic had a heavy week either side of that. Claude’s memory now carries across chat and Cowork, which makes a curated set of brand facts worth writing once instead of re-explaining every session. Claude also stopped asking permission before each action it takes in a browser, and shipped a browser of its own inside Cowork so it no longer has to borrow yours. That second one deserves a deliberate decision rather than a shrug, and there is a decision rule for it below.

LinkedIn produced the week’s other correction. It reported a 40% drop in views on posts its systems classify as AI slop, and that figure is being pinned on the million-odd people who have pressed the “seems like AI slop” button. LinkedIn’s own communications lead says it came from the classifiers instead. Nobody can throttle your reach by reporting you, which changes what you should do about it.

Google starts force-migrating Search campaigns to AI Max on Tuesday, and the two affected groups do not get the same settings, which is where the damage will be. Fourteen months of data from fifty-four advertisers puts a real number on what AI search has taken from organic traffic, and it is harder reading than the “fewer but better visitors” line doing the rounds. And 500 marketers were asked whether their CRM data is fit for AI to use, with only 21% saying yes, which is an awkward thing to learn in the same week your CRM moves inside an assistant.

Let’s get into it…

1. Salesforce put Claude inside the CRM, and shipped nothing for marketers

On 26 August 2026, Salesforce and Anthropic announced Claudeforce, an expanded partnership that makes Salesforce data, workflows, business logic and governance available inside Claude. It was announced the same day Salesforce reported its Q2 FY27 earnings, and it is the first time Salesforce has attached its “force” suffix to another company’s product.

The first product is Salesforce in Claude, a plugin carrying thirty-seven prebuilt skills. It runs on Headless 360, Salesforce’s architecture for exposing the platform to an agent over MCP, the Model Context Protocol, which is the open standard for connecting AI models to external tools and data. It is with select pilot customers now, with an open beta expected in September.

Read the skill list before you plan around this. All thirty-seven are sales skills. The named examples are meeting prep, deal health review and pipeline review. Salesforce’s own Claudeforce page carries a Marketing section stamped “coming soon”, promising to ground campaigns with context, create campaigns that self-correct, build content for agents and humans, and personalise every customer engagement. Service and Commerce carry the same stamp. None of it exists yet.

And nobody will commit to a date. Salesforce’s investor newsroom says additional skills “will begin launching in late 2026”. Several outlets briefed on the same day reported “third quarter” instead. Salesforce runs an offset fiscal year, which is why those two can both be technically accurate and why marketing teams have no firm date at all.

Three things change for a marketer even so, and none of them wait for the marketing skills.

The first is that the CRM interface becomes optional, which promotes data quality from an operations chore to a marketing problem. If an assistant is reasoning over your pipeline rather than a person reading a dashboard, nobody eyeballs the record and silently discounts the field they know is wrong. That pairs badly with the Validity research further down this issue, which found only 21% of marketers say their CRM data is very well prepared for AI.

The second is that permissions become an AI access scope overnight. Access through Claude mirrors existing CRM access, inherited from Salesforce roles under a single admin-managed connection. An over-broad marketing profile that nobody has audited since 2023 is now the reach of an agent rather than the reach of a person who would not have bothered looking.

The third is commercial. This arrives as two line items, not one. Salesforce bills headless API consumption against your licence edition, and Claude inference is contracted separately with Anthropic. Executives have acknowledged it is not a single purchase yet.

Why it matters / what to do: The honest position is that there is nothing to switch on this week, so spend the week preparing rather than waiting.

Start with a data quality pass on the four objects any marketing skill will read first: Campaign, Campaign Member, Lead and Opportunity. Do not attempt a general cleanup, because you will not finish it. Look specifically for fields where two sources disagree, because a skill has to be told which one to trust and right now nobody has written that down.

Then read your marketing team’s Salesforce profiles as an AI permission scope rather than as admin housekeeping. The question is no longer what a person could look up if they were curious. It is what an agent will pull into a summary without anyone deciding it should.

Finally, name the one weekly marketing job you would hand over first. Pipeline-informed content planning and campaign-to-pipeline reporting are the two obvious candidates, and the existing sales skills already prove both are technically possible.

One thing you can act on immediately, if you want to move faster than the roadmap. Ryan Stokes of Salesforce told CIO that pilot users are “leveraging Claudeforce to vibe code their own CRM interfaces, creating purpose-built command centers for how they want to run their teams”. That pattern does not require the pilot at all. Anyone with the Salesforce connector or the Headless 360 MCP server can build a marketing command centre against live data now, without waiting for a skill that has no ship date.

2. Claude now acts in your browser without asking, and brought its own browser

Also on 26 August, and largely buried under the Salesforce news, Anthropic changed how Claude behaves inside a browser. In its own words, Claude “can now also take actions autonomously in the browser, instead of needing approval for every one.”

This is the news, and it is not the same as availability. Claude in Chrome has existed for months. What changed is that it stopped stopping. A classifier now reviews each action Claude is about to take, such as navigating to a new site or entering text into a page, and checks it against what you originally asked for. Probes scan page content for prompt injection, and Claude is warned when an attack is detected. It is on every paid plan, and enterprise administrators can restrict it to approved domains.

Anthropic also gave Claude a browser of its own. Claude Cowork on the desktop app now opens a Chromium browser in a side panel when a task needs the web. Nothing to install. It is a separate browser, so Claude never sees your tabs, bookmarks or passwords.

You can bring logins across per site using an Import cookies step, from Chrome, Edge or Firefox on macOS, from Firefox only on Windows and Linux, and not at all from Safari. Anthropic’s wording on the sensitive categories is precise and worth quoting, because it has been widely misread: “Banking, email, and single sign-on sites stay unchecked by default.” Unchecked, not blocked. If your marketing stack authenticates through Google, Okta or Microsoft single sign-on, you can tick those deliberately, and Claude remembers them for future sessions on that machine.

A follow-up to a caveat we printed on 15 August. We told you then that browser agents remain vulnerable to prompt injection and that Anthropic’s own position was that safeguards “cannot eliminate” the risk. Anthropic has since published evaluations claiming no attack succeeded against Fable 5, Opus 5 or Sonnet 5 in the Cowork harness, even without probes and classifiers running. That is real progress on a real concern, and it does not make the concern disappear, because a browser agent that fills in forms is exactly the surface where an unsuccessful attack and an untested one look identical from the outside.

Why it matters / what to do: There is a clean decision rule here, and it is worth writing down for your team before someone works it out the hard way.

Use the built-in Cowork browser when the task does not need you signed in. Competitor research, public sites, pulling numbers off a published dashboard, filling in a form on a portal that has no integration. Use Claude in Chrome when the work is inside an account you are already authenticated to: your ad platform, your CMS, your email service provider, a vendor portal. If you already run the extension it stays your default, and if you do not, Claude now prefers its own browser. You switch under Settings, then Cowork, then Preferred browser.

The autonomy change deserves one deliberate decision rather than a shrug. Anthropic did not turn this on for everything and neither should you. Give it a job with a bounded blast radius first, something that reads and summarises rather than something that submits, and watch a full run before you leave it alone. And if you run an enterprise plan, the domain allowlist is the control worth setting now, because it is far easier to widen later than to explain afterwards.

3. LinkedIn put a number on the AI slop penalty, and almost everyone is misreading it

On 20 August, LinkedIn’s chief product officer Hari Srinivasan posted that more than a million people have used the “seems like AI slop” button since it launched on 30 July, and that members are “now experiencing 40% less views on what we classify as AI slop from just a few weeks ago”. LinkedIn has also started showing authors a message in Post Analytics reading “Some members told us this post seems like AI”.

Those two numbers are being quoted as cause and effect, and they are not related in the way people think. LinkedIn’s Amanda Purvis told Moneywise that the 40% reduction comes from the content classifiers LinkedIn shipped on the same day as the button, not from members pressing it. The button trains the classifier. It does not throttle your post. A million is the count of unique members who have used the flow, not the number of times posts have been flagged. Srinivasan hedged the figure in his own post, writing “I hesitate to give numbers, as everyone has a different network and feed.”

The practical difference is significant. If flags throttled posts, a competitor or an irritated reader could suppress your reach directly. They cannot. A classifier trained on aggregate signals decides, which means the way to stay out of its way is editorial rather than defensive.

And LinkedIn’s definition of slop is not what most people assume. Sam Corrao Clannon, who leads creator product, describes it as content “potentially sophisticated or polished in its presentation, but lacks substance”. The test is not whether AI was involved in writing it. The test is whether the post says anything.

Why it matters / what to do: Stop optimising for the wrong signal. Nobody needs to strip em dashes out of their LinkedIn posts or avoid AI assistance to stay safe, and a fair amount of advice circulating this week says exactly that.

Run the substance test instead, and run it on your last ten posts rather than your next one. Does the post contain a number, a named example, a specific mistake, or a position someone could reasonably disagree with? If a competitor could publish the identical post with their logo on it and nothing would look odd, that is the definition LinkedIn is describing.

If you see the “some members told us” message on a post, treat it as feedback on that specific piece rather than as a penalty on your account. And be careful about a quiet second-order effect: the message shows the author, which means your team members will start seeing it, and the natural response is to write more cautiously rather than more substantially. Those are not the same thing, and only one of them helps.

1. How an Anthropic field marketer sends a personalised Monday briefing to every sales rep

Adam Ward, a field marketer at Anthropic, was losing Sunday evenings to building slides for a Monday sales stand-up. He replaced it with an automated brief that reaches every rep individually, and Anthropic published the build on 24 August.

The setup. He connected Claude Code to BigQuery, which is fed by HubSpot, Clay and Salesforce, then built two templates: a “top three things for the week” for each rep, and a rollup for managers. Claude pulls each rep’s territory from the CRM and their account updates from Slack. He built the first working version in a one-hour hackathon.

The prompting approach is worth copying on its own. He briefed Claude “as a product manager who deeply understands the business problem” rather than writing a technical specification. He gave it a problem statement, a deliberately fake sample weekly update to act as a template, and a recording of himself explaining the problem out loud, which he then handed over as a transcript.

The part worth copying is how he hardened it. He piloted with ten account executives and converted every correction into an explicit written rule rather than a mental note:

• Claude invented plausible event URLs that led nowhere, so the rule became never invent a URL

• It recommended events to the wrong people, so the rule became check the contact’s title against the event’s stated audience requirements

• Columns in the source sheet kept being rearranged by colleagues, so the rule became read header rows dynamically rather than referencing fixed columns

By the end of the first week the prompt held nine content rules, each one the scar tissue of a specific failure. As the team grew, the prompt moved out of Google Docs and into GitHub so it could be versioned. It eventually ran unattended while he was on holiday, with no approval workflow, and he kept a full message archive so he could audit exactly what each seller received on any given date.

It then scaled sideways cheaply. It went to all account executive segments, then to business development reps by changing only the CRM relationship field, then to customer success and alliances.

The claimed result: registrations for an executive dinner doubled in a week, because the right reps had the right event in front of them on a Monday morning rather than buried in a shared calendar.

Why this leads the section this week. Claudeforce promises marketers CRM-grounded AI at some point in late 2026. This is a marketer who built it himself in an hour, using tools that already exist, and the nine-rule hardening loop is the transferable part regardless of your stack.

2. Buyer research across three tools, because one tool cannot do it

Published 23 August, eighteen minutes, and it answers the obvious objection directly: why not just do all of this in Claude?

The answer is that Gemini Notebook is deliberately tied to the sources you load into it, and Claude is not. That constraint is exactly what you want during synthesis and exactly what you do not want during ideation, which is why the workflow moves between three tools rather than staying in one. Perplexity for research, especially Reddit. Gemini Notebook for synthesis across sources. Claude for final thinking and deliverables.

Step one, two Perplexity prompts in this order:

1. “What Reddit groups might [commercial construction companies] find most useful for understanding their buyers?”

2. “Can you please find a bunch of recent threads from these that might be helpful for understanding buyer behaviour?”

Step two. Paste the thread URLs into Gemini Notebook as website sources. If Perplexity’s reply will not copy cleanly, ask it for raw URLs only, or pass the response through Claude to strip the formatting.

Step three. Apply IDEO’s seven-step design thinking framework as the prompt: frame the question, gather inspiration, synthesis, idea generation, prototyping, testing, storytelling. The demonstrated effect is that it reframes “how do we get more leads” into “what are the friction points and where is the buyer anxious”, which is a materially better question to bring to a content plan.

Step four, the persona prompt verbatim: “Please create evidence-based professional personas from this Reddit corpus. Segment these contributors by context, goals, capability, constraints.” He caps output at three to five segments deliberately rather than accepting twenty or thirty. His construction example produced a facilities manager, a high-stakes institutional developer, a hungry subcontractor and a hands-on operator.

Step five is the Gemini Notebook technique most people never find. Save the note, add it, then convert the note to a source, select only that source, and generate a mind map from it. Every node then traces back to its exact citation, which is the difference between a diagram you can show a client and one you cannot defend.

Step six, synthesis. He uses a problem tree, which separates a central problem from its causes and its downstream effects. The prompt is “From this source material, what might be some good candidates for a problem tree?”, then you request maximum detail on the one you pick. He also names ecosystem maps, journey maps that start at first problem awareness rather than at vendor contact, affinity mapping, the five Ws, and assumption mapping.

Step seven, the deliverable, built in Claude. “Please take the following material and create a nicely designed HTML page to illustrate this problem tree”, with the context separated by triple quotation marks so Claude knows where the instruction ends and the material begins. Two upgrades he demonstrates: give Claude the client’s website URL and ask it to match their branding, and run /loop asking Claude to act as a design expert, which he did three times. Sonnet 5 was sufficient throughout, and he reserves Fable 5 for heavy information crunching.

Step eight. Ideation moves back to Claude using SCAMPER: substitute, combine, adapt, modify, put to another use, eliminate, reverse.

Three caveats he states outright, and they save time. Building a Claude Design system properly takes a couple of hours of loading assets and examples, and in his words “it doesn’t work well if you just try to do it quickly, I learned that the hard way.” Do not let the design get so elaborate that it pulls attention off the content, which he calls a trap a lot of people fall into. And HTML files are easier for AI to create and read than PDFs, Word documents or Google Docs, so making HTML or markdown your default format for client communication removes a genuine amount of formatting work.

He also notes the commercial angle: this framing and discovery stage used to take months at an agency, and it is now cheap enough to run free or reduced as part of a proposal.

He sells cheat sheets and coaching alongside the video, which is worth knowing, though the workflow stands up without any of it.

3. Getting AI to recommend your brand, in three moves

Published 26 August, sixteen minutes. AEO means answer engine optimisation, which is getting your brand named inside an AI-generated answer rather than ranked in a list of links.

Move one, fix your brand identity before anything else. Ask ChatGPT, Gemini and Perplexity the same question: “what does this brand do and who is it for”. If the three answers are vague, wrong, or simply different from each other, that is your first job and nothing else you do will work until it is done.

The fix is unglamorous. Standardise the description word for word everywhere AI looks: your website, your social profiles, your Google Business Profile, directory listings. Her client example is the one that makes the point stick. AI had classified a retailer as a manufacturer, so the brand never surfaced when buyers asked where to buy. No amount of content would have fixed that, because the content was not the problem.

Her framing is useful: it works like a reference check, and if every referee describes you the same way, your odds improve.

Move two, own one lane with genuine depth. Publishing more does not win. She cites Google’s information gain patent, under which a page that repeats what the top-ranking results already say scores close to zero regardless of how well it is written. For most B2B teams this means fewer pieces with original data, first-hand testing or a real position, and considerably less of the middle-of-the-road explainer content that currently fills content calendars.

Move three, engineer mentions you do not own. She cites AirOps that 80 to 90% of what AI engines reference comes from sources the brand does not control, and Google’s own quality rater guidelines, which instruct raters to be sceptical of a site’s claims about itself and to look for independent corroboration. Her Reddit caution is worth repeating to anyone about to get enthusiastic: contribute properly and amplify, do not promote directly, or you will be banned and lose the source entirely.

Her closing three-step is the spine of the whole thing. If you were starting this week: run the brand audit, map your buyer’s questions and pick one lane, then start building mentions you do not own.

She is partnered with HubSpot and demonstrates its AEO product throughout, so treat the tool choices as illustration. The three moves are tool-agnostic and none of them require buying anything.

1. Build a landing page in Claude and send it straight into HubSpot

Announced 26 August. Create a landing page in Claude Design, then push it to HubSpot as a draft using Share, then More formats and apps, then Send. Claude references your products, brand context and existing content when building it. You need the HubSpot connector for Claude and a Pro, Max, Team or Enterprise plan.

Four limits worth knowing before you rely on it. Landing pages are the only supported asset type. Asking Claude in chat to update the page will not push the change, because only the Send button does that. Re-sending after you have published creates a new version rather than overwriting the live one. And if you connected the HubSpot connector before 29 July, most of the newer tools will not appear until you disconnect and reauthorise, because the new permission scopes only load on a fresh authentication.

This is the most immediately usable item on the list for anyone running a HubSpot stack, because it removes the design-tool-to-CMS handoff entirely.

2. Gemini Notebook can now read the books you have paid for

On 27 August Google launched Expert Intelligence, which lets you add eligible ebooks you have purchased from Google Play Books into Gemini Notebook as a source, then question them or generate infographics, audio overviews and quizzes from them. It launches with more than 100,000 titles from Bloomsbury, De Gruyter Brill, Johns Hopkins University Press, Macmillan, O’Reilly Media and Penguin Random House, plus curated notebooks built with more than fifteen authors including Steven Pinker and Michael Pollan.

You must own the book. Share a notebook with a colleague and they are prompted to buy their own copy, otherwise that source is unavailable to them. Check the Tools badge on a title’s Play Books page to see whether it is eligible.

The reason this belongs in a marketing newsletter is the roadmap line rather than the books. Google says third-party subscriptions, business research reports and textbooks are coming next, along with availability in the Gemini app and AI Mode. Analyst reports as a queryable notebook source is the version marketers will care about, and it is a materially different model from the usual argument about AI summarising content it never paid for.

3. X shipped an ads server that can write to live accounts

On 24 August X launched an MCP server exposing twenty-three advertising tools to any compatible AI client, including Claude Code and Grok, with no custom integration work. Thirteen are reads, analytics and targeting lookups. Ten are writes, including creating and updating campaigns, adding targeting and promoting posts, against production accounts funded by real payment methods.

The safety design is worth stealing regardless of whether you advertise on X. Campaigns and line items created through the server always arrive paused, and activation requires a separate explicit call, so a hallucinated targeting criterion costs nothing until something deliberately switches it on. Authentication uses your own OAuth token, so an agent only ever sees accounts you can already access, and there is one grant per application and user pair, which stops an agency spinning up parallel agents against a client’s credentials.

Connect with the read and offline access scopes and omit write access. Read-only is the right first version of this for almost everyone.

4. Gemini Omni 1.1 Flash extends a scene without losing the character

Released 27 August. The model now analyses up to ten seconds of preceding footage, where previous versions saw only the final second, so scenes extend in ten-second increments up to a cumulative forty seconds while holding character, lighting and narrative consistent. You can specify first and last frames, and drop in up to three seconds of reference video to keep a character stable.

The cost mechanic is the useful part for anyone iterating. Draft at 360p up to 60% faster and at roughly a third of the price, then upscale the version you actually want to 1080p or 4K. API pricing runs from $0.03 per second at 360p to $0.30 at 4K. Available to Google AI Plus, Pro and Ultra subscribers in Google Flow. The old gemini-omni-flash-preview endpoint retires on 30 September.

5. Google’s Multimodal Video Creation reaches general availability

In the twelfth monthly Demand Gen Drop, published 27 August, Google’s storyboard-to-finished-asset workflow went generally available. It produces horizontal and vertical cuts in a single pass, powered by Gemini, Veo and Nano Banana, and was first shown at Google Marketing Live in May.

One caveat to carry into any business case: Google’s headline “30% increase in conversions” is its own internal experiment data, not an independent result.

1. AI search took 10.5% of the traffic and gave back about 4% of it

Digiday published Brainlabs analysis on 24 August covering fifty-four advertiser clients across nineteen sectors, using fourteen months of Google Analytics data from January 2025 to April 2026. The sample is 24 US firms, 24 British brands and six elsewhere, with 29 of them employing at least a thousand staff.

The inflection point was September 2025, when Google’s AI Overviews began appearing on at least 30% of US results pages. After that:

• Organic sessions fell 10.5%, from 140.1 million to 125.4 million, with 46 of the 54 clients declining

• Sessions from AI platforms rose 163%, reaching roughly 200,000 a month

• AI-driven key events rose 335%

• Key events occurred at 1.5 times the rate for referrals from ChatGPT, Copilot, Gemini and Perplexity compared with organic search

Three caveats, because this study is going to be quoted badly all autumn.

It is one agency’s client portfolio, not an industry benchmark, so instrument your own data rather than repeating these percentages in a board paper.

It is emphatically not a wash. The sample lost roughly 4.8 million monthly sessions and gained about 200,000, which is a replacement rate of around 4%. “Fewer but better visitors” is true and nowhere near break-even, and anyone presenting this as reassuring has stopped reading too early.

And Brainlabs counts scrolling to the bottom of a page as a key event, alongside purchases and newsletter signups. Anyone quoting “1.5 times the conversion rate” is overstating what was measured.

The B2B-relevant detail is the sector split. Fitness, fintech, insurance and consumer packaged goods saw the largest organic declines, while retail, beauty and entertainment saw the smallest declines and the strongest AI referral lift. The considered-purchase categories nearest to B2B fared worst, which is consistent with AI Overviews triggering hardest on informational and educational queries.

2. Only 21% of marketers say their CRM data is ready for AI, and the seniority gradient is alarming

Validity surveyed 500 marketing professionals across the US, UK, Brazil, Australia and New Zealand for its State of CRM Data Management report, published 25 August.

Only 21% say their CRM data is very well prepared for AI. 62% report losing revenue to poor CRM data quality. Two thirds increased the volume of marketing decisions delegated to autonomous agents over the past year.

The finding worth sitting with is the gradient by seniority. 78% of C-suite respondents and 92% of SVPs and VPs say they have acted on an AI recommendation they later suspected was wrong because of bad underlying data. Among individual contributors the figure is 41%. The people furthest from the data are the most likely to act on it and the least likely to catch the error, which is precisely the wrong way round for a technology being sold on executive decision support.

One more number that explains a lot about internal reporting: 67% of C-suite respondents admit campaign data is sometimes manipulated to look better to leadership, against 38% across the whole organisation.

Read this next to the Claudeforce story at the top of this issue. When an assistant reasons over your CRM rather than a person reading a dashboard, nobody silently discounts the field they know is wrong.

3. 92% of B2B teams are doing GEO and fewer than 15% have anyone in charge of it

GNW Consulting and Demand Metric surveyed 225 B2B marketing and revenue leaders for the 2026 State of GEO in B2B Marketing, published 24 August. GEO means generative engine optimisation, which is optimising to appear inside AI-generated answers.

92% are already experimenting with or operationalising it, and 78% of those investing report measurable return. 22% say AI-driven traffic already exceeds 5% of total site visits, which is well above the sub-1% figure still quoted in most industry commentary.

But fewer than 15% have a dedicated owner for it. And while 88% of agencies claim to offer GEO, 37% admit their offering is loosely defined.

The ownership gap is the story, and unusually for a research finding it is answerable this week. Naming a person costs nothing and is the difference between a discipline and a set of opinions.

4. 40% of B2B buyers trust an AI summary of your product slightly more than your own documentation

Ipsos surveyed 288 US business decision-makers involved in indirect procurement, published 24 August. 70% use AI during vendor selection. 42% use it to build an initial supplier list they would not otherwise have considered. 40% say they trust an AI-generated product summary slightly more than the vendor’s own documentation. Nearly half also use AI to evaluate proposals after receiving them.

Ipsos frames it neatly: AI assembles the guest list, while trusted customers and colleagues still decide who gets invited in. That maps directly onto Grace Leung’s playbook above, because if AI is building the longlist then being described accurately by AI is a shortlisting requirement rather than a marketing nicety.

One caveat to carry: the fieldwork ran between 20 May and 1 June, so publication is recent but the data is three months old.

1. There was no August core update.

Several outlets reported on and after 26 August that Google had begun rolling one out. Google’s Search Status Dashboard lists exactly one August ranking event, the August 2026 spam update, which ran on 18 August for two days and sixteen hours. The last confirmed core update was in May.

2. Google has now bounded the Search Console bug we flagged last week.

Its data anomalies page records a logging error that understated impressions in the Generative AI performance report from 13 to 17 August, with a separate single-day Discover error on 13 August. Logging only, not lost visibility, and the window now has a firm end date so you can annotate it.

3. ChatGPT Computer History is finally available in the UK.

OpenAI lifted the regional block on 20 August for the UK, EEA and Switzerland on the macOS app for Pro, Business and Enterprise. It is off by default, requires Memories to be on, records interaction events rather than screenshots or audio, and deletes event files after 48 hours.

4. Ask Advisor is getting shared memory across Google’s marketing tools.

Carrying goals, past decisions and performance data between Ads, Analytics, Merchant Center and Google Marketing Platform. Announced on 24 August as coming soon rather than shipped.

5. Google will not enforce its site reputation abuse policy in the EEA,

A first for one of its spam policies. Manual actions are replaced by a system that ranks affected site sections independently. Announced 28 August.

6. Google is folding Tag Manager into Google tag.

Every Google tag becomes a full Tag Manager container with debugging and version control, with no change to on-page behaviour. The no-code visual tagging is in beta and currently covers only purchase conversions in Google Ads, so the form submission tracking most B2B teams actually need is still promised for later in the year.

7. Some Meta advertisers report losing placement controls from 25 August, with value rules offered as the substitute.

Value rules cap a bid decrease at 90%, so a placement can be made expensive to win but never excluded outright. Meta has announced nothing and its own help pages still describe manual placement selection as available, so treat this as a sighting rather than a confirmed change.

8. OpenAI cut GPT-5.6 Sol pricing by more than 20% from 21 August…

…to $4 per million input tokens and $20 per million output, guaranteed until 21 November. Separately, o3 left the ChatGPT model picker on 26 August, closing the ninety-day sunset announced in May, though it remains available through the API.

First, memory became the product this week, and three vendors landed on that within seventy-two hours of each other.

Anthropic unified memory across Claude chat and Claude Cowork on 25 August. Google announced shared memory across Ads, Analytics, Merchant Center and Google Marketing Platform on 24 August. Claudeforce’s onboarding reads a seller’s context across Salesforce, Slack and every connected tool to assemble a tailored dashboard before they ask it anything. None of these are model improvements. All three are the same bet, which is that an assistant that already knows your business beats a cleverer one starting from a blank page. That bet has a consequence most marketing teams have not thought about yet. If what the assistant remembers determines the quality of what it produces, then the contents of that memory are a marketing asset in the same way a brand guideline is, and it deserves the same treatment: someone owns it, someone reviews it, and someone notices when it goes stale. Right now, in most organisations, it accumulates by accident from whoever happened to be typing. Anthropic’s memory is inspectable and editable topic by topic, which makes curating it a genuine option rather than an aspiration, and the fact that almost nobody is doing it yet is the opportunity.

Second, the interface is being deprecated, and this week made that unusually literal.

Marc Benioff’s line at the Claudeforce announcement was “the UI is the AI”. In the same week Claude stopped asking permission before acting in your browser and shipped a browser of its own so it no longer has to borrow yours, X exposed ten tools that write directly to live ad accounts, and Google began moving its marketing tools behind a conversational layer that carries context between them. Say it plainly and it is not a small change. Your website, your CRM and your ad platform are all becoming things that software operates rather than things people look at. For marketers that reframes two jobs at once. The first is that the quality of your data and the tightness of your permissions stop being back-office concerns and become the things that determine whether the output is any good, which is exactly what the Validity research measured and found wanting. The second is that being legible to a machine starts to matter as much as being persuasive to a person, which is the whole premise of Grace Leung’s playbook above. Neither of those is a new skill. Both are old disciplines that have just become load-bearing.

REMINDER: MY BEST BUNDLE YET FOR MARKETERS:

Check Out This Limited Time Offer

Tell me: What’s your biggest takeaway from this weeks news from the world of AI Marketing? Love to hear your thougths in the comments.

Leave a comment

Thanks for reading the Marketing with AI Weekly Roundup! If you enjoyed this read, the best compliment I could receive would be if you shared it.

The Marketer's Guide to Setting Up Claude Projects (August 2026)

Thu, 27 Aug 2026 10:07:32 GMT

Every Wednesday, one essay on what's actually working in Marketing with AI - tools, workflows, and mindset shifts you can apply immediately.

Subscribe now

The Marketer’s Guide to Setting Up Claude Projects (Aug 2026)

Something a bit different this week. Instead of a long read, I’ve recorded a video.

And here’s why…

The question I get asked more than any other and the topic I spend the most on in my team trainings, is, how do I set up my Claude Projects properly.

Properly being the key word here.

But the better question is, not just how do you set up one, but how do you structure it so it works for the kind of marketing you actually do every day.

And that’s far easier to answer by showing you, than describing it. So I’ve walked through it on screen this week, including a few slides from inside my Claude for Marketers Masterclass.

One quick thing before you press play. There’s a discount code for anyone reading or watching this: Discount code PROJECTS takes $50 off the course at robertgillespie.me/masterclass.

The video is right here, but underneath is a quick overview of what I cover, in case you want the gist before you watch.

There Are Two Kinds of Claude Projects Now

If you set one up a few months ago and haven’t looked since, you need to really understand the key differences as they will impact older Projects too.

Claude now has local projects and cloud projects, and they work in genuinely different ways.

A local project connects to real folders on your computer. Claude reads the folder, every subfolder inside it, and every file. Nothing gets uploaded. The catch is that it stays on that machine, so it won’t sync and it won’t follow you to your phone.

A cloud project has no folder. You upload files into it manually and they sit in the cloud, which means you can pick the project up in the browser, on the desktop app, or on your mobile.

This is why the split exists.

Cowork originally only ran on local folders and files, and the cloud option is what now lets it work in a browser and on mobile as well.

Telling them apart is easy once you know.

In your project list, a local project carries a small laptop icon and a note saying it works with local files and can only run on that computer. No icon means it’s a cloud project.

Setting Up Each Type of Claude Project

For a cloud project: New Project, give it a name and some instructions, and leave the “Use a folder” option alone.

Once it’s created, the context panel on the right is where you upload your Word docs, spreadsheets and anything else Claude needs.

For a local project: New Project, then click Use a folder and pick the folder you want.

Two things worth knowing here.

Connecting a folder overrides whatever name you typed in, so don’t waste time naming it first, it takes the folder name. And the first folder you connect is permanent. You can add more folders afterwards and remove those freely, but not the original one, because Cowork needs a fixed place to put the documents, decks and spreadsheets it builds for you.

I prefer local wherever it’s practical.

It’s the more powerful of the two, because a folder brings its whole structure with it. If you move between machines constantly, cloud is the trade-off you will have to make.

The reason project ‘structure’ matters so much is what happens when it’s missing.
Upload ten loose files to a cloud project and it’s manageable. Upload fifty and you’ve got fifty single files with no brand folder, no ICP folder, no templates folder.
Claude has to hunt through them every time, and hunting costs you in terms of tokens and output quality.

How to Structure Your Claude Project Folders

This is the part that makes the biggest difference, and it’s where I jump into my Claude for Marketers Masterclass course slides in the video.

There are two project set up patterns you need to grasp.

The focus space. A channel, a campaign or a client. Tight and fast, holding everything Claude needs for one job and nothing it doesn’t. A LinkedIn project, a summer campaign, a client called Google.

The department or business folder. Brand tone of voice, ICP, positioning, product information, templates. The core context that every piece of work draws on, sitting in one place.

The two can work together, and that’s the point.

You can connect the same folder to several different projects. So your winter campaign project holds the winter campaign assets, and you connect the brand folder to it as well.

Your summer campaign does the same. Your LinkedIn space does the same. You maintain your brand voice and ICP once, and every focus space reads from it.

That’s a lot harder to pull off with manual uploads, which is the real argument for going local when you can.

Claude.md and Memory.md

Once the folders are in place, two files do most of the heavy lifting.

Ask Claude, in the chat, to create a CLAUDE.md and a MEMORY.md at the root of the project. Root just means the top level of the project space, nothing more complicated than that.

I describe the CLAUDE.md file as an Airbnb welcome guide. Walk into a well-run Airbnb and there’s a welcome guide telling you the spare towels are in the hall cupboard, the coffee’s in the top drawer, mugs above it. You don’t open every cupboard to find out.

CLAUDE.md is that guide, written for Claude.

It holds a folder map. If I ask about brand, it knows to look in the brand voice folder. Past reports are in examples, and templates are where the templates live.

That has two real benefits, and both are super important.

Output improves, because Claude isn’t guessing where important information sits.
And your usage tokens drops, because Claude isn’t opening every cupboard looking for the mugs.

If you regularly hit your limits, this is one of the first things to fix.

MEMORY.md is the running record of what is going on inside the project.

Claude updates it automatically as you work, so decisions you made, corrections you gave it, a great testimonial a client sent through, the connectors relevant to that project, all of it gets noted.

The next session starts where the last one finished instead of from nothing.

The two prompts to run on any new project

First, get Claude to look around:

“I’ve just created this project. Read every file in the folder. Summarise what you know about this workspace, what’s here, what I probably use it for, and what instructions you’ll follow.”

Read what comes back. If it’s telling you what’s in your brand voice folder and what’s in your ICP folder, it has the full picture.

Then, once you’re happy:

“Create a CLAUDE.md and a MEMORY.md at the root of this project.”

It writes both files and puts them in place. You don’t need to touch them.

Going deeper

The video above goes into all of this in more detail. So please watch it if you want to master your Claude Project set up.

Claude Projects, for me, are key to everything you do with Claude.

If you want the full version of this and much more when it comes to Claude for Marketing, it’s covered properly in my Claude for Marketers course at robertgillespie.me/masterclass, including the Airbnb approach to project folders in a lot more detail, with examples you can copy.

The course covers Cowork, projects, skills, connectors and plugins, marketing workflows, Claude for Excel and PowerPoint, and the Claude Design for Marketers module I added last month. A Claude Code module lands in the next week or so, and How to Build an AI Marketing Team follows in late September.

Use the code PROJECTS at checkout for $50 off.

And if written pieces are more your thing than video, my Inc. column has a couple of new articles live over the past week.

This week’s challenge: open your busiest Claude project and ask it to read every file and summarise what it thinks the workspace is for. If the answer is vague, that’s your folder structure telling you something. Fix the folders, then get it to write your CLAUDE.md.

Any questions at all, leave them in the comments and I’ll come back to you. Thanks for reading Marketing with AI. If you found this useful, sharing it with one other marketer is the best compliment I could receive.

Need 1:1 Help to Embed Claude into Your Marketing?

Let’s talk →

Marketing with AI Weekly Roundup: Google's Bidding Change Lands, the IAB Rules AI Copy Needs No Label, and Your August AI Visibility Report Is Wrong

Robert Gillespie — Sun, 23 Aug 2026 10:03:05 GMT

Welcome to issue #48 of the Marketing with AI Weekly Roundup. Published every Sunday, it shares the top stories, playbooks, data and quick hits from the world of Marketing with AI over the last 7 days.

Subscribe to receive the week's most important AI marketing stories direct to your inbox.

Subscribe now

MY BEST BUNDLE YET FOR MARKETERS (LIMITED TIME OFFER):

Check Out This Limited Time Offer

Last week we led with Anthropic marking everything Claude writes with an invisible watermark. It turns out that was only half the story.

In the same week, and we missed it, Google made its own visible watermark optional on every image, video and song the Gemini app produces, while quietly leaving the invisible one switched on. Two companies, opposite directions, same destination.

On Monday, Google’s Smart Bidding change went live on budget-limited campaigns, the one we made our lead story a fortnight ago. On Tuesday, the IAB published the second version of its AI disclosure framework and drew a line most marketers will be surprised by, because AI-written copy sits on the no-disclosure side of it.

And running underneath the whole week, Search Console’s Generative AI report has been wrongy understating your impressions since 13 August, which means a lot of teams are about to present a decline that never happened.

Elsewhere, Microsoft turned on AI Max globally and started removing the bid ceiling, Reddit lost 86% of its ChatGPT citations in four days, and Gmail finally switched off the domain reputation score you were told to stop relying on two years ago.

Below: four stories to act on, three workflows worth stealing with the real costs attached, and one deadline that closes nine days from now.

Let’s get into it…

1. Google’s bidding change went live on Monday, and anyone showing you what it did is reading noise

On 17 August 2026, Google began rolling out the change we led with on 8 August. Campaigns carrying a “Limited by budget” status while running Target CPA or Target ROAS now optimise toward the target you set, rather than settling wherever the constrained budget happened to leave them.

The mechanics, stated plainly. A Search campaign carrying a £10 target cost per acquisition that has been delivering at £5 will now drift upward toward £10. Total spend does not rise, because the budget still caps it. Volume falls instead, and your reported cost per conversion climbs to meet the number you typed into a settings field, possibly years ago.

What Google’s framing leaves out. Google describes this as making performance more predictable. The practitioner reading, and it is supported by Google’s own documentation, is that the old behaviour was never generosity. Bids were being quietly lowered so that a capped daily budget lasted the full day. That is why a £10 target so reliably produced a £6 result. Yesterday’s overperformance was throttling, and this change removes it.

Three conditions must all hold. Ginny Marvin, Google’s ads product liaison, narrowed the scope more than most coverage has admitted. The campaign must carry a target, must be limited by budget, and must currently be beating that target. Miss any one and nothing happens to you. She also confirmed for the first time that targets set at ad group level are in scope, which matters because lead generation accounts are full of them.

It runs in one direction only. Asked directly whether campaigns missing their targets get a corrective push in the other direction, the answer was no. Google will pull an overachieving campaign up to its target and will not push an underachieving one down. And below roughly seven conversions, Google calculates no recommended target at all, so thin-volume B2B campaigns get no guidance from the tool that is about to change their delivery.

A second change landed the following day, and almost nobody covered it. From 18 August, Google automatically assigns a customer type to every eligible conversion-based Customer Match list that has not been classified manually. The unclassified state no longer exists. Those labels feed straight into Smart Bidding, so if you run a new-customer acquisition goal, the label decides who counts as new. For a B2B account where a conversion is a demo request rather than a purchase, Google’s inference that someone is an existing customer is very likely wrong, and a mislabelled list means paying acquisition bids for people already in your pipeline. Google exposes nine customer-type labels if you set them yourself. Its automatic assignment uses three. The opt-out window has closed, though you can still relabel. This affects only lists Google builds from Enhanced Conversions data, so manually uploaded Customer Match lists are untouched.

The third piece is a deadline. Google has disclaimed its own forecasts. Its guidance says Performance Planner and other budget and bid forecasts may contain inaccuracies during the transition, and advises caution from 17 to 31 August. That window is open right now and closes on Monday week.

Why it matters / what to do: The most useful thing you can do this week is nothing dramatic. There is no credible before-and-after data anywhere, from anyone, and there is a structural reason for that. The rollout is staged over several weeks rather than flipped in a day, and Google’s own guidance is to let one or two conversion cycles run before reading results. For a B2B account with a two to three week sales cycle and weekly offline conversion imports, that means four to six weeks, not five days. First reliable readings land somewhere between mid-September and mid-October, which is to say in the middle of your Q4 planning.

So record a baseline instead of reacting to one. Export your current target-versus-actual performance now, while you still have a clean pre-change read. Then filter for exposure: flag campaigns where actual cost per acquisition is 20% or more below target, or actual return on ad spend is 20% or more above. Use a twelve-month lookback on the budget-limited flag rather than today’s status, because Google’s own notification trigger covers any campaign that hit its cap at any point in the last year, and filtering on today under-reports your exposure badly.

Rank what you find by money at risk rather than by the size of the gap. A 20% gap on a campaign spending £80,000 a month matters more than a 2x gap on one spending £400. When you do move a target, move it 5 to 15% at a time, a week apart, and never change target and budget in the same week or you will not know which one did what. One more trap: the Bid Target Adjustment Tool offers to reset your target to recent actual performance. Applying that blindly locks in whatever number the budget cap happened to produce, not what your business can actually afford per lead. Keep a target only if you can say out loud why that specific number is the goal.

And do not build a Q4 case on Performance Planner numbers before 31 August. Google has told you they may be wrong.

2. The IAB says your AI-written copy needs no label. Your AI-generated images do.

On 18 August 2026 the Interactive Advertising Bureau published version two of its AI Transparency and Disclosure Framework for advertising, replacing the version it issued in January. It is voluntary and carries no enforcement, but it is the closest thing the industry has to an agreed answer to a question creative teams are being asked constantly, which is what actually needs a label.

The test is a single question. Does the AI involvement create a material risk that consumers will be misled? Three criteria sit underneath it: deception potential, material impact, and expectation alignment.

What needs disclosing. Images generated from AI prompts, whether text to image or image to image, and regardless of how much a human refined them afterwards. AI-generated video. Synthetic voices of deceased people, and synthetic voices of living people placed in scenarios that did not happen. Photorealistic AI influencers and avatars. Digital twins of the deceased, and digital twins of living people in events that never took place. AI chatbots get a label too, though theirs reads “AI-powered” rather than “AI-generated”.

What does not. Routine post-production, meaning colour correction, lighting, background cleanup, blemish removal and contrast. Clearly stylised or fantastical imagery. Upscaling and de-noising. Internal workflows. Background music and standard audio enhancement. Generic synthetic voices, and authorised synthetic voices of real people. Obvious cartoon avatars. And, the one that will surprise most people, text and copy.

That last exemption is the whole story for a B2B team. AI-written headlines, slogans, product descriptions, email subject lines, blog posts and translations all sit on the no-disclosure list. That is precisely what most marketing teams are actually using AI for, day to day. The framework treats an AI-written subject line as a non-event and an AI-prompted hero image as something requiring a visible mark, even if a designer spent an hour reworking it.

The mechanics, if you need them. Either the Unicode sparkle at U+2728 in monochrome, or a plain “AI-generated” text label. Either satisfies the framework. Text labels need WCAG AA contrast of 4.5 to 1, plus alt text and captions. Video labels go on the first frame and persist. Audio disclosure is spoken before or immediately after, and repeated once past sixty seconds. All of it pairs with C2PA metadata carrying two new assertions. There is a 24-month roadmap and a recommendation to name an AI Disclosure Lead within 60 days.

Why it matters / what to do: This landed sixteen days after the EU AI Act’s Article 50 transparency obligations became applicable on 2 August. Article 50 requires disclosure but does not prescribe an icon, and the EU Code of Practice finalised in early June is itself voluntary with an illustrative icon that is not finalised. So the IAB framework is currently the most concrete guidance available, and it will function as a de facto standard whether or not anyone adopts it formally.

Take the list above and turn it into a one-page internal rule for your team this week, because the alternative is every designer and copywriter making the call individually. The practical shape of it: your copy is fine, your imagery is not, and refining a generated image by hand does not exempt it.

One caveat to hold onto. The IAB is the advertising industry’s trade body, and a framework that argues against blanket labelling, partly on the grounds that labelling costs clicks, has an obvious interest sitting behind it. Treat it as a well-researched industry position rather than a neutral one.

3. Google made its visible AI watermark optional, and handed you the disclosure liability

We missed this last week, and it belongs directly alongside our own lead story, so here it is.

On 14 August 2026, Google began rolling out a setting that removes the visible watermark from content its AI models generate. It covers images from Nano Banana, video from Omni and music from Lyria. The setting lives at Settings, then Media Watermark, in the bottom left of gemini.google.com, and in the Flow video editor. Support in Google Search is described as coming soon.

What stays. In Google’s own words: “This setting only controls the visible watermark. It doesn’t affect SynthID watermarks or Content Credentials for media you create with Gemini Apps.” SynthID is the imperceptible signal Google embeds in generated media, and C2PA Content Credentials are the metadata standard attached to the file. Both remain, whichever way you set the toggle.

Availability, because this is where most coverage got it wrong. Google’s support page says: “If you are in India, South Korea, or Vietnam, you’ll only see this setting if you have an AI Ultra subscription. Otherwise you’ll automatically see visible watermarks applied to all visual media you create with Gemini Apps.” There is no EU restriction, so UK and European marketers get the setting normally. Work and school accounts are a different matter. Google is explicit that if your access comes through one, “this setting is not available and you will continue to see the watermark”, which will catch a lot of B2B marketers generating from a company Google account.

The watermark remains on by default, so this is an opt-out rather than something that changed under you. It applies only to files created after you change the setting.

How it was announced tells you something. Not through a blog post. Josh Woodward, Google’s VP for Gemini, posted it on X alongside Gemini 3.7 Flash reaching the app. The sharpest comment came from a reply under that post, pointing out that a menu item labelled “Media watermark: off” is misleading when what it means is visible off, invisible still on.

Why it matters / what to do: Read this next to the previous story and next to our own lead a week ago. The IAB has ruled that AI imagery needs a label. Anthropic has made text marking invisible and compulsory. Google has made image marking invisible and optional. The direction of travel is identical in all three cases, which is that the signal telling someone content is synthetic is moving from something a reader can see to something only a detector can read.

The practical consequence is a transfer of liability. When Google stamped every image, the disclosure was handled for you. Remove the stamp and the duty to disclose under Article 50 sits entirely with whoever publishes the asset, which is you. If a piece of generated content misleads someone, Google can point at the SynthID mark still sitting inside the file.

So decide your policy this week rather than leaving it to whoever is generating the asset. Two questions settle it. First, which account are your people generating from, because a work account keeps the watermark whether you like it or not. Second, if you do switch the visible mark off, what replaces it in your own process, given the file still carries provenance data you cannot remove and a disclosure obligation you now own.

4. Reddit’s ChatGPT citations fell 86% in four days

Between 14 and 17 August, Reddit’s share of citations in ChatGPT Search fell from a 3.83% average, measured across 18 July to 7 August, to 0.52%. That is a drop of 86.4%, and it happened over a long weekend.

The contrast is the actual lesson. Reddit’s citation share in Google’s AI products barely moved over the same period, declining gradually from around 2.5% in early July to roughly 2.1% in August, with AI Mode following a similar curve. So this is not Reddit losing relevance. It is one platform changing how it sources answers, without announcement, over four days. The timing loosely coincides with ChatGPT Search changing its query fan-out behaviour on 8 August, though nobody has established a causal link.

Two caveats, and both need to travel with the number. Promptwatch, which produced the data, says plainly that it shows when the shift occurred and not why, and cannot rule out a data collection issue on its own side. Separately, and this matters: Promptwatch is owned by Semrush, which also owns Search Engine Land, the outlet that published the finding. The same corporate parent owns the measurement platform and the publication reporting on it. That does not make the number wrong, but you should know it before you quote it in a deck.

Why it matters / what to do: A great deal of the answer engine optimisation advice written this year reduces to “get mentioned in Reddit threads”. If your visibility plan leaned on that, the ChatGPT half of the payoff may have evaporated in a weekend, and you would have no way of knowing because nothing was announced.

The durable takeaway is not about Reddit. It is that citation sources are platform-specific and can change overnight with no notice, which makes any single-platform AI visibility strategy fragile by construction. If you are tracking your brand’s presence in AI answers, check that you are tracking it across more than one engine, and check that your content distribution is not concentrated in one third-party property whose standing you do not control.

1. The marketer’s AI stack that one prompt builds (Kieran Flanagan)

Published 18 August 2026 on Marketing Against the Grain. Disclosure: this is a HubSpot show, and the episode carries a HubSpot advert. The framework stands on its own regardless.

The premise is that the advice to “become a systems builder” is everywhere and almost never accompanied by an actual system. So he lays out his own, as five layers, and the ordering is the part that matters.

Layer one, intelligence. Everything else reads from here, and nothing works without it. His words: “There is no system if you do not have context, you don’t have intelligence, and your skills are reading from that intelligent context layer in a meaningful way.”

Six context files, each one a markdown document: goals, ICP (your ideal customer profile), competitors, positioning, team, and execs. Goals anchors everything to what you are actually trying to achieve this quarter. Team and execs exist so that skills built on top can handle managing your people and managing upward.

The routing detail is what separates this from a folder of documents. Skills pull only the files they need, not the whole layer. His reasoning: “Imagine a world where every skill was just told to read everything in here before executing it. That’s a really clunky thing. The skill has to hold a lot of context, it’s going to get confused, it needs to have a ton of tokens.” So the people brief reads team and goals. The voice-of-you skill reads positioning and goals. Nothing reads everything.

He puts two maintenance tools in the same foundational layer, which is unusual and sensible. The first is an AI coach, a skill you run weekly that looks across your own usage and returns four things: your three weakest prompt habits, the three skills you should build based on repetitive requests you keep making manually, the context files most in need of an update, and one capability you are not using that fits how you actually work. The second is Record a Skill, the Claude desktop feature we covered on 25 July, used here to capture how you work and turn it into a skill automatically.

Layer two, scale yourself. A people brief that assembles what you agreed at the last one-to-one, follow-ups you owe personally, their KPIs and performance against them, blockers, wins you should call out by name, and two questions worth asking that you would not have thought of. He is candid about why the wins line exists: “I’m really bad at giving people some kudos. I’m always looking for the problems.” Then a voice of you skill so colleagues can get an answer before they interrupt you, which answers from your intelligence layer rather than general knowledge, cites which files the answer came from, and states how confident it is that this is genuinely your view. That last detail is what makes it safe to use. Then a storyteller skill that turns your goals into the narrative you give execs about what marketing is contributing.

Layer three, better thinking. An exec panel built from your execs file, which you pitch to before pitching the real thing, so the holes get found in private. And a challenge me skill running structured thinking methods, including inversion, meaning “what would guarantee this fails”, and a pre-mortem framed as “it is 6 months from now and this has failed, write the post-mortem”.

Layer four, product story, which he argues is task number one for marketers right now because everything sounds the same. A voice of the customer built from your ICP file that you can interview: what are your biggest problems, what would make you happy, what about our pricing frustrates you. A voice of the competitor querying a competitors.md you update monthly with their launches, case study pages, product pages and paid messaging, so you can ask it how it is competing against you. Then an are we really different skill that takes positioning, competitors and ICP together and grades your differentiation from 1 to 10 across your website, product pages and advertising. The question it exists to answer: “Are we going to get lost in the noise?”

Layer five, run your day. A priority list built from goals, team activity, exec requests, email and Slack. A what-did-I-miss pass that sweeps email, Slack and document comments, groups them oldest first, and gives you a one-line draft reply for each. A blocker watcher that scans for people waiting on another team, a missing decision, missing budget, missing headcount, missing data, or repeated slipped dates, then tells you who is blocked, for how long, and with enough context to actually unblock them. And live artifact dashboards aggregating data across your tools.

Do this now: Build layer one only, this week. Six markdown files, written honestly, is a weekend job and it is the layer everything else depends on. Then add exactly one skill from layer five, because a daily-use skill is what makes you maintain the context files. His own warning applies to the rest: “You maybe only really need one of these. This is like pretty overbearing.” Do not build all fifteen skills. Build the intelligence layer properly and one thing that reads from it.

Watch more:

2. Claude Code as a one-person marketing team, with the real costs attached (Nate Herk)

Published 21 August 2026, 38 minutes, and the most complete end-to-end build of the week. He uses a fictional brand, a canned protein coffee called PerkForm, and generates every asset live. Disclosure: Higgsfield is the paid tool at the centre of this and the video is structured around it, so read it as commercially interested even though no sponsorship is declared on air.

It starts with a framework, not a tool. The three Ps: Pain, Person, Promise. What is the pain, who specifically feels it, and how does your product promise to solve it for them. His reason for insisting on this first: “You need all this stuff to make sure that your marketing assets don’t look and feel like generic AI slop. They have to have a purpose. They have to have intent behind it.”

The setup, step by step. Create an empty folder on your desktop. Open Claude Code desktop, sign in, start a new chat, click open folder, select it, and hit trust workspace. Despite the name, there is no coding involved. His framing: “Just because it has the word code, don’t think that we’re going to be coding. The reason why we’re using Claude Code is because Claude Code works out of directories, out of projects.”

Then you dump your business context in and ask Claude to build the project structure. What came back: a CLAUDE.md as the operating document telling every future session what to load, a /context folder holding brand positioning, customer avatar, product facts, voice and copy, a message bank, playbooks and competitive landscape, plus an empty /assets folder and an empty /output folder. The CLAUDE.md itself is short, essentially “this is the project, Nate owns this business, read these files before writing anything”.

A tip worth stealing if you struggle to articulate your own business. He uses a skill he calls grill me, which is just a markdown file whose instruction is to relentlessly interview you about a topic until you reach shared understanding. You drag it into the chat, ask Claude to install it, then run it on your product and marketing. It interviews you question by question and saves the result as context.

Connecting the image and video generation. Higgsfield aggregates multiple models under one subscription, including Seedance 2.5 for video, GPT Image 2 for images, plus Kling, Nano Banana and others. To connect it: in Higgsfield, go to MCP and CLI at the top, copy the Claude URL, then in Claude go to Customize, Connectors, Add custom connector, paste the URL, name it Higgsfield, add, connect and sign in. Higgsfield’s own documentation confirms these steps and confirms it works with Claude on web, in Cowork and in Claude Code, with no API keys to manage.

The debugging habit is the most transferable thing in the video. When he dragged a folder of product shots into the project, the sidebar displayed it oddly and he could not tell whether Claude could actually read the files. Rather than guessing, he asked: can you see all the product shots inside that folder, there should be eleven. Claude read them and returned the inventory. His advice, and it applies well beyond this workflow: “when you’re in a situation like this where you’re a little confused, just ask Claude what it sees.”

What it produced, and what it cost. He fired four prompts in the same session, each specifying the goal, the brand context to use, where to save the output, and a request to report the cost.

Eighteen ad creatives for a buy-one-get-one offer, on brand, using the colour scheme and typography from the guidelines PDF, some deliberately left blank for copy to be added later. Cost: $3.43. His own note on this: he gave it almost nothing to work with. “I literally just said, hey, create me some ads for BOGO. If you’ve been running ads for a while, then apply that subject matter expertise to your prompts and it will be even better.”

A sizzle reel in vertical and landscape, built through Higgsfield Marketing Studio with music, scene changes and edits. Cost: $17.55.

Five Instagram carousels of seven to eight slides each, consistent across the set. Cost: $9.56.

A set of UGC-style video ads, the most expensive of the four, and structurally the most interesting. Claude generated four consistent creator characters as images first, then storyboards for each, then video from the storyboards, then ran a quality assurance pass on its own output, screenshotting frames, rejecting generations with faults it spotted, and only then reporting finished.

He is honest about where it falls short, which is worth quoting because most videos of this type are not. On the UGC output: “they are still very like ad like, they’re very salesy in a way. You can tell it’s not completely organic.” His suggested fix is to have it analyse real UGC first, or write and approve scripts yourself before generation. And on the whole exercise: “you can’t just expect to throw one prompt out there and then go viral and go make a million dollars.”

The last step turns this from a generator into a system. He asked Claude to build a spreadsheet tracking every generation: date, generation type, campaign, channel, primary angle, angles covered, deliverables, number of generations, tool and model used, credits consumed, actual cost, and status. Plus tabs for creative assets, angle performance, a copy bank and a cost summary. It took about eight minutes and it used the brand guidelines to style itself. Once that exists, performance data can flow back in, and you can ask it which angles worked before commissioning the next round.

Do this now: Spend an hour on the folder and the context files before you generate a single asset, because that is the whole difference between branded output and generic output. Then run one small job with a cost report attached to the prompt, so you learn what your own volume actually costs before you commit. And build the tracker early rather than late, because eighteen creatives across four campaigns becomes unmanageable in a folder within a month.

Watch more:

3. Wiring Claude into Search Console and Analytics in under ten minutes (Eliot Prince)

Published 15 August 2026. Seven minutes long and it solves a problem a lot of people have hit: Claude has no native Google Search Console or Google Analytics connector, and browsing the connector directory will not turn one up.

The middleware is Composio. Note the spelling, because automated transcripts of this video mangle it badly. It is Composio, at composio.dev, and it connects Claude to a large catalogue of applications through the Model Context Protocol.

Step one, set up Composio. Create an account and log in. Before doing anything else, check that “For you” is selected in the top left, not “Platform”. He is emphatic about this, and the reason is that the platform view is the developer-facing side and will send you down a much more complicated path than you need.

Step two, connect the two Google properties. Go to Connect Apps, search for Google Search Console, click connect, and complete the authorisation. Repeat for Google Analytics. You can connect multiple accounts, or a single account holding all your properties, which is the useful configuration if you manage several client sites.

A gotcha the video skips, which is worth adding. Composio’s own documentation and other walkthroughs flag that during the Google authorisation step you must tick every permission box, particularly the ones granting the ability to view and manage analytics data. Miss one and the agent will see the account listed but will be unable to pull any reports, which produces a confusing failure that looks like a broken connection rather than a permissions problem.

Step three, connect Composio to Claude. Copy the Composio MCP server URL. In Claude, go to Customize, then Connectors, then Add, then Add custom connector. Paste the URL into the second field, give it a name, and ignore the advanced settings entirely. Click add, then click connect, and complete the authorisation back to your Composio account.

Step four, and this is the bit people get stuck on. You have to name the tool explicitly in your prompt. In his words: “You have to tell it which app you want it to actually pull from through Composio.” So the request looks like: get me the overview stats from Google Search Console for the last 30 days, through Composio, for this domain. His own fix for the repetition is to save a memory rule so that any Search Console request routes through Composio automatically.

What it returns. He asked only for thirty days of overview data and Claude volunteered the analysis unprompted, flagging that traffic was dangerously concentrated, that desktop was the weak spot, and that click-through rate had collapsed. He then asked it to build a live artifact dashboard from the same data with action points, styled to his brand. He keeps a version pinned in the sidebar covering several sites, refreshing on a schedule.

Do this now: Do the connection itself, which genuinely takes about ten minutes, then resist the urge to build a dashboard immediately. Ask it one real question first, something you would otherwise export a CSV to answer, such as which pages lost the most clicks month on month and what they have in common. The value here is not the dashboard, it is that your SEO data becomes something you can interrogate in the same place you write.

Watch more:

1. Meta AI can now analyse and optimise your Meta Ads campaigns

Announced 20 August 2026, and positioned as a small business feature. You connect your Meta ad campaigns to Meta AI and it analyses performance, recommends changes and produces decks, documents and spreadsheets from the results. The part that deserves a pause before anyone clicks through the setup is that it also connects to Google Workspace, meaning Gmail, Docs, Sheets and Slides. Connecting your company email and document store to Meta’s AI is a procurement and data governance decision, not a toggle a performance marketer should flip on a Tuesday afternoon. Worth evaluating on the ads side, worth routing through whoever owns your data policy on the Workspace side.

2. Google Demand Gen view-through optimisation goes video-only, and switches itself on

From 17 August, view-through conversion optimisation in Demand Gen became video-only, with image asset view-throughs demoted to secondary conversions. Two details make this more consequential than the headline suggests. It is enabled by default on all new Demand Gen campaigns, so you have to opt out rather than in. And it expands beyond YouTube and Discover onto the Display Network, where Display video assets move from CPC to CPM billing regardless of whether you use view-through optimisation at all. A billing model change combined with a default-on setting is exactly how a quarter’s numbers move without anyone noticing until the reporting lands.

3. Gmail’s domain reputation score is now actually gone

Confirmed on 21 August, and this is an ending rather than an announcement. Postmaster Tools v1 has finally shut down, the old interface link no longer works, and the Domain and IP Reputation charts have disappeared. The deprecation was announced back in 2024 and v1 was officially retired in September 2025, so the correct framing is that the thing you were warned about has now happened rather than that Google just did something. What replaces it is a Compliance Status dashboard mapped requirement by requirement to Gmail’s bulk sender rules, plus a Deliverability Analysis added in June 2026 that returns a plain-language verdict, including two outcomes that are not failures: not enough volume to determine, and recipients appear to want more of your messages. The buried detail worth the whole entry: Google’s published sender guidelines still cite 0.3% as the spam complaint ceiling, but the v2 API’s failure code is defined at 0.1%. Both thresholds are live simultaneously, checked by different systems, so a sender can be compliant with the published guideline and still receive a negative verdict. Treat 0.1% as your internal alarm. Note that the 0.1% figure is observed behaviour reported by deliverability vendors rather than published Google policy.

4. Google changed how Branded Searches are measured, and dropped Performance Max

A documentation update on 18 August rather than a new feature, but it changes numbers you may already be reporting. The default conversion window is now seven days, down from a thirty-day view-through window, adjustable between one and thirty days. More importantly, the eligible campaign types are now listed as YouTube and Demand Gen only, and Performance Max no longer appears. It sits under the Consideration goal, remains reporting-only rather than a bidding target, and shows up in Results and All Conversions rather than Conversions. If Branded Searches out of Performance Max appears anywhere in your reporting, check this week whether it still exists.

5. Informa TechTarget launches a free B2B AI Authority Index

Released 17 August, a 24-question self-assessment scoring how likely your brand is to be recommended by AI systems during a B2B buying process. It places you in one of four archetypes: The Newcomer, The Contributor, The Referenced, and The Authority. It comes with a companion white paper. Set expectations correctly: this is a diagnostic lead magnet rather than research, and Informa TechTarget sells B2B media and intent data into precisely the gap it identifies. That said, twenty-four structured questions is a reasonable free way to force a conversation about AI visibility with people who have not had it yet, particularly if you need something to put in front of a sceptical stakeholder.

6. Google AI Max gets cross-campaign testing, and keeps your guardrails on

Announced 20 August, with the headline feature arriving in September. From then, a single A/B test can span multiple Search campaigns at once with different budgets and return-on-investment targets, which finally matches the level most B2B accounts actually budget at rather than forcing single-campaign tests. The quieter improvement is available sooner and matters more: AI Max experiments can now run with brand and location controls left enabled, so you no longer have to strip out your guardrails in order to measure the thing you are testing. Performance Planner also gains the ability to forecast bidding and budget changes and apply the recommendation in one click, though see the top story before trusting its numbers this fortnight.

7. Perplexity Computer now runs from your inbox

Shipped 18 August. You can send, forward or copy in computer@perplexity.com to trigger a full Computer agent task without leaving your email client, and the finished work comes back in the same thread. It is the lowest-friction delegation pattern any vendor has shipped so far, because it requires no new tool for anyone to adopt and no change to how a team already works. Two caveats. It needs a plan that includes Computer, and no independent testing has appeared yet. More importantly, emailing an agent means emailing your content to a third party, so the same governance question applies here as with the Meta AI entry above.

Read more: https://www.perplexity.ai/hub/blog

1. Only 11% of brands appear across all three major AI models

Published 18 August 2026 by Fractl, and it is the rare AI visibility study that publishes its full methodology. 96 industry-specific prompts, each run 15 times across three models, producing 4,320 responses and more than 8,500 unique brand references, then mapped against Ahrefs domain authority, organic traffic and keyword rankings.

The headline finding is that traditional search authority and AI visibility align for over 90% of brands, which is reassuring and slightly dull. The number that should change what you do is the split beneath it: only 11% of brands appeared across all three models, while 77% appeared in just one. Roughly 5% were underexposed despite strong SEO, and about 4% overperformed in AI answers on modest traditional signals.

What to do: If you are measuring your AI visibility on a single platform, and most tools default to one, your reported position is systematically overstated. Three quarters of brands are visible in exactly one place, which means a single-engine check will tell most companies they are doing fine. Add a second and third engine to whatever you are tracking before you report a number to anyone. Disclose in passing that Fractl sells digital PR and content distribution, which is the remedy the study implies.

2. Meta is now the biggest AI crawler on the web, and it sends almost nothing back

Published 18 August 2026 by DataDome, with fieldwork covering Q2 2026, so this is genuinely current rather than a recirculated older study. 17.7 billion AI agent requests, up 45% on Q1.

The growth did not come from where you would assume. Neither Google nor OpenAI drove it. Meta-ExternalAgent grew 74% quarter on quarter and Meta-WebIndexer grew 163%, and Meta’s two agents now account for the majority of AI agent traffic DataDome observes. They also return close to nothing in referral traffic.

What to do: Go and read your own robots.txt this week. Almost every AI crawler rule written in the last two years targets GPTBot and Google-Extended, because those were the names in the discourse when the rules were drafted. The largest crawler by volume is now one most policies do not mention, and it is taking content without sending readers back. Whether you block it is a strategic call rather than an obvious one, but making that call by accident because your file is eighteen months old is not. DataDome sells bot defence, so weigh the framing accordingly.

3. AI visibility reached the boardroom faster than the data to support it

Published 21 August 2026. A survey of 400 people responsible for marketing and communications at companies in the US, UK, France and Germany, spanning C-level through to manager.

The adoption numbers are striking. 54% now measure AI search visibility or brand citations in AI-generated content. 46% include AI visibility in C-suite reporting. 42% track which prompts surface their brand’s coverage. For a discipline that barely existed eighteen months ago, that is remarkably fast institutional adoption.

The tension is in what sits underneath. Only 49% are very confident in the accuracy of their data, while measurement influences strategy and budget for 88% of paid social decisions, 87% of paid media and 85% of owned content. Just 35% have fully integrated reporting and 37% are still working from manual spreadsheets.

What to do: If you are one of the 46% now reporting AI visibility upward, put a confidence caveat next to the number before someone else finds the hole in it. Reporting a metric to the board creates an expectation that it is solid, and half of the people doing it do not believe their own data. Pair this with the Fractl finding above, because single-engine measurement is one specific and fixable reason to distrust your own figures. Two caveats on the study itself: the fieldwork date is not published anywhere, which is a real weakness, and it comes from 10Fold, a B2B technology PR and communications agency that sells into exactly the gap it describes.

1. Microsoft gutted the free Copilot tier on 18 August with five days’ notice.

Deep Research is gone for consumers, and its successor Researcher requires Microsoft 365 Premium at $19.99 a month, which Personal at $9.99 and Family at $12.99 do not include. Copilot Podcasts were retired entirely along with the back catalogue, Group Chats converted to one-to-one with other participants’ contributions becoming unreachable, and Copilot Labs wound down.

2. ChatGPT Ads switched automatic advanced matching on for existing web pixels on 17 August, and the opt-out window has closed.

It applies to any account running a ChatGPT web pixel, so a test campaign from the spring that nobody has touched since turned itself on. You can still reverse it per pixel via Tools, Conversions, Data Source, Edit pixel.

3. From 24 August, YouTube counts a view from the first frame of playback across Shorts, long-form and live.

Public view counters will climb faster as a result. Partner Program earnings and eligibility are unaffected.

4. Google Ads API v25.1 landed on 19 August

with 24 new Conversion Lift metrics, brand lift dimensions on an allowlist, category-scoped benchmarks with share of voice, and brand sentiment for creators. The quietly useful addition is original_conversion_value, which shows conversion value before Google’s own value rules adjust it.

5. Google is testing Enhanced Matching for Customer Match,

Spotted on 19 August as an unchecked opt-in setting that would expand list reach by matching your consented first-party records against consented users from publishers. This was found in a screenshot rather than announced, Google has not confirmed it or named any participating publishers, so treat it as a test in the wild.

First, provenance moved this month from something a reader can see to something only a machine can read, and nobody put it that way out loud.

Anthropic began marking everything Claude writes with an invisible watermark, worldwide and with no way to switch it off. Google made its visible watermark optional while leaving SynthID and C2PA embedded regardless. The IAB ruled that AI-written copy requires no disclosure at all, while AI-generated imagery does even after a human reworks it.

Three separate organisations, three different mechanisms, one direction. The signal that content is synthetic still exists in every case, and in every case it has become less visible to the person actually reading it. What that transfers is responsibility. When the label was automatic, the platform carried the disclosure. Now the file carries a mark that only its maker can detect, no third party can verify, and the publisher is left holding the legal obligation. Anthropic’s own detection interface still does not exist, and the EU deadline that forces it out is 2 December.

For a marketer the practical question is no longer whether to disclose AI use, because the regulation already answers that. It is whether your team knows which outputs carry an obligation, and right now the honest answer in most organisations is no.

Second, an entire measurement category appeared in a single fortnight, and every company selling it also sells the cure.

Five vendors shipped AI visibility measurement products on 20 August alone, with Cloudflare and Microsoft doing the same the week before. Meanwhile Fractl found that 77% of brands appear in only one of the three major models, DataDome found that Meta quietly became the largest AI crawler on the web while returning almost no traffic, and 46% of marketing leaders are already reporting AI visibility numbers upward while only half of them trust the data underneath. That combination is worth naming plainly.

Demand for the metric has outrun the ability to measure it, the tools filling the gap were built by companies with a commercial interest in the answer being alarming, and the numbers are now reaching boardrooms. This is roughly where social media measurement sat in 2011, and the lesson from that period is not that the metrics were worthless. It is that the people who did well were the ones who wrote down what they were measuring and why, before a vendor wrote it down for them.

REMINDER: MY BEST BUNDLE YET FOR MARKETERS:

Check Out This Limited Time Offer

Tell me: What’s your biggest takeaway from this weeks news from the world of AI Marketing? Love to hear your thougths in the comments.

Leave a comment

Thanks for reading the Marketing with AI Weekly Roundup! If you enjoyed this read, the best compliment I could receive would be if you shared it.

Marketing with AI Weekly Roundup: Claude Starts Watermarking Everything It Writes, LinkedIn Publishes Its Own AI Search Playbook, and Perplexity Punishes Site Caught Feeding Bots Different Content

Robert Gillespie — Sun, 16 Aug 2026 10:02:14 GMT

Welcome to issue #47 of the Marketing with AI Weekly Roundup. Published every Sunday, it shares the top stories, playbooks, data and quick hits from the world of Marketing with AI over the last 7 days.

Subscribe to receive the week's most important AI marketing stories direct to your inbox.

Subscribe now

NEW: Over the last 6 months, I've asked 2000+ marketers I’m about to train the same question: where could AI save you time? I’ve now turned the data from those answers into a tool that tells you where you too can save time with AI based on your working week.

Get a personalised plan now for your marketing week, built from your answers. 👇

Learn more about the Operator Gameplan

Last week was about what your numbers will say and what your job will look like.

This week is about proof: who can tell where your content came from, and whether the machines reading your website are being shown the same thing your buyers are.

On Friday, Anthropic confirmed it is marking everything Claude writes with an invisible watermark, worldwide, at the model level, on every surface it ships.

On Tuesday, LinkedIn published its first proper guide to getting cited in AI search, and the headline finding is that three quarters of the LinkedIn citations appearing in AI answers come from individual people rather than company pages.

And on the same day, Perplexity blocked Time magazine’s machine-only advertising and called it deceptive, less than a fortnight after we flagged the format here.

Elsewhere, Google quietly gave you control of your own attribution window for the first time, three separate datasets landed on the same uncomfortable conclusion about AI referral traffic, and the creator who runs the most-watched practical AI channel of the week deleted all his skills to see what would happen.

Below: four stories to act on, three workflows worth stealing, and the clearest evidence yet that the number in your analytics is not the size of the effect.

Let’s get into it…

1. Anthropic is now watermarking everything Claude writes, and it follows the text into your CMS

On 14 August 2026 Anthropic published the detail of how it marks AI-generated content. Every Claude model launched on or after 2 August 2026 carries an invisible, machine-readable watermark in its text output from day one. Models released before that date are being retrofitted over the coming months.

This is not a product feature, it is a property of the model. The mark applies across Claude.ai, the Claude Platform API, Claude Code, Claude Cowork and Claude Tag, and wherever Claude is offered through cloud partners including AWS, Google Cloud and Microsoft Foundry. Anthropic’s own framing is that the mark “will be present no matter which Claude product or surface the text comes from.” There is no opt-out, no extra tokens and no extra cost.

How it actually works, because the mechanism matters for what it can and cannot do. It uses Google DeepMind’s SynthID-Text method, published in Nature in 2024. It does not insert invisible characters or hidden metadata. Instead, when the model is choosing between several equally valid next words, it uses a cryptographic key plus the preceding words to settle which one to pick. The text reads identically. Anthropic says there is no practical impact on quality or content.

What survives and what does not. The mark travels with the text when you copy and paste it into a document, a CMS or an email. Anthropic says it “may persist through some editing” but has not defined how much editing that covers. It is not reliably detectable when content is heavily edited, paraphrased, translated, mixed with human writing, or simply too short. Images get C2PA content credentials instead, which is signed metadata rather than an embedded watermark. Code gets minimal marking, mostly in comments where there is arbitrary choice.

The part everyone will get wrong. Detection requires a cryptographic key that only Anthropic holds. A detection API is described as forthcoming with no date attached. Anthropic’s own limiting statement is the sentence to hold onto: detecting a mark “tells you that the content may have been processed by Claude. It does not, on its own, confirm the full provenance of the content.”

The trigger was the EU AI Act’s Article 50 transparency obligations, which took effect on 2 August. Anthropic chose to roll the response out globally rather than scope it to Europe.

Why it matters / what to do: a watermark in your draft is not a penalty, not a search signal and not a detector pointed at you. No search engine has said it reads this mark, and Anthropic has not said it hands the mark to third parties for ranking or moderation. It is a provenance signal that exists at the point of generation, inside a tool you control. So do not spend the week hunting for a way to strip it.

Do three things instead. First, if you have a client-facing or public AI disclosure policy, update it now, because a technical provenance signal exists where one did not before, and the honest version of your policy is easier to write today than to retrofit after a client asks. Second, make the substantive editorial pass a required step rather than optional polish, meaning restructured sections, added sourcing and a real fact-check, not a synonym swap. Detectability drops as a side effect of that work, but the reason to do it is that original data, direct quotes and a specific point of view are what separate a page worth reading from a page that summarises what already exists. Third, ignore any tool claiming to detect Claude’s watermark until Anthropic publishes its own documentation, because without the key nobody outside Anthropic can do it.

2. LinkedIn published its own AI search playbook, and it says your people are worth more than your company page

On 11 August 2026 LinkedIn published its first substantial guide to B2B visibility in AI search, written by Davang Shah, its VP of Marketing. This is first-party guidance from the platform, not another consultant’s teardown, which makes it the most quotable thing published on the subject this year.

The number that should reorganise your plan: 75% of LinkedIn citations appearing in AI answers come from individual member profiles, not company pages. That figure is Meltwater’s 2026 research rather than LinkedIn’s own measurement, and LinkedIn is repeating it in its own guide. Alongside it: articles generate roughly 60% of LinkedIn content citations, with posts accounting for the other 40%.

The supporting numbers are worth carrying because they set the stakes. 94% of B2B buyers now use generative AI for research (6sense). Gartner still forecasts a 50% decline in traditional search volume by 2028. LinkedIn’s own non-branded search traffic is down 60%. Profound puts the median time for a new page to earn its first citation at 6.81 days, with 90% of pages taking up to 37 days. And eMarketer found only 34% of B2B tech marketers feel prepared to execute an AI strategy.

LinkedIn calls the model a credibility stack: executive thought leadership, employee advocacy, third-party media coverage and contributed articles, partnerships with niche creators, and consistent reinforcement of the same core themes across all of it.

The useful part is that it comes with actual specifications, which is rare in guidance like this:

● Pick 3 to 4 core topics to own, and build clusters of content around each one rather than posting across everything.

● Posts: 200 to 300 words, keywords in the opening line, structured as a question and answer.

● Articles: 800 to 1,200 words, clear headers, a TL;DR summary at the top, and a question-driven title.

● Post 2 to 3 times a week on those core topics, publish long-form articles on the priorities, and launch a newsletter.

● Activate the people, not just the page: executives, product experts and employees posting regularly.

It also splits measurement into three tiers, which is a more honest structure than most vendor dashboards offer. Leading indicators move quickly and tell you little on their own: impressions, reactions, comments, shares. Outcome metrics move slowly and are what you actually report: citation counts, share of voice, citation rank, sentiment. Diagnostic patterns tell you what to change: which formats get cited, which authors get cited, which prompt types surface you.

Why it matters / what to do: most B2B teams have spent two years optimising a company page that, on LinkedIn’s own numbers, generates a quarter of the citations their people do. Action for this week: pick your three named experts, give each of them one of your core topics, and get one 800 to 1,200 word article published under a personal profile rather than the brand account. If you only do one thing, make it that, because the article format is doing 60% of the citation work and almost nobody in B2B is publishing them.

One caveat worth stating plainly. LinkedIn has an obvious commercial interest in concluding that the answer to AI search is posting more on LinkedIn. The specifications are still useful, and the 75% figure comes from Meltwater rather than from LinkedIn’s own data, but read the guide knowing who published it.

3. Google Analytics will finally let you set your own attribution window, and B2B has been quietly punished by the old one

On 14 August 2026 Google Analytics replaced its fixed preset lookback windows with custom ones. Click-through conversions can now be set to any value from 1 to 90 days. Engaged-view conversions can be set from 1 to 30 days, having previously been locked at three. The settings live under Advertising, then Conversion management, then Settings, and in linked Google Ads conversion management.

This is the least dramatic story in this week’s issue and probably the one that changes your reporting most.

Why it matters / what to do: if your sales cycle runs 90 days and your attribution window has been sitting at 30, you have been systematically under-crediting every campaign that starts a deal and over-crediting every campaign that happens to be running when it closes. That is not a rounding error, it is a structural bias in favour of bottom-of-funnel activity, and it has been quietly shaping budget decisions in B2B accounts for years because the setting was not adjustable.

The action takes about twenty minutes and needs your CRM rather than your analytics. Find your actual median time from first touch to closed-won, not your average, because a handful of very long deals will drag an average somewhere useless. Then set the click-through window to match that median. Then, and this is the part people skip, write down what your reporting looked like before you changed it, because the numbers will move and in six weeks you will want to know whether the campaign genuinely improved or whether you simply started counting differently.

One caveat: Google did not state whether the change applies retroactively to historical data, and we could not confirm it either way. Assume it does not until you have checked your own account, and do not present a before-and-after to your leadership until you know.

4. Perplexity blocked Time’s machine-only ads and put a number on the punishment

On 11 August 2026 Perplexity blocked all markdown advertising from Time.com from influencing its agents and its user-facing answers. That is less than two weeks after Digiday broke the original story, which we carried in On Our Radar on 8 August.

The recap, briefly, because the mechanic is the point. Time had partnered with adtech vendor Mobian to serve a stripped, machine-readable version of its pages to AI crawlers, containing sponsored content written in the advertiser’s own marketing language that no human reader ever sees. Request the page as a browser and you get roughly 303,000 bytes of normal HTML. Request the identical URL as ClaudeBot or PerplexityBot and you get about 13,000 bytes of markdown with the sponsored material baked in.

Perplexity’s Chief Communications Officer Jesse Dwyer said the company works continuously to protect users from deceptive practices, and that publishers deploying “deceptive advertising like markdown ads” face a reputational downgrade including a hit to their trust score in Perplexity’s search index. Time declined to comment. Mobian co-founder Jonah Goodhart called the overall response “extremely positive” and questioned why Perplexity would block factual brand information, arguing that accurate real-time facts serve users better. No other AI company has publicly responded.

Why it matters / what to do: the line on serving different content to machines than to humans has now been drawn, and it was drawn by a platform rather than by a standards body or a regulator. It came with a named consequence attached to a scoring system nobody outside Perplexity can see. That is a materially different risk profile from the one that existed a fortnight ago, when this looked like an interesting new revenue line.

The action is the same one we gave on 8 August and it has got more urgent rather than less. Fetch one of your own key pages six times, changing only the user agent: a normal browser, then Googlebot, ClaudeBot, OAI-SearchBot, PerplexityBot and GPTBot. Compare the six responses. Most marketers assume this is not their problem because they are not selling agent ads, and that is exactly the assumption worth testing, because your CDN, your CMS or a well-meaning performance plugin may already be serving assistants something different from what it serves buyers. You want to discover that yourself rather than when an answer engine summarises the wrong version of your pricing page to a prospect.

1. The three levels of AI memory, and why yours is stuck on the first one (Jeff Su)

Published 12 August, and it explains the single most common complaint marketers have about AI, which is that it forgets the brand voice you taught it last week. Disclosure: the video is sponsored by Granola and he sells a Cowork course, so treat him as an operator with a commercial interest. The framework survives that.

Level one, global memory. Account-level facts your chatbot generates automatically. In ChatGPT it lives under Personalization, then Memory, then Manage. His core insight is that this layer is deliberately kept thin, because anything saved at account level gets dragged into every future conversation, and in his words, “if something wrong lands there, it poisons everything.” So it knows your role and your writing preferences and nothing about the campaign you worked on yesterday.

You can force entries in by saying “update your memory, I have a high-stakes presentation on October 6th” mid-chat and watching the memory page update. But you have to remember to do it every time, and it still lands in the global profile where it will follow you into unrelated work.

Level two, project memory. Projects draw a boundary around one work stream, so entries can be specific without contaminating everything else. Global memory still cascades in.

His failure example is the useful part. He pasted a screenshot of a calendar invite into a project chat, and Claude read it correctly at the time. Asked a week later for the confirmed attendee list, it returned a partial one, because it had decided the names were not worth keeping. His verdict: “the AI is still the author. It decides what makes the cut.” The fix is to say it explicitly: “update project memory, John Turnip is now also confirmed to attend.”

Level three, a memory system that lives in your own files. Supported by Claude Cowork and Claude Code, ChatGPT Work and Codex, and Gemini Spark.

The mechanic is simpler than it sounds. At the start of every task the system loads one small file, a `root memory.md`, which is nothing more than a routing table of active projects. You say “I want to continue working on the X presentation”, it checks the table, routes to the folder that owns that project, reads those files and loads nothing else. He says this holds up with more than twenty active projects.

The session wrap-up ritual is the part worth stealing outright. At the end of a working session you type “I’m done with the session, let’s wrap up.” It scans the session for decisions, learnings and progress, then splits them into two lists. An input needed list that requires your approval, which in his demo meant turning a one-off correction (”don’t use acronyms in slide headlines”) into a permanent rule. And an approval exempt list of things it has already done, such as recording the edits and updating the routing table so tomorrow’s session starts where today’s ended.

The line that frames the whole thing: “everything you’ve seen so far has the same root issue. The AI is still in control. It decides what gets remembered, where it lives, and when it gets written.”

His own honest caveat is that level three “does take some real effort to set up, and they work differently than the chat window you might be used to.”

Timely addition he could not have included. OpenAI shipped a related change on 14 August: an existing unshared project can now be switched between default memory and project-only memory without starting over, via the three-dot menu, then Project settings, then Memory. If you have a project that has been quietly absorbing your global memory, that is a two-click fix.

Do this now: create a single `root memory.md` listing your three live projects and the folder each one owns. Then end your next working session with “I’m done with the session, let’s wrap up” and see what it proposes to remember. The test of whether it worked is opening a fresh session on Monday and asking “where are we on the Q4 campaign and what is next”, then seeing whether the answer is correct.

Watch more:

2. Delete your skills, then rebuild only the ones that were doing formatting (Nate Herk)

Published 12 August and the highest-viewed practical video of the week. Worth being straight with you: we flagged the underlying idea on 25 July in On Our Radar, when Anthropic’s Thariq Shihipar wrote about carefully written instructions working against you. What is new here is a working test attached to it, plus an instruction on camera that the blog post never gave.

The instruction comes from Boris Cherny, the creator of Claude Code, interviewed at Y Combinator’s Startup School. Quoted directly: “for people that aren’t building agentic products, but you’re using Claude Code, every 6 months delete your Claude MD. Delete your skills. Delete your hooks. See what the model does and it might surprise you. And actually for Opus 5, this is something we really do recommend.”

On what replaced prompt engineering: “the skill nowadays is less about prompt engineering and more about figuring out how do you give Claude a hard task that seems a little bit too hard. And then how do you make it possible for Claude to verify its work along the way. And the verification, I think, is probably the single most important thing that people do not get right.”

The test is what makes this actionable. Herk duplicated his setup, stripped out his instruction file and every skill, and gave both versions the same job: turn a YouTube interview into a resource guide.

● With skills: nine pages, properly formatted, brand header image, channel links, correct colours. Structurally exactly what he had specified.

● Without skills: messier, no header, no branding. But it broke the content into ideas with timestamps entirely on its own, which he preferred. His words: “arguably I like the content of this version better.”

The conclusion to build on: keep the instructions that specify look, brand assets and constraints. Delete or loosen the ones that dictate the thinking. He would rewrite his skill as “make the resource guide however you see fit, but put this image in the header and link to my channel at the top.”

His own caveat is good editorial hygiene and we are running it: “you should probably be taking advice from people who are using the AI systems the same way you want to.” Cherny designs agent harnesses and works across enormous code bases. A marketer producing research, documents and deliverables should not blanket-apply his advice.

Do this now: take your single most-used skill or custom instruction set. Copy it. From the copy, delete every line that tells the model how to think, and keep every line that tells it what the output must look like. Run both on the same brief and compare. If the stripped version thinks better, you have been paying for formatting with your quality.

Watch more:

3. Give Claude a shared whiteboard it can actually draw on (Blazing Zebra)

Published 9 August. The problem it opens with is one every marketer recognises: ask an advanced model a simple planning question and you get back “14-page essays for even doing the simplest thing.” The fix is an editable, infinite whiteboard that both you and Claude can read from and write to. Disclosure: his prompt cheat sheet sits behind a Patreon, though the setup is described fully in the video.

Setup. Open the Claude desktop app, switch to Cowork, point it at a new empty folder. The setup prompt asks Claude to get TLDraw running locally with a file-backed bridge so both parties can read and write to JSON files, and specifies that it “should be only about 40 lines of code.” It runs on localhost. Test it by writing something on the board and asking Claude to read it. In his demo, Claude replied on the board itself.

Then drop your messy source material into the same folder, so PDFs, emails, call transcripts and documents, and tell Claude to get familiar with it.

The sequence of requests, each of which maps onto a real marketing job:

1. A mind map of the project and its major work streams. His steering phrase is worth borrowing: “keep things as simple as possible, but no simpler.”

2. A dependency map from today to the launch date, showing the critical path and anything blocking it. Blockers render in red.

3. Change one date and ask it to update the critical path. Everything downstream turns red, which is the moment the diagram earns its keep.

4. A Kanban board: “take that task CSV and build it out as a Kanban with the owner’s initials on each card.”

5. Move the cards yourself, then say “I made some changes directly to the board, tell me what changed and replan anything that needs replanning.”

6. A confidence heat map colour-coding how sure Claude is about each item, which you can then overrule.

7. Per-stakeholder versions: share, download, export to PDF, strip detail for different audiences.

Why a whiteboard rather than Claude’s usual output, in his words: “Unlike HTML, sometimes when these things build HTML, you can’t get in there and manipulate that. You have to use the AI to manipulate it.”

Do this now: if you have a campaign launching in the next quarter, build the dependency map first and then move one date. Watching the critical path recalculate in front of you is the fastest way to find out whether your launch plan was ever real.

Watch more:

1. Ahrefs launched Letaido, an agent workspace that keeps running after you close the tab

On 12 August Ahrefs launched Letaido, a separate always-on workspace where agents run multi-step jobs, build dashboards and reports, monitor your site and your competitors, and execute recurring workflows on a schedule, with native access to Ahrefs data. Integrations named at launch: Notion, Slack, HubSpot, Google Ads and WordPress. The differentiator against a chatbot is persistent hosting, meaning automations survive between sessions rather than dying with the conversation. Pricing was not published. The launch case study claims Foundation Marketing cut keyword research and bottom-of-funnel content audits from 40 hours a week to about 60 minutes, and that number is a customer quote inside a vendor press release, so attribute it as a vendor claim rather than a benchmark. Worth setting against a figure from the same coverage: 81% of martech leaders are piloting AI agents, and 45% say vendor-supplied agents have not delivered what was promised. This is the SEO tool category following Ahrefs Brand Radar and HubSpot AEO, which we covered on 8 August, into agent workspaces. The buying question is whether you want your agents living inside a vendor’s workspace or inside your own file system.

2. Claude Cowork moved into the Chrome side panel, and took your skills with it

On 12 August the Claude in Chrome side panel became a full Cowork session. The practical change is that your existing skills, connectors and plugins now work inside the browser with no additional setup, conversations save to Claude history, and a task started in a tab can be finished on desktop, web or mobile. It can navigate pages, click, type and complete forms using the logins you already have. Live on Max and Team now, rolling out to Pro over the coming weeks, and off by default on Enterprise where admins can restrict it to approved domains. Chrome only, not other Chromium browsers, not mobile. This matters most for the tools that have no integration and never will: ad platforms, your CMS, your email service provider, vendor portals and internal dashboards. Anthropic’s own caveat is worth repeating, that browser agents remain vulnerable to prompt injection and its safeguards “cannot eliminate” the risk, so this is not the place to let it act unsupervised on anything irreversible.

3. Google put AI agents on the front page of Ads and Analytics

Announced 10 August. Google Analytics now shows AI Overviews on the homepage summarising what changed since you last logged in, with optional email or phone notifications at a frequency you choose. Google Ads gets a redesigned homepage with personalised AI insight cards and a prompt box for questions such as how competitors are affecting your impression share, and you can click a card to route its context straight into Ask Advisor. A new Dashboards feature in Google Ads builds visualisations from a text prompt, with Analytics support described as coming soon. Ask Advisor will also now benchmark your campaigns against anonymised averages from similar businesses. Caveats matter here: this is beta, English-language accounts only, and Google’s own label on it is that generative AI is experimental. No timeline for general availability. The Monday-morning “what happened in the account last week” job is the one most B2B marketers do manually and least well, so the honest test is to ask it what changed before you open a single report, then check its answer against your own read.

4. Search Console’s generative AI report is now on effectively every account

We have covered this report twice, on 13 June when it launched and on 8 August when we warned its data was more misleading than it looks. What changed on 11 August is that the rollout finished. A pop-up now appears on the Search Console home screen pointing you at it, and it sits as an expandable tab beneath the main Performance report. You get impressions inside AI Overviews and AI Mode, separated from your normal search performance and broken down by page, country, device and date, with a matching view for Discover. You still get no clicks and no queries, and neither has appeared since June. The detail that makes this urgent rather than interesting: the data starts on 18 May 2026, nothing before that exists, and nothing backfills. Every week you do not record a baseline is a week of history you cannot recover. One honesty note: Google has not formally announced the full rollout, its help documentation still says a subset of website owners, and John Mueller has said it is not yet every domain and only appears once a site has enough AI-feature impressions.

1. ChatGPT referrals to B2B websites are up 303% in a year

Published 12 August from Labs by Demandbase platform data, drawn from $123 million in media spend across 75,645 campaigns. Monthly ChatGPT-referred visits to the B2B sites Demandbase tracks went from roughly 645,000 in June 2025 to 2.6 million in June 2026, with a clear inflection in May 2026 when volume more than doubled against prior months.

The competitive split is the part worth noting. ChatGPT accounts for the overwhelming majority of that growth. Perplexity referrals actually declined, while Gemini and Claude stayed flat. No conversion or engagement-quality figures were published, so this is a traffic story rather than a pipeline story.

What to do: if you are building an AI visibility programme and splitting effort evenly across four engines, this is the number that says stop. Weight your prompt testing and your measurement toward ChatGPT until your own data says otherwise. Demandbase’s own caveat is the honest one to carry: much of the activity happens before a buyer ever reaches your site, so referral counts undercount influence rather than measuring it.

2. Only 1.1% of publisher visits carry an AI referrer, and that is not the same as AI not working

Published 13 August, from a Scrunch study using a privacy-safe opt-in panel with fieldwork running February to June 2026, covering millions of searches, AI conversations and publisher visits.

AI referrals accounted for 1.1% of publisher visits. But readers were 20.5 percentage points more likely to visit a publisher in the week after a news-related AI conversation. Roughly 75% of post-chat visits arrived through direct navigation and about 9% through traditional search, meaning the visit happened but carried no AI referrer with it. Major publishers took 97% of follow-through visits while accounting for 82% of AI mentions, while mid-sized publishers had about 17% of mentions and a far smaller share of the visits.

What to do: read this next to the Demandbase number above, because the two datasets are measuring different halves of the same effect and both point the same way. Your AI referral figure in GA4 is not the size of the effect, it is the visible corner of it. The practical fix costs nothing: add a “how did you hear about us” field to your forms that names ChatGPT, Claude, Gemini and Perplexity explicitly, and compare what people tell you against what analytics recorded. Where those two numbers diverge is your actual blind spot. Caveats: this is correlation rather than causation, it measures visits rather than revenue, and Scrunch sells AI visibility software.

3. ChatGPT is now serving ads on a quarter of commercial prompts, and one in seven misses the topic

Published 10 August in Search Engine Land, from SE Ranking’s analysis of more than 50,000 commercial prompts across 20 niches, with data collected on 23 July 2026.

Sponsored placements appeared on 25.94% of commercial prompts, putting ChatGPT close behind Google AI Mode’s 29.45% from the same research series, which we covered on 25 July. 14.35% of the ads were semantically unrelated to the prompt they appeared alongside, roughly one in seven. And in 96.37% of placements, the advertiser was not cited anywhere in the answer copy above their own ad. Advertisers also currently have no visibility into which prompts triggered their ads.

What to do: the last figure is the strategically important one and it will be missed in most coverage. Being the advertiser is not the same as being the source. In 96% of cases, someone else’s content was informing the answer while your ad sat underneath it. If you are eligible to buy ChatGPT ads at all, which most B2B readers are not yet, the earned side of that placement is worth more than the paid side, and it is available to everyone regardless of category eligibility.

1. The IAB published the first industry standard for measuring AI visibility.

Released 3 August, “Measuring Visibility in the AI Era” is a 36-page framework built on four levels, Presence, Prominence, Portrayal and Persuasion, and it splits measurement into directional and decision-grade tiers across nine criteria, treating anything under 50 queries per programme as exploratory. It states plainly that directional data is not sufficient for budget allocation, and notes that more than 20 companies now sell AI visibility tools on methodologies that can return different answers for the same brand. Its own headline number: only 16% of brands systematically track AI visibility today.

2. Google will auto-upgrade some Search campaigns to AI Max on 1 September.

An email went to advertisers on 5 August confirming that campaigns using automatically created assets or the campaign-level broad match setting will be converted. The defaults differ by trigger: automatically created assets brings search term matching and text customisation on, while campaign-level broad match brings search term matching only. To avoid it, switch the legacy setting off or enable AI Max yourself before then. Dynamic Search Ads are not part of this wave, that migration was pushed to February 2027.

3. Three Meta Ads reports went dark on 6 August without an error message.

The device, hourly and frequency breakdowns became opt-in per ad account with Marketing API v26.0. Requests on accounts that never opted in return HTTP 200 with an empty data array, so no tool alerts you and your dashboard simply renders zeroes. The toggle sits under Additional Breakdowns in Ads Manager’s reporting view, there is no Meta help centre article for it, and both Supermetrics and Improvado issued breaking-change notices to customers. Most sources say the missing data does not backfill, so the sooner you switch it on the less you lose.

4. ChatGPT Atlas stops working tomorrow, 9 August.

Bookmarks, open tabs and history do not transfer automatically. Cookies and passwords can be exported to the ChatGPT desktop app and bookmarks to Chrome. It launched in October 2025 and does not reach its first birthday.

5. Google search rankings moved sharply between 1 and 3 August,

…Picked up by third-party trackers and heavily discussed in the SEO community. As of 5 August Google had not confirmed an update and its Search Status Dashboard listed no incident. That is the fourth unconfirmed tremor since the confirmed spam update in late June.

6. Search Console’s new generative AI performance data is more misleading than it looks.

Roughly eight weeks after the reports launched, the core problem is that the central-tendency formula treats position one inside an AI Overview as equivalent to position one in classic organic results, which distorts any trend you try to read from it. Presence metrics also arrive with no click or query detail attached.

7. A correction to something we ran on 25 July.

When we covered Claude’s Record a skill feature we flagged that Anthropic had published nothing about what happens to your recordings. That gap has now been filled by hands-on documentation rather than by Anthropic: Claude discards the raw audio and video after processing and keeps only the screenshots it needs for context, which live inside the Cowork task and disappear with it. Two details we should have carried at the time: recordings cap at about 10 minutes, and the feature is Mac only, with no Windows support and no availability on Free or Enterprise.

8. HubSpot’s Agent Hub credit pricing has firmed up…

…closing a question we left open on 2 August. Published rates are $0.50 per resolved conversation for the Customer Agent, $1.00 per recommended lead for the Prospecting Agent, $0.10 per answer for the Data Agent, and around $10 per piece for the Content Agent. Included allowances run 500 credits a month on Starter, 3,000 on Professional and 5,000 on Enterprise, pooled account-wide with no rollover. Agent Hub itself remains free for Professional and Enterprise and is still in public beta with no general availability date.

9. Anthropic has added noindex and nofollow to shared Claude conversations…

which resolves the problem we reported on 2 August when shared chats began appearing in Google Search. The original cause was a robots.txt block sitting in front of a noindex header, so Google never crawled the pages to see the instruction.

10. Claude Code’s auto mode becomes the default permission setting on 14 August…

…for Pro, Max and Team users, unless a user or admin has pinned something else. It replaces repeated approval prompts with a classifier that checks each action for anything irreversible or destructive, falling back to manual approval after repeated blocks. Anthropic says it will not charge for the extra tokens the classifier uses.

First, provenance stopped being a policy question and became infrastructure.

Anthropic is now marking Claude’s output at the model level, globally, on every surface, because of a European law that took effect thirteen days earlier. Perplexity started downgrading a publisher’s trust score for showing bots different content from humans. LinkedIn is penalising AI-written comments, as we reported last week. Three different mechanisms, three different companies, one direction of travel. The question is shifting from “can anyone tell this was AI” to “can you show where this came from”, and that is a better question for anyone doing serious work, because it rewards the parts of the job that were always the point. Original data, named sources, a real opinion and a person willing to attach their name to it are not detection-avoidance tactics. They are the things that make content worth citing, and they now happen to be the things that make provenance easy to demonstrate.

Second, memory turned out to be a file-system problem, and the teams winning have already worked that out.

Six separate developments landed on the same question this week: who decides what your AI remembers. Jeff Su’s three levels, the skills that now follow you into Chrome, OpenAI letting you change a project’s memory mode after the fact, Computer History watching what you do, Nate Herk finding his instructions were constraining thinking rather than improving it, and the wider pattern of practitioners moving their context into plain files they can open and edit.

The unifying insight is Su’s, that at the first two levels “the AI is still the author”, deciding what gets kept and what gets dropped without telling you. What is striking is how unglamorous the fix is. Writing down your ideal customer profile, your pricing, your services, your team structure and your standard operating procedures is not an AI project. It is the thing you have been meaning to do for three years, and it is now the highest-leverage AI work available to most teams. This is the fourth consecutive week some version of that conclusion has appeared here, which is either a very persistent coincidence or the actual answer.

Third, the models are being handed more judgment at the exact moment marketers are downgrading their own.

Boris Cherny is telling people to delete their carefully written instructions and let the model think. Anthropic stripped most of its own system prompt. Nate Herk’s test found the unconstrained version produced better thinking and worse formatting, which is a trade most people would take.

Set that against the finding we ran on 8 August, that marketers rated critical thinking, communication, collaboration and adaptability as less important in 2026 than they did in 2025. Those are the four things the same research places furthest from automation. The uncomfortable version of this week’s news is that the industry is being told, from two directions at once, that judgment is the scarce input. One direction is telling you to stop over-instructing the machine. The other is quietly recording that we have decided judgment matters less than it used to.

Tell me: What’s your biggest takeaway from this weeks news from the world of AI Marketing? Love to hear your thougths in the comments.

Leave a comment

Thanks for reading the Marketing with AI Weekly Roundup! If you enjoyed this read, the best compliment I could receive would be if you shared it.

Marketing with AI Weekly Roundup: Google Resets Your Bid Targets in Nine Days, the AMA Names the Marketing Jobs AI Is Taking, and LinkedIn Starts Marking Down AI Comments

Robert Gillespie — Sun, 09 Aug 2026 12:29:28 GMT

Welcome to issue #46 of the Marketing with AI Weekly Roundup. Published every Sunday, it shares the top stories, playbooks, data and quick hits from the world of Marketing with AI over the last 7 days.

Subscribe to receive the week's most important AI marketing stories direct to your inbox.

Subscribe now

NEW: For 6 months I've asked every marketer I train the same question: where could AI save you time? Most aren't sure. So I turned everything those trainings taught me into a $5 tool that tells you. Your tasks, your hours back, your first workflow to build.

Learn more about the Operator Gameplan

Last weeks edition was about disclosure, whether you have to tell people you used AI. This week the questions got more personal: what your numbers will say on 18 August, and what your job will look like in two years.

On 17 August, nine days from now, Google changes how Smart Bidding treats campaigns that are limited by budget. If you run a lead gen campaign that has quietly been beating its target for months, that campaign is the one that gets hurt, and nothing will land in your inbox on the day it happens.

On Tuesday the American Marketing Association published the clearest data anyone has produced on which marketing roles AI is actually eating, and its answer is not the one the industry keeps telling itself.

And on Wednesday the most widely followed independent LinkedIn researcher reported that the platform has started scoring individual comments, with AI-written ones counting against you.

Elsewhere, Adobe put more than seventy of its tools inside ChatGPT with a free tier that needs no Adobe account, three Meta Ads reports went dark without an error message, and the IAB published the first industry standard for measuring AI visibility, arriving a good while after the money did.

Let’s get into it…

🔥 Top Stories

1. Google resets what your bid targets mean on 17 August, and the campaigns beating their targets are the ones that get hurt

From 17 August 2026, Google Ads changes how Smart Bidding behaves on campaigns that are limited by budget. Today, a budget-capped Target CPA or Target ROAS campaign often overperforms, delivering acquisitions at a cost well below the number you typed in, because the budget cap is doing the constraining rather than the target. From 17 August, the system optimises toward the target you actually set, so those campaigns drift upward toward it.

Two things this is not. Campaigns with no budget ceiling are unaffected. And the auction itself is not changing, this is a bidding change only. Google has publicly pushed back on the reading that the update amounts to letting the system spend more money, stating it will not lead to spend increases.

In scope: Target CPA, Target ROAS, and Target CPC for Demand Gen, across Search, Shopping, Performance Max, Demand Gen and Travel campaigns, and across Google Ads, Search Ads 360, Display & Video 360, Google Ads Editor and the API. Not affected: App campaigns, Video reach and Video view campaigns, and any campaign not limited by budget.

The change was announced back on 15 June by Google Ads Product Liaison Ginny Marvin, and we have carried it in Looking Ahead since 11 July. What arrived this week is the operational guidance, a Q&A from Marvin answering the questions advertisers were actually asking, covered on 6 August. The deadline is now nine days away and the how-to finally exists.

Why it matters / what to do: picture a lead gen campaign permanently capped by budget, delivering leads at £22 against a £35 target you set eighteen months ago and have not looked at since. Google is about to start treating £35 as the goal rather than the ceiling. That is a 59% increase in your cost per lead, arriving with no email on the day. Here is the actual to-do list. First, filter your account for campaigns with the status Limited by budget running Target CPA or Target ROAS, which is your exposure list. Second, put a column for actual CPA or ROAS over the last 60 days, not last week, next to the stated target. Third, flag every campaign beating its target by a wide margin, because where the two numbers are already close the impact is minimal and you can leave it alone. Fourth, decide per flagged campaign: lower the target to match what it genuinely achieves, set a number that reflects what works commercially, or keep the target because it is still right and accept the drift. Fifth, write down what you expect to happen before 17 August, so you are not reconstructing the comparison afterwards from memory. Then leave it alone, because the rollout takes a few weeks with a calibration period and panic-fixing on 18 August will teach you nothing.

Three caveats. Google’s own Bid Target Adjustment Tool, live since 6 July, will suggest changes for you, but do not apply them blindly, because it has no idea what your business can actually afford per lead. Performance Planner forecasts are explicitly unreliable between 17 and 31 August while the tools recalibrate, so do not commit budget off those numbers. And Google’s own FAQ warns against reactive fixes such as data exclusions or bid limits applied purely because of this update.

2. The American Marketing Association put numbers on which marketing jobs AI is taking, and the profession is deprioritising exactly the skills that survive

On 5 August 2026 the American Marketing Association released its 2026 State of Marketing Careers Report, a mixed-methods study built on a survey of 1,412 marketing practitioners conducted in December 2025 and January 2026, supplemented with interviews, job-posting analysis and secondary research.

The share of marketing job postings mentioning AI nearly doubled across 2025, from 8% in January to 15% in December. Over the same period total marketing job postings fell by more than 8%, and US marketing employment remains roughly 27% below its March 2020 level. The roles that lost the most ground were social media coordinator, copywriter and SEO manager. Senior leadership held up considerably better than middle management and individual contributor roles.

The AMA’s own summary is the line worth carrying: “Companies are not hiring less judgment. They are hiring less execution.”

The report scores marketing tasks against Stanford’s Human Agency Scale, running from H1, where AI handles a task alone, to H5, where AI cannot function without continuous human involvement. The disciplines it places at highest disruption are email marketing, SEO and paid media, on the grounds that they are rules-based and repetitive. Routine reporting, data collection and campaign monitoring sit at the automatable end, alongside scheduling and social monitoring.

Then there is the finding that should genuinely bother you. Marketers rated critical thinking, communication, collaboration and adaptability as less important in 2026 than they did in 2025. Those are precisely the four capabilities the AMA’s own analysis places at the top of the human-involvement scale. The profession is actively demoting the things that are hardest to automate, at the exact moment the execution work is being automated out from under it.

Why it matters / what to do: this is the closest thing to an honest answer to “am I going to be fine” that anyone has published this year, and the answer is conditional rather than reassuring. Action: look at the report’s H3 zone, where humans and AI working together outperform either alone, which the AMA names as content making and data storytelling. That is where your next quarter of learning should go, rather than into another tool. And if your job title is one of the three that lost ground, the practical move this month is to attach yourself to a decision rather than to a deliverable, because the deliverable is the part being commoditised. Three caveats worth holding: the AMA is a professional body with a commercial interest in marketers continuing to invest in professional development, the job-posting analysis is US-based, and the skills-importance figures are self-reported perception rather than measured outcomes.

3. LinkedIn has started scoring individual comments, and AI-written ones now count against you

On 5 August, Richard van der Blom published his latest LinkedIn algorithm findings, the research series most practitioners in this space treat as the standard independent reference. Four things in it matter.

Reach and engagement fell again. July was down 18% in reach and 16% in engagement compared with June.

LinkedIn is now rating individual comments, not just posts. More than ten factors decide whether your comment gets distribution or quietly dies, and one of them is whether the comment was partly or fully written by AI, which is now a negative signal.

Send in DM has overtaken saves and reposts as the strongest distribution signal. Testing across more than 800 posts and 2,400 DM sends found that posts sent privately saw roughly a 60% uplift, and the effect pushes distribution outside your own network.

And “Celebrate an Occasion” posts are being boosted hard, up to 2.6 times the reach they got in March 2026. That looks like a deliberate rebalance back toward network signals after months of complaints that the interest-based feed had buried people’s actual connections.

Why it matters / what to do: LinkedIn is the single most important organic channel for most B2B marketers, and last week it gave every member a button to report a post as AI slop. This week we learn the same logic has reached the comment box, which is where most B2B teams do their AI-assisted volume work. Comments are the part of LinkedIn activity people most often automate, and it is now the part with an explicit penalty attached. Action: stop running AI-generated commenting of any kind this week, and if you use a tool that does it, turn it off. Then change what you optimise for. Asking someone to like your post is worth much less than writing something a reader would send to a colleague, so make the post useful enough to forward rather than easy to applaud.

One caveat, and it is important. This is independent practitioner research published by a consultant on LinkedIn, not an announcement from LinkedIn, so treat the figures as well-sourced observation rather than confirmed platform mechanics. Worth adding, because a lot of it went around this week: the wider claims circulating that LinkedIn has just changed its algorithm to favour expertise and consistency mostly trace back to agency blog posts with no LinkedIn attribution at all, and the underlying feed rebuild they describe actually shipped in March 2026.

4. Adobe put more than 70 tools inside ChatGPT, and the guest tier needs no Adobe account

On 6 August 2026 Adobe released a single Adobe plugin in ChatGPT, replacing the three separate connectors it shipped in December 2025 with one plugin reaching more than 70 tools across Adobe Express, Photoshop, Firefly, Premiere, Acrobat, Lightroom, Illustrator, InDesign and Adobe Stock. You describe the outcome in plain language and the plugin picks the tools. It is available globally from launch, on ChatGPT web and the desktop app, and works across ChatGPT, ChatGPT Work and Codex.

The setup path: open ChatGPT, go to Plugins in the left toolbar, search “Adobe”, install the one labelled “Adobe Design, combine, and edit”, then start a new chat and type @Adobe. You can continue as a guest. Signing in with an Adobe account unlocks Firefly image and video generation, access to your Creative Cloud files, and saving work between sessions.

The workflows Adobe itself names are unusually marketing-shaped: apply one consistent style across a whole batch of campaign photos and save them to Creative Cloud, turn images into product or packaging mock-ups, cut a long video into a highlight reel and reformat it for Shorts and Reels, customise a template’s copy and colours then export a print-ready PDF, or upload a dataset and get back a finished PDF such as an event badge set or a catalogue. It can also search your Creative Cloud library by loose descriptions like style or mood, which is probably the feature you would actually use daily.

Why it matters / what to do: on a small B2B team the constraint on shipping campaign assets is rarely the idea, it is the twenty minutes of resizing, recolouring and exporting that sits between the idea and the post. That work has just moved into the same window where the brief was written. Action: take one asset you resize every single week, a webinar promo or a conference banner, run it through @Adobe as a guest, and compare the output against what your designer or your Canva template produces. Decide on the evidence rather than on the demo. Three caveats: it is web and desktop only at launch with no mobile, how much you can generate depends on your ChatGPT plan and Adobe has not published the per-tier limits, and guest access saves nothing between sessions.

📋 The Playbook

1. Turn your team’s internal conversations into a daily LinkedIn content engine

Cody Schneider returned to Greg Isenberg’s channel on 5 August with a second system, and this one needs no engineering at all. He sells a platform and implementation services for exactly this kind of work, so treat him as an operator with a commercial interest rather than a neutral tutorial. The method survives that.

The rule the whole thing rests on: never ask a model to invent the content. His words: “If you go and you try to just have the agent like think about this, you’re like, ‘Write good LinkedIn content.’ It’s going to be the most mid thing, or you’re going to get flagged for AI slop by LinkedIn’s new feature that just released this morning.” That is a direct on-camera reference to the story we ran last Sunday, and it is the same conclusion from the other direction.

The workflow, step by step:

1. Generate source material on a cadence. His method is a weekly one-to-one call with each person on the team, deliberately unstructured: “just like tell me everything that like you’ve learned in the last week.” No agenda required.

2. Or mine the source material you already have. Sales call recordings, Gong transcripts, the sales channel in Slack, Notion, internal comms. His example of what is hiding in there: a prospect explaining why they did not buy, which he calls “an unbelievable piece of content.”

3. Extract the insights, then write one post per insight. He is candid that the writing model barely matters here, saying Claude Sonnet is “probably good enough on the writing side”, because the raw material is doing the work.

4. Publish across several named people, not just the company page, since personal accounts carry organic reach that brand pages do not.

5. Feed the performance data back into the loop, so the next writing cycle knows which insights earned attention. His two instruction words, used literally: snowball and remix.

6. Recycle the winners on a 90-day cycle. “If you look at my Twitter post as an example or even my LinkedIn, it is the exact same thing remixed every 90 days, like full stop. I can’t post it every day. You post it every 90 days.”

The number that makes this a budget conversation. He prices earned reach against paid: “on LinkedIn, it’s like $22 per thousand impressions is the average. Every post that you get, even with an account that’s like 500 followers, you can get a thousand impressions. That’s like $20 that you just like put into your pocket for free.” That is a defensible way to put a figure on organic in a planning meeting, though the $22 CPM is his number rather than a published LinkedIn rate, so attribute it to him.

The provocation, included because it deserves arguing with. “I actually think the social media manager job, like full stop, it’s already dead.” Isenberg pushes back and they land on evolved rather than dead. Set that against this week’s AMA finding that social media coordinator is one of the three roles losing the most ground, and it is less throwaway than it sounds.

And the alternative for anyone who will not build a personal brand: build a topic page instead of a personal one. Isenberg’s worked example is Julian Shapiro, who ran his agency Demand Curve behind an account called @GrowthTactics rather than a branded handle.

Do this now: record your next internal team call, pull three insights out of the transcript, and post them under three different named people this week. That is the entire system at its smallest possible scale, and it costs nothing.

Read more:

2. Set up Obsidian as the shared file layer that both Claude and ChatGPT can read and write

Paul J Lipsky published this on 5 August, and it is the most useful beginner build of the week because it solves a problem most people have not noticed they have: everything you teach an AI tool currently lives inside that tool. No sponsor is disclosed and the Obsidian link in his description is the plain product URL.

The argument in his words: “Obsidian operates entirely on local markdown or text files stored directly on your computer. This gives you complete ownership and control over your notes and you could switch to another tool at any time.” And on why it pairs so well: “tools like Claude Cowork and ChatGPT love markdown files. And since Obsidian uses markdown files too, it’s kind of a match made in heaven.”

The setup. Download from obsidian.md and create a vault, which is simply a folder. Save it somewhere backed up, he uses iCloud Drive and notes Google Drive or Dropbox work equally well. Formatting is plain markdown: # plus a space for a heading, - plus a space for a bullet, - space [ space ] space for a checklist item. Organise with folders and subfolders, right-clicking a folder to create a new note or subfolder inside it.

The thing that surprises people: the vault is an ordinary folder in Finder, so you can drag any file straight into it and it appears in Obsidian. PDFs and images render in place.

One settings fix you will need. Excel files do not show up by default. Go to Settings, then Files and links, and switch on Show all file types. They still will not open inside Obsidian, but at least you can see they exist. While you are there, turn on automatic updates under Settings, General.

Connecting Claude. With Cowork selected, use the dropdown, choose add a folder, navigate to your vault, select the whole vault (or a single subfolder if you want to limit what it can reach), open, then always allow. His demonstration prompt is worth copying: “create a new note for me in this vault that is an index of everything in the vault.” Claude writes a vault index file back into the folder, with working internal links that click through to the real notes.

Connecting ChatGPT Work. Select Work at the top, choose your project, new project, select your vault folder, open, then create project. His framing: “projects are just folders.”

The trick that makes it usable day to day. When you want the assistant to work on the specific note you are looking at, click the three dots at the top right of the note, hover copy path, click from vault folder, and paste that path into the chat. His example instruction after pasting: “Restructure this file with nicer headings, emojis, checklists, etc.”

Do this now: create a vault, drop in your positioning document, your ideal customer profile and your brand voice notes, point Claude at it and ask for an index. You have just built a knowledge base that both Claude and ChatGPT can read, that lives on your own machine, and that neither of them owns.

Read more:

🛠️ New Tools & Features

1. HubSpot published a head-to-head against Ahrefs, with real prices on both

On 6 August HubSpot published a direct comparison of HubSpot AEO against Ahrefs Brand Radar, the two mainstream tools for tracking whether AI answer engines mention your brand. HubSpot AEO is $50 a month, or $45 annually, covers ChatGPT, Gemini and Perplexity only, and is wired into the CRM so its prompt suggestions come from your actual customer data. There is a free 28-day trial with 25 prompts and no credit card. Ahrefs Brand Radar covers seven platforms, adding Google AI Overviews, AI Mode, Microsoft Copilot and Grok, and is built on a database of 406 million real monthly search prompts. Standalone it runs $199 to $699 a month, but it is bundled free into Ahrefs plans from $29 a month, so if you already pay for Ahrefs you effectively have it. Worth knowing: Ahrefs added Claude tracking to Brand Radar at the end of July, as a custom-prompt-only source that costs eight checks per update against one for other platforms. Neither product tracks everything, and the honest split is breadth versus action. Caveat, and it is the obvious one: this is HubSpot’s own comparison of its own product against a competitor, and while the Ahrefs pricing checks out against Ahrefs’ published tiers, the framing does not.

2. Microsoft Clarity now groups your AI citations by topic

Announced in Microsoft Advertising’s first monthly product newsletter on 4 August, Topic Insights inside Microsoft Clarity’s AI Visibility reporting groups your AI citations by subject rather than listing them individually, so you can see which topics AI systems associate with your brand and where the holes are. Three metrics come with it and the vocabulary is worth learning. Grounding queries are the retrieval searches an AI system runs before it writes an answer, which is not the same as what the user typed. Citation share is how often your domain appears as a source. Share of authority is how often you are cited compared with competing sources, which is a genuinely competitive number in a category where most tools only report your own presence. Clarity is free, which makes this the cheapest topic-level AI visibility data available without buying a subscription.

3. Perplexity put Statista, CB Insights and PitchBook inside its research agent

On 6 August Perplexity announced Premium Sources, bringing paywalled professional research directly into Perplexity Computer. At launch that means Statista, CB Insights and PitchBook, with CB Insights alone contributing more than 8,000 research reports. Market sizing, category trend data and competitor funding history are three things B2B marketers routinely approximate in a deck because the real sources cost five figures, so a Statista or PitchBook citation in a positioning document is a different quality of evidence. The pricing genuinely does not reconcile and you should check your own plan before relying on it. Perplexity’s enterprise guide describes Premium Sources as built into your subscription at no extra cost, while a separate Perplexity post describes a much tighter allowance of three premium searches a month on free, five on Pro and ten on Enterprise Pro, and names a different set of providers. Those two descriptions cannot both be right.

4. Gemini Notebook’s cloud computer finished rolling out, and Workspace can now keep it fed

Two connected updates. On 4 August the upgraded Gemini Notebook experience was confirmed as fully rolled out to all Google AI Pro subscribers on the web, nineteen days after it was announced. That means every notebook now has a secure cloud computer that can write and run real code against your own uploaded sources, returning charts, PDFs, spreadsheets and slide decks rather than descriptions of them. The 20-second test for whether your account has it: upload a table and ask for a chart or a spreadsheet, and if you get a downloadable file it is active. Then on 6 August Google Workspace added an “Add a source to Gemini Notebook” step to Workspace Studio, so an automation can keep a notebook current on a schedule by feeding it Drive files, YouTube URLs or web pages. The second half is the more interesting one for marketers: a competitor-monitoring notebook that refills itself rather than going stale after a fortnight. Caveat: Workspace Studio steps are a Workspace feature, so that half does not apply to personal Gmail accounts.

5. Microsoft’s Ad Preview Hub reached Performance Max, with Bing search previews

Also in Microsoft Advertising’s 4 August product newsletter: the Ad Preview Hub now supports Performance Max campaigns, and adds a Bing search results preview alongside the existing MSN and Outlook placements. You get a shareable public link that expires after 30 days. This is the least glamorous item here and it is on the list anyway, because seeing what your ad actually looks like across placements before it goes live is a real marketing job that Performance Max has made unusually hard. The shareable link is the useful part for anyone who has to get creative signed off by someone without a Microsoft Ads login.

📊 Research and Data

1. Forrester: 88% of B2B marketing organisations have adopted AI, and the foundations have not kept up

Published 5 August and drawn from Forrester’s Marketing Survey 2026, covering more than 1,000 B2B marketing decision makers, this is the largest and most credible B2B sample of the week. 88% of B2B marketing organisations have adopted AI tools or built their own. Alongside that, 85% say they run integrated campaigns and 90% of those say integrated programmes deliver more value than separate initiatives. Yet the same leaders report unclear AI strategy, difficulty measuring impact, data infrastructure problems and uncertainty about where AI should even be applied. Forrester’s framing: “marketing’s ambitions, priorities, and accountability are expanding faster than the underlying capabilities required to support them.” The named top challenges are poor data quality and accessibility, difficulty measuring performance, and insufficient insight for decision-making.

What to do: adoption has stopped being a differentiator, because almost everyone has adopted. The recommendation that follows is unglamorous and correct: treat data infrastructure, credible measurement and clear ownership as strategic investments rather than housekeeping. Practically, if you are choosing between buying another AI tool and spending a month making one dataset clean and reachable, the second is now the competitive move. Caveat: Forrester sells advisory services on precisely this problem, and the full methodology sits behind the report.

2. TransUnion: 89% will spend more on AI, 36% think their data can take it

Also published 5 August. TransUnion commissioned United Talent Agency’s brand advisory division to survey 100 senior marketing and technology leaders, all director level or above, all at US enterprise brands with marketing budgets of at least $50 million, and all already using AI in core marketing workflows rather than just for productivity. TransUnion calls the result the AI Confidence-Readiness Paradox.

89% expect to increase spending on AI-enabled marketing over the next 12 to 24 months and 64% are confident of hitting their AI goals. But only 42% rate their people readiness as high and only 36% say the same of their data and processes. Just 13% have fully scaled AI across the enterprise, and 53% report meaningful ROI. 65% measure AI’s impact primarily through estimated cost or time savings.

The split that tells the story: 75% say AI has reduced manual effort or time spent, but only 44% have reduced media waste, 35% have improved conversions or response rates, and 30% are reaching more of their target audience at similar spend.

What to do: AI is reliably saving time and unreliably making money, and two thirds of teams are measuring the thing it is good at rather than the thing they are paid for. Action: pick one AI-assisted activity and measure it on a business outcome instead of hours saved. If you find you cannot, that is itself the finding, and it is a more useful one to take to your leadership than another efficiency number. Three caveats, all worth stating: the sample is small at 100, it is enterprise-only with $50m-plus budgets so it does not describe a ten-person B2B team, and TransUnion sells identity and marketing data, which makes “marketers have a data problem” a conveniently placed conclusion.

3. Marketers are now putting 24% of search and content budgets into AI visibility

Reported by Digiday on 7 August, pulling together several vendor surveys. From Fractl: marketers route roughly 24% of search and content budgets to AI visibility work, 82% have allocated at least something, 18% allocate nothing at all, and 43% put more than a fifth of that budget into it. By function, performance marketing at 32% and SEO at 31% spend the highest share, while brand at 21% and content at 20% spend the least.

Two other numbers in the same round-up are worth pairing with it. 10Fold found 52% of B2B marketers now rate AI answer engines as their most effective distribution channel, against 29% for organic search, while 41% have only a quarter to a half of their content optimised for AI discovery. And Similarweb data put AI search at 35% of initial product discovery against 13.6% for traditional search, with the gap narrowing at the purchase stage to 24.3% against 22.1%.

What to do: this is the “am I behind” number for the week, and the 10Fold split is the sharper half. B2B marketers are already rating answer engines above organic search as a distribution channel while most of their content is not built for it, which is a gap you can close with editing rather than budget. Caveats: Fractl and 10Fold are both agencies with a commercial interest in AI visibility work, neither sample size nor fieldwork dates were published in the coverage, and the Similarweb figure is a behavioural estimate rather than a survey. Attribute all three to the Digiday round-up and be explicit that the samples are unstated.

💡 Quick Hits

1. The IAB published the first industry standard for measuring AI visibility.

2. Google will auto-upgrade some Search campaigns to AI Max on 1 September.

3. Three Meta Ads reports went dark on 6 August without an error message.

4. ChatGPT Atlas stops working tomorrow, 9 August.

5. Google search rankings moved sharply between 1 and 3 August,

6. Search Console’s new generative AI performance data is more misleading than it looks.

7. A correction to something we ran on 25 July.

8. HubSpot’s Agent Hub credit pricing has firmed up…

9. Anthropic has added noindex and nofollow to shared Claude conversations…

10. Claude Code’s auto mode becomes the default permission setting on 14 August…

👀 On Our Radar

Time is selling ads that only AI crawlers can see, and every bot read is billed as an impression.

Digiday broke the business story on 30 July, reporting that Time had partnered with adtech vendor Mobian on a format they call agent ads. On 5 August developer Vincent Schmalbach published the technical teardown that proved the mechanic, and The Register independently replicated it. Request a Time page as Chrome, Safari or Googlebot and you get the normal HTML, roughly 303,000 bytes. Request the identical URL in the same second as ClaudeBot, PerplexityBot or OAI-SearchBot and you get a stripped 13,000-byte markdown version with sponsored content baked in that no human ever sees. What sits in the machine-only version is a full Ally Bank FAQ, labelled sponsored, written in Ally’s own marketing language, with structured data and campaign-tagged tracking links. Ally Bank and the Project Management Institute are the first two advertisers.

The mechanic is the interesting part. Response headers carry a fresh unique ID on every single request paired with a no-store cache instruction, so every bot read is logged as a distinct ad impression, and a second header names the unit being counted: tokens fed into a model. Not a person, not a pageview. Mobian’s chief executive put the pitch plainly, that the aim is to influence the assistant rather than the reader, because with a human you only influence one person.

Ignore the buying side, which is not available to you. The useful part is free and takes ten minutes. Fetch one of your own key pages six times, changing only the user agent, as a normal browser and then as Googlebot, ClaudeBot, OAI-SearchBot, PerplexityBot and GPTBot, and compare the six responses. If your CDN, your CMS or a plugin is already serving assistants something different from what it serves buyers, you want to find that out yourself rather than when an assistant summarises the wrong version to a prospect. Two caveats: this is one documented example, and Time deliberately excludes Googlebot, which reads more like avoiding a cloaking penalty than a principled position.

First, the money moved into AI visibility before anyone agreed how to measure it. This week produced the cleanest piece of arithmetic we have run in months. 82% of marketers now put budget into AI visibility, at an average of 24% of search and content spend. 16% of brands systematically track it. Those two numbers describe the same people. On the same day, the IAB published the first standard for what good measurement even looks like and drew a line most vendor dashboards do not cross, saying in plain language that directional data is not good enough to allocate budget against. Microsoft then shipped topic-level citation reporting for free, and HubSpot and Ahrefs went head to head on price. The industry is closing the gap, but it is closing it in the right order only by accident, and the practical consequence for you is that you can now hold a vendor to a document instead of a demo.

Second, the honest measure of an AI programme is what you count, not what you adopted. Forrester found 88% adoption across more than a thousand B2B marketing decision makers. TransUnion found 13% fully scaled and 36% who trust their own data. Underneath both sits the number that should sting: 65% of marketing leaders measure AI’s impact primarily by hours saved, while 75% report saving time and only 35% report better conversions. This is the third consecutive week this newsletter has landed on some version of “everyone adopted, almost nobody rewired”, and it is worth saying that out loud rather than dressing it up as a new insight. What has changed is that the framing can now be sharper. Adoption is finished as a story. The thing separating teams is whether they are measuring AI on the outcome marketing is actually paid for, and most are not, because hours saved is the easier number to produce.

Third, the work being automated and the work being valued are pulling apart, and marketers are backing the wrong one. The AMA put numbers on it: execution roles losing ground, judgment roles holding, and the profession itself rating critical thinking, communication, collaboration and adaptability as less important than it did a year ago. Read that against LinkedIn penalising AI-written comments, and against a week in which Adobe moved seventy production tools into a chat window. The production layer of marketing is being commoditised in public, quickly, and the response from a lot of teams has been to get better at production. The uncomfortable implication is the same one the research keeps arriving at from different directions: the highest-return thing available to most marketers right now is not a tool at all. It is being the person who decides what should exist and why.

🔮 The Big Picture

First, the money moved into AI visibility before anyone agreed how to measure it.

This week produced the cleanest piece of arithmetic we have run in months. 82% of marketers now put budget into AI visibility, at an average of 24% of search and content spend. 16% of brands systematically track it. Those two numbers describe the same people. On the same day, the IAB published the first standard for what good measurement even looks like and drew a line most vendor dashboards do not cross, saying in plain language that directional data is not good enough to allocate budget against. Microsoft then shipped topic-level citation reporting for free, and HubSpot and Ahrefs went head to head on price. The industry is closing the gap, but it is closing it in the right order only by accident, and the practical consequence for you is that you can now hold a vendor to a document instead of a demo.

Second, the honest measure of an AI programme is what you count, not what you adopted.

Forrester found 88% adoption across more than a thousand B2B marketing decision makers. TransUnion found 13% fully scaled and 36% who trust their own data. Underneath both sits the number that should sting: 65% of marketing leaders measure AI’s impact primarily by hours saved, while 75% report saving time and only 35% report better conversions. This is the third consecutive week this newsletter has landed on some version of “everyone adopted, almost nobody rewired”, and it is worth saying that out loud rather than dressing it up as a new insight. What has changed is that the framing can now be sharper. Adoption is finished as a story. The thing separating teams is whether they are measuring AI on the outcome marketing is actually paid for, and most are not, because hours saved is the easier number to produce.

Third, the work being automated and the work being valued are pulling apart, and marketers are backing the wrong one.

The AMA put numbers on it: execution roles losing ground, judgment roles holding, and the profession itself rating critical thinking, communication, collaboration and adaptability as less important than it did a year ago. Read that against LinkedIn penalising AI-written comments, and against a week in which Adobe moved seventy production tools into a chat window. The production layer of marketing is being commoditised in public, quickly, and the response from a lot of teams has been to get better at production. The uncomfortable implication is the same one the research keeps arriving at from different directions: the highest-return thing available to most marketers right now is not a tool at all. It is being the person who decides what should exist and why.

Tell me: What’s your biggest takeaway from this weeks news from the world of AI Marketing? Love to hear your thougths in the comments.

Leave a comment

Thanks for reading the Marketing with AI Weekly Roundup! If you enjoyed this read, the best compliment I could receive would be if you shared it.

The Marketer's Guide to Claude Code Goals and Loops

Wed, 05 Aug 2026 12:01:26 GMT

Every Wednesday, one essay on what's actually working in Marketing with AI - tools, workflows, and mindset shifts you can apply immediately.

Subscribe now

The Marketer’s Guide to Claude Code Goals and Loops

Your email went out on Tuesday promising 20% off.

The landing page it linked to said 25%.

The social post had the old CTA, the one pointing at a page that was archived back in March.

Nobody noticed until a customer replied asking which discount was real.

None of it was badly made, though. Every asset was fine on its own. Six pieces were produced in sequence, each correct, and no one read them against each other at the end.

The mistake didn’t live inside any single content piece. It lived in the gap between them.

That’s the most common failure in marketing execution, and it has almost nothing to do with talent. It’s structural.

When work is produced in separate pieces or silos, what goes wrong is how the pieces line up, and that is the one thing not everybody checks.

AI has the exact same failure mode, and this year Claude Code got two commands that fix it.

They’re called /goal and /loop.

And unless you follow the developer world closely, you’ve probably not heard of either, because they launched as coding tools.

But they aren’t.

And they’re the most useful thing to happen to marketing execution in months, and I want to show you why.

First, why Goals and Loops is not just another feature

Think about where most of us have actually landed with AI.

You’ve learned to prompt. You write a good brief, you give context, you know a vague ask gets a vague answer.

You’ve maybe built a few Skills, set up a Project, and connected Claude to your files.

The good news…that already puts you ahead of most marketers.

But there’s a limit to prompting, and it’s easy to miss because it’s not obvious.

Every prompt, however good, gets one attempt. You ask, Claude does a pass, Claude hands back a response. If it covered eleven of your twelve points and quietly skipped the twelfth, nothing in the system notices. The response ends because responses end, not because the work is done.

The marketers pulling ahead now have made a different move.

They’ve stopped writing better instructions and started defining better outcomes.

Less “here’s what to do,” more “here’s what finished looks like, and don’t stop until it’s true.”

And this is all now possible inside Claude Code.

Not Cowork, not the chat window.

If you’ve only ever used Claude through those, this is the feature worth opening Claude Code for. I wrote a full guide to Claude Code for non-technical marketers earlier in this series, so if the name makes you nervous, start there and come back.

You don’t write code.

You describe outcomes in plain English, and that’s the whole job.

What /goal actually does

/goal is a command you type at the start of a task. You follow it with a condition, in plain English, describing what “done” means.

/goal every deliverable in this campaign uses the same offer, price and
CTA link, shown asset by asset in a table before you finish

That condition becomes a standard the work has to meet.

Claude can’t end its turn and hand back to you until the condition is actually satisfied. It produces the work, checks it against your condition, spots what’s missing, goes back and fixes it, and only stops when the thing you asked for is genuinely true.

Two details make it feel different in practice.

Typing the command starts the work immediately, so the condition is the brief and you don’t need a separate prompt. And when the condition is met, the goal clears itself, so you’re not managing anything.

That might sound like a small thing. But it isn’t, and the reason becomes clearer when you really think about it.

Every AI answer you have ever received ended for the wrong reason. It ended when the model reached the end of its response, not when the work reached the end of its requirements.
Those are two completely different finish lines, and most of the time you can’t tell them apart, because a stopped-early answer looks exactly like a finished one.

Looks finished and is finished are not the same.

A campaign where the email and the page disagree on price looks finished.

A site audit that covered nine of your twelve pages looks finished.

Three social posts where two open the same way look finished.

/goal is the difference between asking Claude to do something and telling it what done means.

Watch this 60-sec explainer video to help you better understand Goals and Loops

NB: This explainer video was created in less than 10 minutes using Claude Design...and I have no video creation skills whatsoever.

If you want to create all kinds of on-brand assets for your business, from animations to slides, brochures, diagrams, social graphics and so much more…I’ve just added a full 2-hour Claude Design training to the Claude for Marketers Masterclass to show you how.

Learn More About Claude for Marketers

There is genuinely a second pair of eyes

This is the part that separates goals that work from goals that frustrate you, so it’s worth thirty seconds to get clear on this section.

When Claude finishes a turn, your condition and the whole conversation so far get passed to a second, smaller model whose only job is to answer one question: has the condition been met, yes or no? A “no” sends Claude back to work, with the reason attached as guidance for the next attempt. A “yes” clears the goal and hands your session back.

So the work and the marking are done by two different models. This is the key.

The one writing the campaign is not the one deciding whether the campaign is finished, which is exactly why a goal catches things a detailed prompt doesn’t. It’s the same reason you don’t proofread your own copy.

Now the single most important thing you need to understand about this checker: it cannot use tools.

It can’t open your files, run a check, or visit a link. It reads the conversation and nothing else, which means it can only confirm what Claude has actually written into the transcript.

That has one very practical consequence…

If your condition is “every link in this email works,” the checker has no way to test those links itself. It’s looking for evidence in the conversation that Claude checked them and reported back.

A goal is satisfied by visible proof, not by the work having been done.

This is why every good goal asks Claude to report, list or show its findings. It isn’t thoroughness for its own sake, it’s structurally required.

The rule to remember: state the check, not just the outcome.

Weak: “The pricing is consistent everywhere.”

Strong: “The pricing is consistent everywhere, shown asset by asset in a table before you finish.”

Same intent. The second one gives the checker something to actually read.

Prompt versus goal, and why this fixes the “80% done” problem

There’s an obvious objection here, and it’s a fair one…to a point…

If I can list twelve things I want checked, why not just put all twelve in a really detailed prompt?

You can, and it does help.

But it doesn’t solve the problem, and the reason is the whole point of this piThere’s a part in the article which creates a table look which doesn’t work in my Substack, which is the column headings say ece.

A detailed prompt still gets one attempt. Claude reads your twelve checks, does a pass, produces a response. Miss the fourth, half-do the ninth, and nothing notices.

A detailed prompt tells Claude what to check. A goal makes Claude prove it did.

One marketer, Brandon Redlinger, framed the distinction better than I probably can: a prompt tells the AI what to do, a goal tells it when to stop. His blunter version is that AI bails early on prompts, but it can’t bail early on goals.

This is exactly where the “80% done” problem gets solved.

The single most common way AI work disappoints isn’t the model doing something stupid. It’s the model doing three-quarters of the job well and presenting the result with total confidence.

A goal removes the option.

It doesn’t hand you 80% dressed up as 100%, because an 80% answer fails the check and goes straight back for the missing 20%. It keeps refining until your definition of finished is actually met.

I want to be precise about “finished,” though, because it’s where people sometimes trip up. A goal drives to your checkable standard, not to some vague notion of perfect.

That distinction is the whole skill, and it’s the next section.

How to write a goal that actually works

A goal works when the checker can read the conversation and answer yes or no.

It fails when the condition is, well, just a ‘vibe’.

“The campaign is good” is not a goal. There’s no finish line, so you get either an arbitrary stopping point or Claude endlessly polishing things nobody asked it to polish.

“Every deliverable uses the same offer, price, dates and CTA link, listed asset by asset” is a goal. It’s checkable, and the proof lands where the checker can see it.

Four tests before you commit to one:

1. Is it checkable? Could a stranger read the output and tell you whether the condition holds, without needing your taste or your context?

2. Is it countable? “Three drafts,” “every page,” “ten examples,” “all twelve links.” Numbers and completeness turn a wish into a condition.

3. Is it bounded? Does it have an end state, or just a direction? “Make it better” is a direction. “No page is missing a meta description” is an end state.

4. Is the proof visible? Have you told Claude how to demonstrate it, and does that demonstration land in the conversation?

You’ve got plenty of room to do this properly…

A condition can run to 4,000 characters, so a goal can carry a real checklist, not a single sentence. Most people I’ve seen using this so far write them far too short.

And ALWAYS give it an escape hatch.

Add something like “or stop after 15 turns and tell me what’s still outstanding” to the end. Without it, a goal built on something Claude genuinely can’t satisfy will keep going, and going, quietly spending money while it does. With it, you get an honest report instead of a spiral. If you take one habit from this piece, take that one.

From finishing once to improving continuously: /loop

Everything so far fixes a single piece of work. You set a finish line, Claude works until it’s reached, and you get something genuinely done rather than nearly done.

But marketing isn’t a single piece of work.

It’s the same checks, over and over, as the inputs keep changing.

New posts publish, competitors reprice, and content that ranked in January quietly drops by March. A campaign that was good to go on Monday is out of sync by Thursday as three people edit four assets.

This is where the second command comes in, and where the two of them together become more than the sum of their parts.

/loop re-runs a prompt on an interval, inside your session.

Example:

/loop 30m check the campaign folder for any asset whose offer or CTA no longer matches the brief, and flag it runs that check now, then again every thirty minutes while you work.

You can give it a fixed interval, or leave it off and let Claude choose one based on what it’s seeing, longer gaps when nothing’s changing, shorter when things are moving.

Hold the goals and lopps side by side and the benefits become clearer.

/goal controls when work is allowed to stop.

/loop controls when it starts again. One drives a piece to done. The other keeps coming back to check it’s still done as things change around it.

That’s continuous improvement.

Inside a single run, a goal already improves in a loop of its own: produce, check, “not yet, here’s why,” improve, check again, until the standard holds. Across days and weeks, /loop is the outer loop, re-running the check on a schedule so the work stays at your standard even when you’re not looking at it.

Get something to your standard with a goal. Keep it at your standard with a loop. That’s the engine.

One thing to note.

/loop runs while your session is open, so it’s for repetition you want during a working stretch: polling a build, watching a folder, re-checking a campaign as it comes together.

For a check that should run every morning whether or not you’re at your desk, that’s Routines, the cloud-scheduled cousin I covered a few issues back. Same idea, different clock. /loop for while you’re here, Routines for while you’re not.

What Goals and Loops look like in practice

Four goals worth trying this week.

1. The campaign consistency check

/goal Every deliverable in this brief is drafted and cross-checked against the others: same offer, same price, same start and end dates, same CTA URL, same product name. Before finishing, show me a table with one row per asset and a column for each of those five, so I can see for myself that they match. Flag anything you changed. If something can’t be resolved without a decision from me, stop and ask rather than guessing. Stop after 15 turns either way.

This is the highest-value use of the feature, because cross-deliverable consistency is the one thing no single writer can check.

Each person makes their piece correctly. The mismatch is a property of the whole set, and it only surfaces if someone reads all six things side by side at the end, which nobody ever does. Run this and you’ll usually get back two or three genuine catches. Rarely the price.

Usually the archived link, the product name that changed mid-campaign, and the date that says Friday in one place and the 14th in another when the 14th is a Thursday.

2. The content batch with actual variety

/goal Three LinkedIn posts are drafted for Monday, Wednesday and Friday. Each uses a different hook type, no two open with the same sentence structure, and none reuses a story or statistic from my last month of posts. Before finishing, list the three hook types side by side and the opening line of each so I can see they’re genuinely different. Stop after 15 turns.

Batching content is efficient and it quietly kills variety.

Write three posts in one sitting and they sound like the same post three times, because you’re in one mood, drawing on one recent set of thoughts, stuck in one sentence rhythm. The variety clause is the whole point.

Without it you get three good posts your audience experiences as one repeated post, which is worse than three mediocre posts that at least feel different.

3. The pre-send email QA

/goal Every link in this email resolves, every UTM parameter follows our naming convention, the preview text is under 90 characters and doesn’t repeat the subject line, and every merge field has a fallback value. Print each link with its status, and each merge field with its fallback. If a link can’t be checked, say so explicitly rather than treating it as fine. Stop after 15 turns.

Email is the channel with no undo.

Every marketer has sent a broadcast with a broken link, and it happens because link-checking is boring, mechanical work done at the exact moment you’re most eager to hit send.

The merge-field one is the quiet killer. “Hi {{first_name}},” renders as “Hi ,” for every subscriber who came in through the form where you never collected a first name, which is usually the form that brought your best segment.

4. The research that doesn’t give up at six

/goal I have ten real competitor onboarding email subject lines, each listed with the company, the exact subject line, and a one-line note on the angle it uses. Ten, not eight. If you genuinely cannot find ten, list where you looked and what you found instead, rather than inventing any to make up the number. Stop after 15 turns either way.

Research is where AI stops early most often, because six good examples genuinely look like a complete answer.

You asked for examples, you got examples, the response reads as finished. The “ten, not eight” is deliberately blunt and it works. The final clause matters just as much, because it gives Claude an honest exit that isn’t quietly padding the list with things it made up to hit the number.

Then, once one of these earns its place, wrap it in a loop.

A weekly /loop that re-runs your competitor subject-line pull keeps your swipe file current without you remembering to do it.

A /loop over your published posts that re-checks internal links catches decay as it happens rather than at the next audit.

The goal makes the check trustworthy. The loop makes it continuous.

And this isn’t theory.

One marketer I know ran a goal across thirty blog posts, requiring every one to carry at least two internal links, with the count documented in a file as proof. It finished in about four minutes, unattended, and the checking model cost roughly two cents. A boring, important, easily-skipped job, done completely, for the price of not thinking about it.

Keep the finish line at “ready,” never “sent”

One rule I’d apply to every goal and every loop you write.

Drive work to ready. Not to sent.

It’s tempting to write “the campaign is published.” Don’t.

The moment a condition can only be satisfied by an irreversible action, you’ve handed away the part of the process where your judgement does the most work.

Keep the finish line at drafted, cross-checked, staged and waiting for you. Publishing, sending and scheduling stay manual, deliverable by deliverable.

This is the 10/80/10 principle I keep coming back to.

You do the first 10%, writing the goal and defining what done means. AI does the middle 80%, the production and the checking. You do the last 10%, reading it and deciding what ships.

Goals and loops don’t change that split.

They make the middle 80% far more reliable, and they stop it handing you 90%-complete work dressed as finished.

4 things worth knowing about Goals and Loops

1/ A goal does not change permissions, and this catches almost everyone.

Setting a goal removes the need to prompt Claude each turn. It does not remove the approval prompts on individual actions. Set a goal, wander off for a coffee, and you may come back to find it waiting on you to approve something on turn two. To genuinely leave it running, pair the goal with auto mode. Walking away is two settings, not one.

2/ Turns cost money.

The checking model is cheap to the point of irrelevance. The turns it triggers are not. A goal that runs fifteen times has done fifteen turns of real work, and a condition that can never quite be satisfied announces itself through your bill rather than through an error message. That’s the whole reason for the “stop after N turns” clause, and the same caution applies doubly to loops, which by design keep going.

3/ It’s the wrong tool for subjective work.

“This essay is compelling” has no finish line, and a second model is no better placed to judge that than the first one was. Creative judgement is precisely the part you shouldn’t hand over. Goals are at their best on the mechanical, checkable layer underneath the creative work, which is a bigger share of marketing than most of us like to admit.

4/ The checker only sees the conversation.

Worth repeating, because it’s the mistake that costs most. A confident line saying “all links checked and working” reads exactly like a real check to a model that can’t open a browser. Ask for the itemised proof, and spot-check it yourself the first few times until you trust the pattern.

The bigger picture

Every other layer in that table makes Claude capable of more, or gets it moving faster.

/goal is the only one that makes it stop being satisfied too easily, and /loop is the one that keeps it coming back.

Together they’re the closest thing marketing has to a self-correcting production line: get it right, then keep it right.
That’s the more interesting angle for me, because capability stopped being the constraint on AI marketing a while ago. The models can write the email, build the page, run the audit and pull the research. What they’ve been poor at is knowing when the work is genuinely finished, and when to come back and check it hasn’t drifted. Those two commands are the answer to both.

And once you’ve written a goal or a loop that earns its keep, it’s worth saving.

Drop it into a Skill, or a shared repository your whole team pulls from, and the check you wrote once becomes a check everyone runs. I made the case for that shared home in the guide to GitHub for marketers, and this is exactly the kind of asset it was built to hold.

The marketers getting the most from AI right now aren’t writing the cleverest prompts.

They’ve got specific about what finished means, written it down, and stopped accepting work that merely looks like it.

That’s a briefing skill, not a technical one.

It’s the same skill that separates a good creative director from a bad one, and it always has been.

This week’s challenge: take the next campaign or content batch on your list and write the goal before you write anything else. One condition, specific enough that a stranger could check it, with a “stop after 15 turns” on the end. Run it, and see what comes back on the list of things it found and fixed. Then, if it earns it, wrap it in a weekly loop. Reply and tell me what it caught. I’m genuinely curious whether it’s the price mismatch or the dead link.

Thanks for reading Marketing with AI. If you found this useful, sharing it with one other marketer is the best compliment I could receive.

Need 1:1 Help to Embed Claude into Your Marketing?

Let’s talk →

Marketing with AI Weekly Roundup: LinkedIn Turns AI Slop Into a Button, the AI Disclosure Deadline Lands on Sunday, and Agents Learn to Log In

Robert Gillespie — Sun, 02 Aug 2026 10:02:32 GMT

Welcome to issue #45 of the Marketing with AI Weekly Roundup. Published every Sunday, it shares the top stories, playbooks, data and quick hits from the world of Marketing with AI over the last 7 days.

Subscribe to receive the week's most important AI marketing stories direct to your inbox.

Subscribe now

Learn more about Claude Operator System

For two years the question about AI in marketing has been what you can make with it.

This week, three different forces converged on a different question: whether you have to tell people you used it.

On Thursday, LinkedIn gave every member a button to report a post as AI slop, killed its own AI writing feature, and started building a test that will tell you privately when your audience thinks your posts read like a machine wrote them.

On Tuesday, Google shipped AI content labels across its ad platforms.

And on Sunday, the day this lands in your inbox, the European Union’s AI transparency rules become legally enforceable.

Those are not three coincidences.

A regulator, a platform and an audience arrived at the same demand in the same week, which is worth paying attention to even if you sell nowhere near Europe.

Elsewhere, the access question moved.

ChatGPT’s agents can now log into the tools you use and stay logged in, which quietly removes the wall that stopped most useful marketing automation at the sign-in page.

Voice stopped being dictation and became a way to actually move work. A

nd three separate pieces of research landed on the same uncomfortable finding: almost everyone has AI in production, and almost nobody has the plumbing to support it.

Below: four stories to act on, two workflows worth stealing, and the clearest data yet on why your launch dates have not improved.

Let’s get into it…

🔥 Top Stories

1. LinkedIn hands every member a button to report AI slop, and kills its own AI writing tool

On 30 July, LinkedIn’s Chief Product Officer for Ecosystem, Hari Srinivasan, announced six changes aimed at what the company calls AI slop, which it defines as low-effort AI-generated content that sounds polished but carries no real perspective or substance. The headline change is a “Seems like AI slop” option in the three-dots menu on any feed post. Selecting it hides the post from you and feeds LinkedIn’s classification models. Alongside it, LinkedIn is deploying new classifiers that identify AI-slop and low-quality posts and reduce their reach in suggested content and in content shown outside your network. It says it now catches hundreds of thousands of automated comment attempts every day.

Two further changes matter more than the button. LinkedIn is testing a feature that will privately flag in your own analytics dashboard when members feel your post came across as inauthentic or heavily AI-written. And it is removing the “enhance your post” AI writing feature entirely, replacing it with a proofreader that fixes grammar and spelling without rewriting your words into a generic AI voice.

Why it matters / what to do: LinkedIn is the single most important organic channel for most B2B marketers, and it has just turned “does this read like AI wrote it” into both a ranking signal and a button in the hands of your readers. The author-side flag is the part with teeth, because for the first time you will be told, in your own numbers, that your audience does not believe you wrote your own posts. Action: stop using any tool that produces a finished LinkedIn post end to end, and move to a process where AI generates raw material and you supply the specific detail, the numbers and the first-hand experience a model cannot invent. The irony worth noticing is that LinkedIn has reached the same conclusion about its own product, which is why the writing feature is being deleted rather than improved. Two caveats: the button is not on every account yet, and 404 Media’s screenshot shows the option also appearing on sponsored posts, sitting between “Hide this ad” and “Report ad”, with LinkedIn so far saying nothing about how that signal is treated on paid placements.

2. Google ships AI content labels five days before the EU deadline that explains them

On 28 July, Google began rolling out AI content labels in Asset Studio, letting advertisers disclose when an image or video asset was generated or modified with AI. The setting spans Google Ads, Display & Video 360, Campaign Manager 360, Merchant Center and Google Ads Editor. You switch on a new AI label setting and can then apply text or visual labels to eligible creative manually, and Google says it may also apply labels automatically to certain assets made with its own AI tools. Related plumbing arrived the same week, with a synthetic content attestation field added to the Display & Video 360 API on 23 July and AI content labelling added to both the Campaign Manager 360 and Display & Video 360 APIs on 28 July.

The reason this shipped now sits in the calendar. Article 50 of the EU AI Act becomes legally enforceable on 2 August 2026, the day this newsletter reaches you. Two of its four transparency duties land directly on marketing teams. Anyone publishing AI-generated or AI-manipulated images, audio or video that realistically depicts real people, places or events must disclose that it was artificially generated. And AI-generated text published to inform the public on matters of public interest must be labelled, unless a named person or company took meaningful human editorial responsibility for it. Law firms spent last week telling advertising and PR teams specifically that this reaches product shots, AI-generated backgrounds, synthetic characters, press releases and ad copy. Disclosures must be clear, distinguishable, accessible and delivered no later than first exposure. Fines reach 15 million euro or 3% of worldwide annual turnover.

Why it matters / what to do: the obligation is legal but the switch is now sitting inside the ad platform you already use, which makes this unusually easy to act on. Action: this week, open Asset Studio, turn the AI label setting on, and audit which of your live image and video assets were generated or touched by AI. Then write down, in one line, who at your company takes editorial responsibility for AI-assisted published copy, because that named responsibility is the exemption that removes the text-labelling obligation entirely. Most B2B teams already qualify and simply have not documented it. Three details worth holding onto: content both generated and published before 2 August needs no retroactive labelling, text drafted in July but published in August is caught because the trigger is the publication date, and Google itself explicitly warns that using its AI label setting does not guarantee compliance with local regulations. The visual obligation has no editorial-responsibility exemption.

3. ChatGPT agents can now log into your tools and stay logged in

Announced on 25 July by OpenAI, a ChatGPT Work agent can now reach websites that require an account. The mechanism is a handover: you take over the agent’s cloud browser, sign in yourself, then hand control back and the agent continues the job. Screenshots are disabled during the takeover and OpenAI says it never sees your password. Cookies persist afterwards, so you sign in once per site rather than every session, and the login steps do not count against your message limits. It is available on Pro, Plus, Business, Enterprise and Edu plans.

Why it matters / what to do: this is less exciting than it sounds and more consequential than it looks. Almost everything a marketer would genuinely want an agent to do sits behind a login: Google Ads, GA4, LinkedIn Campaign Manager, your CMS, your email platform, your CRM. Until now the agent walked confidently up to the sign-in page and stopped, which is why so many agent demos quietly use toy examples. That wall is now gone. Action: pick one recurring, read-only job that lives behind a login, such as pulling last week’s LinkedIn Ads numbers or checking your CMS publishing queue, run it once with the takeover, and see whether it repeats unattended next week. Do not give an agent write access to a live ad account on the first run. Two caveats worth weighing: this is sourced to OpenAI’s developer account rather than a full product post or help-centre article, so treat the specifics as launch-note quality. And persistent cookies mean an authenticated session to your ad account is now sitting in OpenAI’s cloud browser, which is a real question to put to whoever owns security at your company before you connect anything that spends money.

4. Voice stopped being dictation, and the most useful version landed on the Mac

On 29 July, Google added Speak to Window to the Gemini app for macOS. Long-press the Fn key and speak into any window on your desktop, then release and the text lands at your cursor. Double-tap Fn for hands-free recording on longer passages. By default it is intelligent dictation, which strips the ums and ahs and catches your mid-sentence corrections rather than transcribing them.

The part that changes the job is a toggle you have to switch on yourself. Under the Speak to Window settings there is an opt-in for Gemini reasoning. With it on, Gemini reads what is on your screen, so you can highlight a paragraph and ask for a different tone, pull specific information out of a selected document, PDF or image, or generate an image straight into whatever you are working in. Pressing both Command keys sends the whole foreground window across as context. It needs an Apple Silicon Mac, macOS Sequoia 15.0 or later, and Gemini for macOS version 1.88 or later, and Google describes it as rolling out globally in English with more languages to follow.

This caps a fortnight in which every major assistant did the same thing. Claude’s voice mode moved off Haiku on 24 July and now runs on Opus, Sonnet or Haiku, and reaches connected apps mid-conversation, with Gmail, Google Calendar, Slack, Canva and Notion named. Two days before that, OpenAI brought voice into the ChatGPT desktop app, where it can coordinate several agents at once and be driven from your phone through a feature called Remote connections. In one practitioner walkthrough this week, the Gemini hotkey is used to highlight five unopened files in a folder and turn them into a finished email without opening a single one.

Why it matters / what to do: the honest marketer’s version of this is smaller than the demos suggest, because the demos are mostly personal admin. What is genuinely new is that briefing has become the interaction. Talking through what you want is faster than typing it, and briefing is the part of the job where your judgement actually lives, which makes it a better thing to speed up than drafting. Action: pick the next piece of work you would normally write a brief for, highlight the source material on screen, and talk the brief instead. Then compare how much correcting the output needs against your usual. Three caveats. Only the Gemini release falls inside this week, with the Claude and OpenAI changes landing on 24 and 23 July respectively and included here as context for the pattern. Speak to Window needs Apple Silicon, so an Intel Mac or a Windows machine is out. And the practitioner who demonstrated it is candid that with reasoning switched on it often tries to be an assistant when you only wanted it to type, which is a real irritation rather than a small one.

📋 The Playbook

1. What a real marketing agent actually is, and the ads loop one operator runs on it

This is the most useful thing published this week on agents, mostly because it starts by ruling things out. Cody Schneider runs companiesgraph.com and states on the record that this is what his company does, so treat it as an operator with a commercial interest rather than a neutral tutorial. The framework survives that.

His three tests. Something is only a marketing agent if it has all three, and if it does not, it is a linear automation wearing a costume:

1. Unified data across the whole pipeline, not a single tool’s export.

2. Autonomous decisions on a cadence, with a thinking loop that reads results back in.

3. Hosting in the cloud, not a script on your laptop.

His own words on why the third one matters, and why most “agents” fail the second: “I want something that is doing a process for me and then looking at the data and then basically improving based off of the data that’s getting back from it.”

The loop he runs for a live client, step by step:

1. Research pain points in the buyer’s own words. He uses Perplexity to pull from Reddit, then asks it to rank the results by most referenced to get the top three pains. Reddit specifically, because it is people complaining without a filter.

2. Generate the statics. Feed an existing competitor ad in as a structural example, generate the new versions, then put a vision model over the output to check it against your brand style guide, fonts, colours and text legibility, before anything ships.

3. Generate the video. He still gets results from AI-avatar video for user-generated-content style ads, and is candid about the constraint: current clips cap out under about nine seconds, so stitching them into a 30-second ad is the actual hard part.

4. Publish through the platform API, for writes only. His correction of the common complaint about account bans is blunt: “The agent is not the reason it got banned. They violated TOS and they spammed the API.” Use the API to publish, pause and promote. Do not use it to pull hundreds of millions of rows.

5. Build the data layer. Open-source and self-hostable throughout: a pipeline tool feeding a fast warehouse, with ad platform data, analytics, product analytics, your CRM and your payment processor all landing in one place so the agent can trace an individual ad through to actual revenue.

6. Run the loop. For one client: two ad sets per day, five ads per ad set, uploaded automatically. Each gets a two to three day window for initial signal. Kill the worst, and move winners into a winners pool that competes for budget. Store the actual prompts and ad scripts in a database so the system learns which creative recipes produce outcomes rather than just which ads won.

7. Solve for entropy. This is the part almost nobody discusses. The agent gets stuck thinking the same way and creative output converges. His two fixes: pull competitor ads from the platform’s public ads library to inject new ideas, and mine YouTube and podcast transcripts in your category for angles to build ads from.

The claim worth arguing with: he says Facebook has become the best B2B ads channel that exists right now, because Meta’s ranking system reads your creative and your landing page to decide who sees the ad, so you no longer build interest-based targeting lists. You write creative that names a specific pain and the platform finds the people who have it.

Do this now: most readers will not build the warehouse, and that is fine. Three things here work without any of it. Apply his three tests to the next “AI agent” a vendor pitches you. Run the winners pool manually, because a two to three day read and a budget-competing shortlist needs no infrastructure at all. And take the entropy problem seriously, because it is the reason your AI-generated creative starts strong and gets samey by week three.

Watch more:

2. A 14-step system for updating decayed pages without destroying what already ranks

Published on Search Engine Land on 29 July by Alex Galinos, this is the rare AI-and-SEO piece that is a process rather than a tool pitch, and it targets the highest-return content work most teams skip entirely.

The problem it solves is familiar to anyone who has run a content programme for more than two years. Existing pages quietly lose rankings and revenue over time. Updating them recovers that performance far more cheaply than publishing something new. But rewriting too aggressively wipes out the equity the page had already accumulated, so the update makes things worse and nobody can explain why.

The 14 steps cover diagnosing which pages have actually decayed rather than which ones feel old, making targeted changes that preserve the existing signals, measuring whether the update worked, and then turning the whole sequence into a repeatable system driven by Claude Code rather than a one-off project. The worked example runs on a marketplace with pages spanning airports, resorts and routes across multiple countries and languages, which is a harder case than most B2B sites and makes the method easier to trust on an easier one.

Do this now: before touching anything, pull the list of pages that have lost the most traffic over the last twelve months and sort by the revenue or pipeline they used to produce rather than by traffic volume. That list is almost always shorter and more valuable than the content calendar you were about to write. This one requires Claude Code, so it sits a step above the usual entry point, but the diagnosis half works manually.

🛠️ New Tools & Features

1. Google Search Console Platform properties, now available everywhere

On 29 July Google rolled Platform properties out to every Search Console account worldwide and published a guide to reading the data. Platform properties let you track how your posts on Instagram, TikTok, X and YouTube perform in Google Search, Discover and News, with clicks, impressions, average click-through rate and average position for each connected account. It works even if you do not own a website. For B2B teams this fills a genuine measurement gap, because it shows which social posts are earning visibility in Google rather than only on the platform where you published them. It covers Google surfaces only, so it complements your native analytics rather than replacing them. Free.

2. StackAdapt Ivy Studio

Announced 28 July. A conversational workspace inside StackAdapt’s advertising platform that combines planning, forecasting, analysis, optimisation and execution. You describe the outcome you want in plain language, and agents grounded in StackAdapt’s own campaign, audience, creative and performance data surface context, identify opportunities and recommend next steps, with the marketer keeping approval. Worth noticing as a pattern rather than a product: this is the third “describe the outcome, agents propose, human approves” advertising product in six weeks. Pricing is not published.

3. Nielsen Ad Intel AI, and you can query it from your own tools

Launched 27 July. Nielsen turned Ad Intel, its competitive advertising database covering 5.5 million brands and 4.6 million advertisers across 23 media types in more than 90 markets, from a reporting tool into something you ask questions of. It surfaces competitor creative strategies, emerging trends and spend shifts. The detail worth noting is that it can be exposed through MCP, the Model Context Protocol, which is the open standard that lets an AI assistant connect directly to an external data source, meaning your own assistant can query it inside your workflow. Enterprise-priced with no public pricing, so for most readers this matters less as a tool to buy and more as evidence that MCP access is quietly becoming a martech buying criterion.

4. OpenAI ships two new transcription models

Announced 29 July: gpt-transcribe for files and gpt-live-transcribe for live streams. Transcription is the unglamorous foundation under most content repurposing, turning a webinar into a blog post, a sales call into a messaging document, a podcast into a fortnight of social. Better and cheaper transcription lowers the floor on every one of those workflows. These are developer-facing models rather than a product with an interface, so the practical benefit arrives when the tools you already use adopt them.

Watch more:

📊 Research and Data

1. Most marketers cannot tell real AI capability from a sales pitch

StackAdapt and B2B agency Ledger Bennett published research on 27 July, with fieldwork by B2B research specialist NewtonX, surveying 426 marketers across the US, UK and Asia-Pacific. 77% say AI-related questions are not asked rigorously enough when evaluating a vendor. Only 23% assess AI against defined evaluation criteria. More than 60% say their RFP process cannot separate meaningful AI capability from marketing hype. The same survey found 63% cannot effectively measure cross-channel performance despite having clear targets, 76% are managing six or more platforms, and 59% still combine reporting data manually.

What to do: every martech vendor you speak to this year will claim AI capability, and by marketers’ own admission most buyers currently have no way to test the claim. That means a lot of budget is being committed on impression rather than evidence. The fix is cheap. Write a five-question AI block into your next RFP and reuse it: which model or models does this use, what happens to our data, what does the human approval step actually look like, what does it cost per action rather than per seat, and can you show us the output on our own data before we sign. Doing that once puts you ahead of roughly three-quarters of your peers. One caveat to hold: StackAdapt is itself an AI advertising platform, so this is a vendor publishing research about how difficult it is to evaluate vendors.

2. AI made the drafting faster, and your launch dates did not move

Knak published “Marketing Production in the Age of AI” on 28 July, surveying more than 300 enterprise marketing leaders. 85% missed at least one planned campaign launch date in the past twelve months because of workflow constraints, and 1 in 10 miss more than five a year. 70% have deployed AI in production, but 88% say the output still needs moderate to substantial human editing. The revealing split is where AI is actually being used: 64% for first-draft copy and 56% for image generation, but only 25% for building and coding the emails and landing pages, which is where the hours genuinely go. On process, 60% of campaigns involve four or more people, 54% juggle three to five separate tools, and 69% need two to three rounds of revisions.

What to do: this is a budget-direction finding, and it points away from where most teams are spending. The instinctive answer to “we are too slow” is to buy another AI writing tool, and this data says drafting was never the constraint, so that purchase will not move a single launch date. Handoffs, reviews and rebuilds are the constraint. Action: time one campaign end to end this month and record where the hours actually land, then redirect the next automation effort at whatever that measurement tells you rather than at whatever is easiest to automate. It also gives you a straight answer for the stakeholder asking why AI has not made the team faster yet. Caveat: Knak sells a marketing production platform, so the finding that production is the bottleneck is directly convenient for them, and the sample is enterprise.

💡 Quick Hits

1. Google Ads is switching to passkeys on 5 August.

From that date, new sign-ins that generate OAuth refresh tokens for the Google Ads API, Google Ads Editor and Google Ads scripts will require passkey authentication.

2. Google indexed people’s shared Claude conversations.

Shared Claude chats began appearing in Google Search on 28 July. The cause was a robots.txt block sitting in front of a noindex header, so Google never crawled the pages to see the noindex and indexed them from links instead. A painful reminder that disallow and noindex are not the same instruction.

3. Gemini Spark can now drive Chrome on your own machine, and most of you cannot use it.

Announced 30 July, Google’s background agent now controls desktop Chrome using your logged-in accounts and saved passwords rather than a remote browser, handing sensitive actions such as payments back to you. Google AI Pro allows up to 20 multi-step Chrome tasks a day, Ultra up to 200. The catch is geography, and it is worth stating plainly: Google’s own documentation says Spark is unavailable in the European Economic Area, the United Kingdom, Switzerland and Nigeria regardless of plan, and the Chrome integration is United States only at launch.

4. Anthropic shipped the biggest MCP update since launch.

The 28 July specification makes the protocol stateless so servers can run on serverless and edge infrastructure, standardises interactive interfaces rendered inside a Claude conversation, adds long-running tasks that do not block the chat, and hardens authorisation for enterprise identity providers. Claude now lists more than 950 connectors.

5. ChatGPT Atlas stops working on 9 August.

OpenAI is deprecating its standalone agentic browser and folding the capabilities into ChatGPT and Codex. Hard cutoff, no extension announced.

6. Independent testing of Claude Opus 5 has started arriving, and it is mixed.

Artificial Analysis ranks it first of 170 models on its Intelligence Index, and ARC Prize independently published 30.16% on ARC-AGI-3. But a separate evaluation that feeds requirements in sequential checkpoints rather than handing over the whole specification up front returned a 24% strict pass rate with quality warnings on 93% of the output, which is a notable gap from the launch benchmarks Anthropic published itself.

7. You have until Tuesday to try Google’s video model for nothing.

Google announced on its Gemini account on 29 July that you can create ten videos at no cost until 11:59pm Pacific on 4 August, through Create video in the tools menu. Two details decide whether this applies to you: the offer is for users without a Google AI subscription, so paying subscribers are not the target, and while uploading your own footage to edit is blocked in the EEA, Switzerland, the UK and some US states, generating from a text prompt is not. Ten in total rather than ten a day, and 18 or over.

8. OpenAI cut the price of its cheap models by up to 80%.

From 30 July, GPT-5.6 Luna dropped 80% to $0.20 in and $1.20 out per million tokens, and GPT-5.6 Terra dropped 20% to $2 and $12. Sol is unchanged but gains a Fast Mode running up to 2.5 times quicker at double the price. Worth being precise, because the coverage has not been: this is an API price cut. Inside Codex and ChatGPT Work nothing gets cheaper, the usage is simply counted more favourably, so the same subscription stretches further. The practical effect is that high-volume, low-judgement jobs like classifying inbound enquiries or tagging survey verbatims became roughly five times cheaper to run at scale.

9. ChatGPT Ads grew up, and most B2B advertisers still cannot use it.

Announced 24 July and live from 27 July: a Conversions objective that creates conversion-optimised cost-per-click campaigns, geographic exclusions, automatic advanced matching using hashed customer data, AppsFlyer and Adjust integrations, bulk operations in the API, and daily budgets becoming an average across a rolling seven-day period. The catch sits in OpenAI’s own ad policies, which limit the test period to consumer verticals such as lifestyle and household goods, local services, travel and digital products, allow financial services, healthcare and legal case by case, and state that all other categories are disallowed at launch. B2B software and professional services are not on the list.

👀 On Our Radar

OpenAI appears to be building ads whose destination is a chatbot rather than a landing page.

Search Engine Land reported on 31 July, from observations inside ChatGPT Ads Manager, on a format under test that replaces the landing page entirely with a business-specific AI agent. The system automatically profiles your business by crawling your website, you configure the agent with custom instructions and data sources, and the campaign sends people into a conversation with that agent rather than to a web destination. The agent can answer questions, make recommendations, schedule appointments and qualify leads before anyone reaches your site. Search Engine Roundtable separately noted the same day that ChatGPT Ads Manager is testing “Agent” campaign types, which is consistent.

Treat this as unconfirmed. OpenAI has not announced it, it appears limited to selected advertisers, and nobody has observed the end-user experience, so there is no way to judge how prominent these would be. It is worth watching because of what it would change rather than what it is today. Every conversion habit most marketers have was built around a page you control, optimise and measure. If the destination becomes a conversation, the asset you are optimising stops being a layout and starts being a set of instructions and a data source, which is a genuinely different job and one almost nobody has practised.

Tell me: What’s your biggest takeaway from this weeks news from the world of AI Marketing? Love to hear your thougths in the comments.

Leave a comment

Thanks for reading the Marketing with AI Weekly Roundup! If you enjoyed this read, the best compliment I could receive would be if you shared it.

Marketing with AI Weekly Roundup: Claude Learns Your Job by Watching You Do It, HubSpot Hands Every Marketer an Agent Builder, and Google's AI Ad Machine Leaves Beta

Robert Gillespie — Sun, 26 Jul 2026 10:00:41 GMT

Welcome to issue #44 of the Marketing with AI Weekly Roundup. Published every Sunday, it shares the top stories, playbooks, data and quick hits from the world of Marketing with AI over the last 7 days.

Subscribe to receive the week's most important AI marketing stories direct to your inbox.

Subscribe now

Course Update:

The Complete Guide to Claude Design Now Included Inside Claude for Marketers Masterclass
(Pro/Elite Editions Only)

See what's inside the masterclass

For two years, automating a marketing task has meant writing down how you do it. That is the part almost everyone gets stuck on, because the jobs most worth automating are usually the ones hardest to describe: pulling numbers out of a portal with no API, formatting the same client report every month, logging into six dashboards in a particular order for reasons you have never written down.

On Tuesday, Anthropic removed that step.

You now record your screen while you do the job, talk through what you are doing and why, and Claude turns the recording into a saved, reusable automation. Every major practitioner channel published a walkthrough within 72 hours, which tells you how significant they think it is.

It was a big week elsewhere too.

HubSpot opened no-code agent building to every Professional and Enterprise customer, which puts “build an AI agent” within reach of a marketing ops person rather than a developer.

Google confirmed its AI-driven Search campaign type is out of beta with more than 500,000 advertisers on it.

And Anthropic shipped Claude Opus 5, which you will find in Quick Hits rather than up top, because the price and the capability jump matter more than the version number.

Running underneath all of it is a gap worth naming. The tools took a real step forward this week. New research says the plumbing did not: 71% of marketing organisations rate their ability to use their own first-party customer data as ineffective or underdeveloped.

Below: three stories to act on, three workflows worth stealing, and the clearest data yet on how winnable AI search still is.

Let’s get into it…

🔥 Top Stories

1. Claude can now learn a task by watching you do it once

On 21 July, Anthropic added Record a skill to Claude Cowork. You open the Claude desktop app, switch the Cowork toggle on, click the + menu and hit record. Then you do a task the way you always do it, narrating what you are doing and why as you go. Claude captures your screen, your clicks, your typing and your voice, and converts the whole demonstration into a skill, which is simply a saved instruction file it loads and runs on command. It is available on Pro, Max and Team plans, on the desktop app only. Until now, teaching Claude a repeatable process meant writing that process out. Now you can show it instead, and you can schedule the result to run on its own.

To give a sense of scale: in one of this week’s walkthroughs, a 200-second recording of a monthly client SEO report became a skill that pulls the data, writes the analysis, builds a branded PDF and drafts the client email, running automatically on the 22nd of every month.

Why it matters / what to do: The bottleneck in marketing automation has never really been the AI, it has been the handover. Writing a good process document is a genuine skill and most people never do it, which is why so many recurring tasks stay manual. Doing your own job while talking about it is something everyone can already do. That change in the input is what makes this the most consequential release of the week for ordinary marketers, more so than any model launch. Action: pick the most repetitive, least judgement-heavy task in your week, record yourself doing it once, then test the resulting skill on a fresh example before you trust it. Two practical warnings. Claude’s own in-product notice tells you not to type passwords or display sensitive information or private conversations while recording, so close anything confidential first. And there is no Anthropic help page documenting recording retention or storage yet, so treat anything client-sensitive with caution until there is.

2. HubSpot opens no-code agent building to every Pro and Enterprise customer

On 23 July, HubSpot launched Agent Hub and Agent Builder, both in public beta and both included for Professional and Enterprise editions. Agent Hub is a single console showing every AI agent in your account with live status and recent results, organised by outcome rather than by product: build demand, win deals, delight customers, scale growth. Agent Builder is the more significant half. It is a no-code canvas where you describe an agent in plain English and assemble it from the customer data already sitting in your CRM, with multi-source triggers, workflow orchestration and approval guardrails built in. Custom agents consume HubSpot Credits when they take actions, and credits are included at those tiers with the option to buy more.

Why it matters / what to do: Most AI agent coverage this year has assumed you are willing to work in Claude Code, n8n or a developer tool. This is the first time the CRM that a large share of B2B marketers already pay for has let them build their own agent rather than switch on the vendor’s pre-built one, grounded in their own account data. That is a meaningful shift in who gets to do this work. Action: if you are on Marketing Hub Professional or Enterprise, open Agent Builder and build one deliberately narrow agent for a recurring, low-judgement job, such as enriching lead-routing notes or assembling meeting-prep briefs. Watch the credit consumption on a small volume before you scale it, because that is the real cost and it is not yet clearly priced. Bear in mind this is a public beta, and HubSpot’s headline customer proof point is its own case study.

3. Google AI Max leaves beta with 500,000 advertisers on board

On Alphabet’s Q2 2026 earnings call, reported on 23 July, Google confirmed that AI Max is out of beta and now used by more than 500,000 advertisers. AI Max is Google’s AI-driven Search campaign type: instead of matching against your keyword list, it interprets longer, conversational queries and matches them against your landing pages, creative and assets. Google’s Chief Business Officer said it is opening “billions” of previously unmonetised searches to advertising. The company also said advertisers using AI Max or Performance Max see an average 15% lift in conversions or conversion value at similar return on ad spend, that Gemini has improved ad relevance on complex queries by around 20%, and that it is testing Highlighted Answers, a format that places labelled sponsored links inside AI Mode’s generated list responses.

Why it matters / what to do: Long, specific, high-intent B2B queries were always the hardest thing to buy against with keywords. Nobody was bidding on “best compliance platform for a 200-person fintech” because you cannot build a keyword list that anticipates it. That inventory is now being opened up, and the matching burden has moved off your keyword list and onto your landing page copy, your assets and your feed. Action: check whether AI Max is switched on in your Search campaigns and, more importantly, read your search terms report for query types you would never have bid on manually. If the matches look wrong, that is usually a landing page problem now rather than a keyword problem. One caveat to hold onto: every performance figure here is Google’s own, delivered to investors on an earnings call, and the 15% figure blends AI Max with Performance Max. Treat it as directional.

📋 The Playbook

1. Turn your monthly client report into a recorded, scheduled skill (Eliot Prince)

This is the clearest end-to-end demonstration of Record a skill, and it automates the kind of job that eats a week every month.

The setup. Open the Claude desktop app, which is required because Claude may need to drive your computer. Switch the Cowork toggle on, then click the + button. Record a skill only appears when Cowork is on. The dialogue tells you your screen, clicks, typing and voice will be recorded and turned into a repeatable skill. Hit Start recording.

The recording, step by step. The job is a monthly SEO client report, and the whole thing takes 200 seconds:

1. Open the browser and go to Google Search Console. Use the left dropdown to select the client property.

2. Go to Performance, then More, then Compare, then last 28 days year over year. He narrates why: comparing year over year removes seasonality from the picture.

3. Export to an Excel file, saved into a new desktop folder named for the client and the month.

4. Say the analysis rubric out loud as you go: five executive points covering the good, the bad and the unchanged, each with data behind it, plus three recommended actions for the month. He also states the tone rule, which is to stay positive and frame problems as opportunities.

5. Open a document, add a title, put the five bullets at the top with commentary below, and export as PDF.

6. Open Gmail, compose, write a two-line overview, attach the report, and set the subject line.

After you hit done. Claude asks clarifying questions before it builds anything, so answer them deliberately: use browser automation rather than waiting for a manual export, produce a branded PDF, draft the email but do not send it. Then edit before saving. His own correction is the one worth copying: the first build used generic branding, so he told Claude to use his existing brand skill rather than inventing a look, which points the new skill at work he had already done.

Running and scheduling it. Invoke it with a slash command or in plain English. Claude then drives Chrome hands-free, applies the filters, exports, analyses, builds the PDF and drafts the email. To automate it entirely, ask Claude to turn it into a scheduled task that runs on a set date each month, and it appears under Scheduled tasks in the sidebar.

The honest limitation, which he shows rather than hides: Claude could not attach the finished PDF to the Gmail draft through the connector. Everything up to that point worked. Expect to finish the last step yourself for now.

Do this now: pick one boring, repetitive task, show Claude click by click, and narrate why you are doing each step. The narration is what becomes the skill.

Watch here:

2. When to record a skill, and when not to (Brock Mesarich)

This was the most-watched video of the week on any of the channels we track, and it is more useful than the enthusiastic walkthroughs because it tells you when the new feature is the wrong tool.

The decision framework, which is the reusable part:

• If the task has a native connector, use the connector. Connecting directly is more reliable than narrating your way through an app’s interface.

• If there is no connector but you can do the job in a browser, record it. This is where Record a skill genuinely earns its place, on the manual work you do on your own computer that no integration covers.

• If there is no connector and the browser work is fragile, use the Zapier MCP connector instead, which reaches around 9,000 apps. His examples of tools with no native Claude connector include Skool and beehiiv.

The demonstration. He deliberately picked a task with no available connector: copying comments out of YouTube Studio into a tracking spreadsheet and having Claude judge which ones are worth turning into a video. While clicking, he narrates the rules, including filtering to long-form videos only, pasting into a specific column and wrapping the text. Claude processed 73 steps and roughly 76 screenshots plus a full transcription of his narration, and built the skill in about two minutes. The resulting file specified the browser tool to use, the navigation steps, a visual of the spreadsheet columns, a yes or no rubric for video-worthiness, a report-back format and edge cases. It also inferred a rule he never said out loud, which was never to reply to, like or delete the comments.

His caution on privacy, quoted: if you are worried about privacy, maybe do not use this, and at minimum make sure not to show banking information or anything confidential.

Do this now: before you record anything, ask the one question that saves the most time. Can Claude connect to this application directly? If yes, build the skill by talking to Claude instead. Recording is for the work that has no other route.

Disclosure: he links the Zapier MCP connector in his description as an affiliate and recommends it over the free Anthropic feature in some cases, so weigh that recommendation accordingly. The decision framework itself holds up.

3. Three files that make Claude Design produce on-brand decks instead of generic ones (Jeff Su)

Most Claude Design tutorials tell you to pick a template and start prompting, which is how you end up with something generic and then waste more time fixing unusable slides. This is the version that front-loads the setup so the output is on-brand from the first deck.

Input 1: a design file. This is a plain text rulebook covering colours, fonts, spacing and component styling, and you do not write it yourself. There is a public GitHub repository of design.md files from major brands. He downloads one, uploads it to any chatbot with the instruction to remove the proprietary content and replace it using the model’s own judgment while keeping everything else exactly as is, then renames the result. The reason for the swap is practical: Claude Design will not copy a real brand’s guidelines, so you need a file that is detailed enough to be useful and safe to use.

Input 2: a design system. Start a new chat in Claude Design, select the Opus model, set effort to max, choose Create design system, name it, and upload the design file at the bottom. It takes 10 to 20 minutes and produces mockups and sample components to review. Refine using the feedback button in plain English. Then upload the extras that make it yours: a voice principles file so the copy sounds like you, your logo, and the icons you use often. These become permanent and you never upload them again.

Input 3: a template. Ask it to build a slide deck template for your use case. The design system controls how things look, the template controls how a deck is laid out. His feedback round is worth copying almost verbatim: apply the gradient to every white slide, remove the colour backgrounds he did not want, scale text up to fill empty space, and add the logo to the footer.

The compounding trick, and the most valuable part of the whole workflow: when you give feedback, add an instruction telling Claude to make the edits and also write a claude.md file it will read for all future slide decks. It then saves every correction to a file at the project root and reads it before every new design. His related tip: start a new chat inside the same project rather than a new project, so your accumulated learnings carry forward.

Three editing modes worth knowing: Edit lets you click an element and change it yourself. Annotate lets you literally draw a circle on a slide and type an instruction about that area. Tweaks creates toggles for decisions that affect every slide, such as showing your logo or slide numbers, so you can flip them across the whole deck at once.

Do this now: the same three-input setup produces social carousels and on-brand email visuals, not just decks. Build the design system once, then reuse it. Export options include PowerPoint and PDF, though he prefers standalone HTML because it opens in any browser and can be sent to anyone.

1. Webflow MCP 2.0 puts guardrails on agents editing your website

Webflow, the visual website builder used by a lot of B2B marketing teams, shipped version 2.0 of its MCP server on 21 July. MCP, the Model Context Protocol, is the standard that lets an AI assistant safely operate an outside tool. The 2.0 release is entirely about control rather than capability: reusable Agent Instructions that encode your voice and tone guidelines, legal requirements and brand rules; enforcement of your existing account roles and permissions with granular access per site, CMS collection, page and locale; and every action logged in the site activity log with AI attribution, so you can see what an agent changed and when. This is the first credible answer to the objection that stops most teams letting AI near their website, which is that you want the speed without the risk of something going live that legal has not seen. Webflow also disclosed a number worth noting: more than 30% of its enterprise customers actively use its MCP, usage is up more than fourfold since January, and nearly 90% of those MCP users connect through Claude.

2. Otter.ai Live Assist coaches you during the call, not after it

Otter.ai, the AI meeting transcription tool, launched Live Assist on 22 July, described as a real-time coaching agent that joins any call and surfaces glanceable tip cards visible only to you while the conversation is happening, ticking off objectives as you cover them. You build it in Otter’s Agent Builder and ground it in your own material: playbooks, standard operating procedures, past meetings and key resources. It works across Zoom, Google Meet and Microsoft, and connects to Claude and Salesforce. The obvious use is sales coaching, but the marketing use is sharper. Ground it in your positioning and messaging document and it becomes live messaging enforcement on customer calls, which is exactly the point where carefully built positioning usually falls apart. It is Otter Enterprise only, which rules it out for smaller teams for now.

3. Ask Claude how your industry actually uses AI

On 22 July, Anthropic made its Economic Index queryable directly inside Claude through a connector. The Economic Index is a public dataset measuring how AI is genuinely being used across the economy, and you can now ask which occupations use it most, what tasks people are automating, or how a specific role or region uses Claude, and get an answer grounded in the data with sources attached. Enable it from the connectors directory in Claude, with nothing to install. For marketers this is a free instrument for answering the “are we behind?” question with something better than a vendor survey, and it is a decent source for content backed by real usage data rather than a press release. The obvious limitation: it is Anthropic’s own usage data, so it skews towards how Claude users work.

4. Gemini Notebook adds Collections

Google’s Gemini Notebook, renamed from NotebookLM last week, is rolling out Collections, which let you group related notebooks the way you would build a playlist rather than filing them into rigid folders, with a single notebook able to sit in several Collections at once. It was reported on 21 July and is rolling out gradually. If you have built separate notebooks per client, product or campaign, this is the first real organisation layer for them. One limitation to know before you plan around it: Collections cannot be shared with other people, so they organise your own workspace only.

📊 Research and Data

1. More than half of all categories have no brand owner in ChatGPT yet

Semrush, working with independent growth adviser Kevin Indig, analysed 1,094 US categories in ChatGPT every month from January to June 2026, covering more than 50,000 brands, 220,000 domains and 600,000 citations. Rather than testing single prompts, they grouped five prompts per category to mirror how a buyer actually researches: a definition question, a comparison, an alternatives question, a use case and a buying question. A brand counted as the category owner only if it appeared in at least four of those five prompts and led the runner-up by at least five percentage points.

The findings are unusually encouraging. Only 15.2% of categories have a clear owner. Another 31.2% have an emerging leader without a decisive lead, and 53.7% are completely unsettled, with no brand appearing reliably across the category. Ownership is scarcest exactly where the demand is: only 11.3% of the most-asked categories have an owner, compared with 19% of the least-asked, which means 89.3% of estimated AI-search demand sits in categories nobody has locked down. Two further findings matter. Being cited is not the same as being recommended, and the two rarely coincide: only 21% of the most-cited domains in a category were also the most-mentioned brand. And traditional SEO strength barely predicts ownership, with branded search volume the only statistically significant metric and even that separating the owner from the runner-up just 55.7% of the time. Where a brand does establish ownership it tends to hold, leading in 90.4% of month-on-month comparisons, though narrow leads are fragile.

A separate study published the same week by Writesonic, based on 631,999 prompts, points the same way: the brand holding the top position in ChatGPT changes on 52% of reruns of the same prompt, with only 8% of positions genuinely locked.

What to do: run your own five prompts this week, one for each stage of that buyer journey, and count how many name you in the answer text rather than in the citations underneath. That distinction is the point, because the answer is what your buyer reads. If nobody consistently owns your category, that is an open lane and your existing search rankings will not win it for you. If you are leading narrowly, you are one competitor’s content push away from losing it. Worth flagging on the source: Semrush sells the AI visibility toolkit that produced this data, so it has a commercial interest in you caring about the result. The methodology is unusually well documented, and the finding is checkable in your own category in about twenty minutes.

2. Only one in four chief marketers rate themselves as advanced at adopting martech

The CMO Council released initial findings from its 2026 Marketing Transformation Performance Audit on 22 July, a self-assessment completed by 200 members. Only 1 in 4 rate themselves as highly advanced, adaptable and agile at adopting emerging marketing technology, with nearly half describing their stack as functioning but capable of better. 34% admit to fragmented environments and integration problems, and 37% say they are lagging at identifying, specifying and deploying new solutions. The number that should stop you: 71% rate their organisation’s ability to use first-party customer data as ineffective or underdeveloped, and 80% say the same of sourcing and integrating third-party data. Fewer than a quarter believe they have excellent data quality and a real-time data foundation. On people, the picture is brighter, with nearly 1 in 4 reporting deep, capable teams and 43% describing them as good with gaps.

What to do: read this against the rest of this week’s issue. The tools took a visible step forward, and the constraint they run into is not model capability, it is whether the data those tools need is actually reachable. Before you buy or build anything else, run the cheap version of this audit yourself: pick the AI tool or agent you are most excited about, list the data it needs to be useful, and check whether that data is accessible, current and clean. If the answer is no, that is the project, not the tool. Three caveats on the research: it is self-assessment rather than measured performance, the sample is small at 200, and it covers CMO Council members, who by definition self-select as people already engaged with this problem.

💡 Quick Hits

1. Anthropic released Claude Opus 5 on 24 July at the same price as its predecessor, $5 per million input tokens and $25 per million output. It is now the default model on Claude Max and the strongest model available on Claude Pro, with a 1 million-token context window, a new xhigh effort setting between high and max, and an optional Fast mode that runs about 2.5 times faster for twice the price. Read more: https://www.anthropic.com/news/claude-opus-5

2. Claude Fable 5 became a permanent part of Max and Team Premium plans on 20 July, usable for up to 50% of weekly limits. Pro and Team Standard subscribers do not get permanent inclusion and instead received a one-time $100 usage credit, after which Fable 5 runs on pay-as-you-go credits at API rates. Read more: https://support.claude.com/en/articles/15424964-claude-fable-5-on-your-plan

3. Google dropped the $50,000 lifetime spend requirement on Lead Form assets on 20 July, the in-ad forms that let someone submit details without leaving the results page. It also added Zapier as an official lead-delivery method alongside webhooks and the Ads API, plus eligibility in more than 25 new countries. Note that this was a documentation change spotted by Search Engine Land rather than an announcement from Google. Read more: https://searchengineland.com/google-drops-50k-ad-spend-requirement-for-lead-form-assets-482759

4. Text ads now appear on 29.45% of commercial queries in Google’s AI Mode, according to an SE Ranking analysis of 50,032 keywords, rising to 53.56% on keywords with a cost per click above $10. In 88% of cases where an ad ran, the advertiser’s own domain was not among the sources the AI answer cited. Read more: https://www.searchenginejournal.com/google-ai-mode-shows-ads-on-1-in-3-commercial-keywords/582976/

5. OpenAI launched Presence on 22 July, an enterprise platform for deploying voice and chat agents that answer questions, verify callers, take approved actions and escalate to a human when needed, wrapped in your own policies, guardrails and pre-launch simulations. OpenAI runs it on its own phone support line, where it says it resolves 75% of inbound calls without human involvement. It is not self-serve, with deployments led by OpenAI engineers or selected integrators and no published pricing. Read more: https://openai.com/index/introducing-openai-presence/

6. Anthropic launched four role-based Claude certifications on 23 July, covering Associate and Developer foundations plus Architect foundations and professional levels, proctored through Pearson VUE with a digital badge. Registration requires a company email tied to a Claude Partner Network organisation, so personal addresses are not accepted. Read more: https://claude.com/blog/four-role-based-claude-certifications

👀 On Our Radar

Your carefully written AI instructions may now be working against you.

On 24 July, Anthropic published new guidance on context engineering for its latest models, and the headline finding is striking: it removed more than 80% of Claude Code’s system prompt for Opus 5 and Fable 5 with no measurable loss on its own evaluations. The explanation is that many of those instructions were scaffolding written for weaker models, and on current models they actively create friction through conflicting rules, over-specified edge cases and context that crowds out the model’s own judgment. Anthropic lists six reversals of what were treated as best practices, including moving from rules to judgment, from front-loading everything to loading the right context at the right moment, and from stuffing memory into instruction files to letting the tool handle it.

This matters to marketers because everyone building a serious AI setup has been told to write a detailed instructions file, and many of those files have been growing for months. The emerging advice is to strip them back to what the model genuinely cannot work out for itself, which for a marketing team means your brand voice, your pricing floors, your approval authority and your confidentiality boundaries, and to move anything procedural into separate skills that load only when needed. Nobody should tear up their setup on the strength of one post from the vendor whose models benefit from the framing.

(This is precisely how to work with AI in 2026 and I teach this is exact set up inside Claude for Marketers Masterclass, including my now infamous Airbnb folder set up structure for optimal marketing results)

🔮 The Big Picture

First, the input to automation changed from describing to demonstrating.

This is the week’s real story and it is easy to undersell, because Record a skill sounds like a convenience feature. It is not. The reason most marketing work is still manual is not that the AI could not handle it, it is that nobody ever wrote down how the job is done, and writing that down well is a rare skill. Doing your own job while talking about it is not rare at all. When the barrier to automating something drops from “author a precise process document” to “narrate what you already do,” the set of tasks that can realistically be automated expands enormously. Notice that Brock Mesarich’s framework in the Playbook is really about the same thing from the other direction: the value is highest exactly where no integration exists, which is where the undocumented, unautomatable work has always lived.

Second, capability moved this week and plumbing did not. Opus 5 arrived at the same price with a real jump in capability. HubSpot handed agent-building to any Professional customer. Webflow put governance rails around agents editing live websites. In the same week, the CMO Council found that 71% of marketing organisations cannot effectively use their own first-party customer data and a third are wrestling with fragmented systems. Those two facts sit awkwardly together, and they point at where the constraint now lives. The gap between marketing teams is going to be decided less by which AI they use and more by whether their data and processes are in a state the AI can act on. That is unglamorous work, and it is now the highest-leverage work.

Third, AI search is still genuinely winnable, and your SEO history will not win it. The Semrush research found more than half of all categories have no established brand in ChatGPT, that almost 90% of AI search demand sits in those unowned categories, and that traditional search strength barely predicts who ends up owning one. Set that alongside the finding that buying an ad in Google’s AI Mode does not get you cited in the answer, and a consistent picture emerges: paid placement, organic ranking and being named in AI answers are three separate channels that do not feed each other. Most brands are treating them as one. The opportunity is that this is an unusually open field for a channel this large, and the window will not stay open indefinitely, because the same data shows that once a brand does establish ownership it keeps it nine times out of ten.

Tell me: What’s your biggest takeaway from this weeks news from the world of AI Marketing? Love to hear your thougths in the comments.

Leave a comment

Thanks for reading the Marketing with AI Weekly Roundup! If you enjoyed this read, the best compliment I could receive would be if you shared it.

Marketing with AI Weekly Roundup: China's Kimi K3 Cracks the Model Market Open, CMOs Stop Betting on One AI, and Your Ad Accounts Move Into the Chat Window

Robert Gillespie — Sun, 19 Jul 2026 12:20:52 GMT

Welcome to issue #43 of the Marketing with AI Weekly Roundup. Published every Sunday, it shares the top stories, playbooks, data and quick hits from the world of Marketing with AI over the last 7 days.

Subscribe to receive the week's most important AI marketing stories direct to your inbox.

Subscribe now

Create Stunning, Animated
‘5-Minute Slides’ Like These — Using Claude

Grab these Slide Packs where I hand you the exact skill files and show you, step by step, how to turn any script into a beautiful, scroll-stopping slide deck — with no designer, no PowerPoint, no Canva, and zero design skill required.

Limited Time Offer: Only $27 for 6 Slide Packs

Learn More About 5-Minute Slides

This was the week the “which AI is best?” question stopped being the right question.

On Thursday, a Chinese startup you may not have heard of released the most powerful open model the world has seen, at roughly 40% of the price of the frontier US models, and the AI industry lost its composure over it.

In the same few days, new research into 100 CMOs said the smartest marketing leaders have already moved on from picking a winner: they now run a stack of models, sending each job to whichever one does it best.

And a wave of tools quietly shipped connectors that let you run your Google Ads, LinkedIn Ads and CRM from inside your AI assistant, with a human approval step before anything goes live.

The thread running through all of it: your durable advantage is no longer the model you happen to use, because the models are converging and getting cheaper by the month. It is the system you build around them, the context, the workflows, the judgment about which tool does which job, and your ability to move that system to whatever wins next.

This week’s three best practitioner videos say exactly that, in three different ways.

Below: the four stories to act on, three steal-this workflows for building a portable AI system, the tools worth your time, and a reality check on what AI search is actually doing to your traffic.

Let’s get into it.

🔥 Top Stories

1. Kimi K3: the cheap, open Chinese model that rattled Silicon Valley

On 16 July, Chinese startup Moonshot AI released Kimi K3, and by Friday it was the only thing the AI world could talk about. The headline numbers: 2.8 trillion parameters, making it the largest open-weight model ever released, a 1 million-token context window, and multimodal understanding across text and images. In blind testing on the Arena evaluation platform, developers preferred Kimi K3 over every leading US model for front-end coding, including Anthropic’s Fable 5 and OpenAI’s GPT-5.6 Sol, and it outranked the standard Claude Opus 4.8 on broader text tasks. Moonshot says it still trails Fable 5 and GPT-5.6 Sol on overall performance, but it beat Opus 4.8 and GPT-5.5 on coding and agent benchmarks. The part that matters commercially: it costs roughly 40% less than the top US models, about half the price of GPT-5.6 Sol, and it is open-weight, with the weights due to be released publicly on 27 July so anyone can download, inspect and run it. Arena’s CEO called it “the single biggest release of the year.” For context, Chinese open-weight models now occupy the top five spots for weekly usage on the developer marketplace OpenRouter.

Why it matters / what to do: You almost certainly will not switch your marketing work to Kimi this week, and you should not. The weights are not even out until 27 July, the benchmarks are early and vendor-led, and its standout skill is coding, not marketing. The real takeaway is strategic, and it is the through-line of this whole issue: frontier-level AI is getting cheaper and more open at a speed that makes vendor loyalty a liability. Keep your prompts, your brand context and your workflows portable, so that when a model wins on price or quality next quarter you can move your system to it in an afternoon rather than rebuilding from scratch. If you run high-volume AI work (bulk content variants, classification, large-scale research) or you handle sensitive data you cannot send to a US cloud, a capable open model you can access cheaply or self-host is now a genuine option worth tracking. Action: put a note in your calendar for 27 July, watch what independent testers say about real-world reliability, and read the two Playbook items below on keeping your setup model-agnostic.

2. NotebookLM is now “Gemini Notebook,” and it just grew a data-analysis brain

On 16 July, Google renamed NotebookLM, the research tool used by more than 30 million people and 600,000-plus organisations, to Gemini Notebook, folding it into the main Gemini brand. The rename is cosmetic, but the upgrade underneath it is not. Every notebook now gets a secure cloud computer that can write and run its own code to do real data analysis, plus more than 100 curated “skills” and a wider set of output formats. Previously this power was limited to the top Ultra tier; it is now rolling out to all Pro users on the web over the coming weeks, and notebooks are becoming viewable in the Gemini app and, soon, inside AI Mode in Search.

Why it matters / what to do: Gemini Notebook is already a workhorse for marketers who use it to synthesise research, build briefs, mine voice-of-customer language and repurpose content. The cloud computer changes what it can do with numbers. It turns a cheap research tool into something that can take a messy spreadsheet of campaign data, GA4 exports or Search Console figures, actually crunch it, and hand back an analysis with charts, rather than just talking about your data in the abstract. Action this week: update the name in any internal docs or skills that reference NotebookLM, then feed it one real data file, last month’s ad spend or a GA4 export, and ask it to find the three things worth acting on and chart them. It is the fastest way to feel the difference the upgrade makes.

3. The “AI stack” era: 100 CMOs reveal they now match the model to the job

New research from Profound and Listen Labs, released this week and first previewed at Cannes Lions, is built on interviews with 100 CMOs across technology, financial services, healthcare and professional services. The findings map neatly onto Kimi’s arrival. 90% of these marketing leaders now use large language models daily, and 85% say their usage climbed in the past six to twelve months. Crucially, they are not standardising on one platform. They are building AI stacks, routing each task to whichever model handles it best: Claude is preferred for strategy and long-form writing (44%), ChatGPT for content creation and copywriting (29%), Gemini for Google-ecosystem and data work (22%), and Copilot for Microsoft-centric tasks (12%). Three more numbers stand out. 22% of CMOs now begin vendor research inside an LLM, ahead of the 16% who start with traditional search (peer recommendations still lead at 49%). 47% have moved beyond one-off prompting into structured workflows, automated reporting and custom agents. And the org chart is shifting: 28% now require AI fluency when hiring, and 32% expect AI-driven restructuring or smaller teams. (Worth noting: Profound sells AI-search-visibility tools, so it has a commercial interest in the “buyers now research inside LLMs” finding. The model-by-task breakdown is still the most useful snapshot of how marketing leaders actually work right now.)

Why it matters / what to do: This is the data behind the whole “stop betting on one model” idea, and it hands you a ready-made map. Stop asking “which AI is best” and start asking “which AI for which job,” then assign your recurring tasks accordingly: strategy and long-form to Claude, high-volume copy to ChatGPT, anything living in Sheets or Google Ads to Gemini. Then pick the single workflow you run most often and move it from ad hoc prompting to a saved system, a reusable skill, a scheduled task or a documented process. That one shift is what separates the 47% pulling ahead from everyone still typing prompts from scratch. The three Playbook items below show you exactly how.

4. Run your ad accounts from inside Claude and ChatGPT: the ad-platform MCP wave

On 13 July, a company called Markifact launched a hosted Google Ads MCP server. In plain terms, it connects your Google Ads account to AI assistants like Claude and ChatGPT so you can ask, in normal language, for performance reports, investigate wasteful search terms, and prepare campaign, budget, keyword and ad changes, with every change that writes to your account held for your approval until you review it. Markifact runs the same kind of connector for Meta Ads and Google Search Console, covering 500-plus operations across 30-plus platforms. It lands on the heels of AdKit’s LinkedIn Ads MCP (announced 9 July), which does the equivalent for LinkedIn: build audiences, launch Sponsored Content, flag wasted spend and draft client reports, all held as drafts for sign-off. (MCP, the Model Context Protocol, is simply the standard that lets an AI assistant safely call an outside tool or data source.)

Why it matters / what to do: For B2B marketers, this collapses paid-media reporting and routine optimisation from a dashboard chore into a conversation, with no engineering required and a safety gate built in. The right way to start is deliberately cautious. Connect a read-only Google Ads or Search Console MCP first, ask it for a plain-English review of last week’s performance and where you are wasting budget, and only turn on write access once you trust what it is telling you. Even used purely for reporting, it can save a marketer hours a week and surface the search-term waste that dashboards bury.

📋 The Playbook

1. Build one portable AI marketing system that runs on any model (Grace Leung)

If Kimi K3 taught us anything this week, it is that you should not hard-wire your work to a single AI. Grace Leung’s walkthrough is the clearest guide yet to building a marketing setup you can pick up and move between Claude, OpenAI’s Codex and whatever comes next. Her core idea: the model matters less every month because the frontier models are converging, so your durable asset is the “harness,” the portable workspace around the model, not the model itself.

The five parts of a portable system. Think of your setup as five components, all of which live in plain files you own: (1) knowledge and context, what the AI knows about you and your brand; (2) instructions, the rules for how it should behave (in Claude, this is a CLAUDE.md file); (3) skills, reusable workflows it can trigger; (4) sub-agents, saved role definitions it calls when needed; and (5) tools, live connections to real systems like your analytics or Search Console via MCP.

Why it’s portable. Grace splits those five into three buckets. Your context and knowledge are just files, so they move anywhere instantly. Your instructions, skills and sub-agents are the same content but each platform packages them slightly differently, so they need light translation. Only your tool connections need genuinely reconnecting when you switch platforms. Understanding this is what stops platform lock-in from ever trapping you.

The build, step by step.

1. Create a local folder called marketing with subfolders for each channel (paid, content, email, SEO). Add a core context file describing your product and offerings, a file listing your MCP connections, and a CLAUDE.md with your standing instructions.

2. Inside it, create a .claude/skills/ folder and drag in skill folders to install them; you then call each one with a slash command.

3. Build your first sub-agent, a data analyst, and give it a data-visualisation skill with memory switched on. Now give it a real task: “How has the website performed in SEO over the last two weeks?” It pulls your Search Console data through the MCP connection, runs the data-viz skill, and saves on-brand charts straight into your SEO folder.

4. To prove portability, load the exact same folder in OpenAI’s Codex. It auto-detects the Claude configuration and maps it across. Build a second sub-agent there, a growth specialist, that opens your homepage in a browser, runs a full multi-resolution UX audit, and outputs a Word-document report.

5. Keep the two in sync with Grace’s free agent-migration skill, which converts your setup between Claude Code and Codex. Install it at the user-account level so every project can use it, and always let it back up first.

Do this now: create the folder, the CLAUDE.md and one shared brand-context file today, and add a single skill. That foundation is the thing that compounds. (Disclosure: the video is sponsored by HubSpot, which offers a free “Marketer’s Agent Skill Pack” of five portable skills. The method works with any tools.)

Watch here

2. The four-level Claude Cowork setup that runs your admin for you (Brock Mesarich)

Where Grace’s system is about portability, Brock Mesarich’s is about getting a non-technical marketer to a working setup fast. Cowork is the middle ground between Chat (thinking and text) and Code (for developers): it works inside folders on your own computer and uses your connected apps. Brock builds up in four levels.

Level 1, the foundation. First, create a new empty workspace folder and grant Cowork permission to work inside it (choose “always allow” so it stops asking). Cowork can only touch folders you give it. Then build a CLAUDE.md so you never have to re-explain yourself: use his “teach Claude who you are” prompt, which has Claude interview you with seven questions (who you are, who you serve, your voice and writing rules, your folder structure) and write the file for you. Test it by pointing it at a messy folder with a “clean up this folder” prompt; it sorts everything into subfolders in about a minute.

Level 2, skills and connectors. Skills are reusable workflows. The three Brock runs daily: a visualise skill (”visualise how X works” produces an interactive HTML dashboard), a PDF-guide skill (research a topic, get a branded PDF explainer he uses as a lead magnet), and an email-voice skill that reads your last 30 sent Gmail messages, builds a profile of your greetings, sign-offs, length, tone and phrases, and drafts new emails that sound like you. Install a skill via Customize, then Plugins, then Add, then Upload. Add connectors via Customize, then Connectors, then Browse connectors; he connects Gmail and Google Calendar, then prompts Claude to pull an inbox-and-calendar rundown and draft replies. For any app without a native connector, the Zapier MCP connector unlocks 9,000-plus apps (he uses it for Beehiiv newsletter stats and his Skool community).

Level 3, automation. Scheduled tasks are prompts Claude runs on a timer. His live examples: a morning admin dashboard at around 8am (Gmail inbox plus calendar plus AI news, as an HTML page), a 7am community scan that reads his Skool community and surfaces posts needing attention plus content-gap ideas, and a daily 7am briefing that finds the five most important stories in his field. Build one by writing the prompt, then choosing “Schedule” (you can also “Run now”).

Level 4, your operating system. These are live mini-apps pinned inside Claude that pull from your connected apps: a competitor-content tracker showing rivals’ top-performing videos with live view counts and an outlier score, and a “Cowork OS” dashboard built from a single prompt (”Build me a live dashboard I can open every morning that pulls from email and calendar and shows today’s meetings, anything I need to prep for, and which emails I still need to reply to”).

Do this now: create the workspace folder and run the seven-question CLAUDE.md interview, then build the email-voice skill. That alone changes how your mornings run. (Disclosure: Anthropic partnered on the video; Brock also offers a free prompt pack with copy-paste prompts for each step.)

Watch here

3. Match the Claude model (and effort level) to the job so you stop burning usage (Eliot Prince)

The Profound study says match the model to the task. Eliot Prince shows you how to do that inside one tool, prompted by a painful story: his fiancée accidentally left the most powerful model running for everyday work and burned through her entire monthly Pro-plan usage. Here is his ladder, from cheapest and fastest to most capable and expensive.

● Haiku 4.5, the intern. Instant answers, very efficient, little deep reasoning. Use it to summarise an email thread or a webpage, pull action items from a transcript, or power small scheduled tasks.

● Sonnet 5, the everyday employee. Versatile knowledge work with a 1 million-token context (roughly a dozen books in one chat): research, writing, light analysis, a few multi-step jobs. It makes slightly more mistakes on hard reasoning. Use it to turn a YouTube transcript into a blog post, draft invoices from a template skill, or prep a one-pager before a sales call.

● Opus 4.8, the senior specialist. Deep reasoning, fewer mistakes, strong at following rules, but slower and pricier. Use it for pricing critiques (feed it your revenue, offers and competitors and ask where you are undercharging), contract-risk analysis, or important financial spreadsheets.

● Fable 5, the external consulting team. Top-tier and near-autonomous: give it a goal and it plans, splits into sub-agents and executes end to end. Use it for your hardest problems, for example “build me a course” (in the demo it pulled customer emails and LinkedIn profiles via connectors, built an ICP analysis, researched competitors and outlined the whole course) or cleaning up 2,000 messy CRM records.

On effort levels (found on the Effort tab for Sonnet and Opus): Low is fast and burns the least usage; High is the balanced default and his recommendation when unsure; Max is more thorough but slower and burns your limits fastest. Most people wrongly default to Max thinking it is “best,” so save it for genuinely hard problems. If Sonnet is over-thinking a simple task, turn thinking off for a faster, cheaper answer.

His pro pattern: use Opus or Fable to do the deep analysis and planning, then switch to Sonnet to execute the simple steps it laid out. The expensive model does the hard thinking; the cheap model does the cheap work.

Do this now: build a one-page cheat sheet of “which model and effort level for which task” for your five most common jobs. The payoff is immediate: you stop wasting premium usage on work a cheaper setting handles fine. Timely note: Fable 5 is only included in standard Claude plans until 19 July; after that it draws on extra credits, so this is the week to point it at your hardest problem.

Watch here

🛠️ New Tools & Features

1. Dovetail customer-segment digital twins

Dovetail, the customer-intelligence platform used across a large share of the Fortune 500, added simulated “digital twins” of your target segments (14 July) for on-demand research. You can pressure-test a message, a positioning line or a campaign concept against a synthetic version of your buyer segment before you commit budget to real fieldwork. Treat the output as directional rather than a replacement for talking to actual customers, but as a fast first gut-check on messaging, it is a useful new step.

2. GetWhys turns buyer interviews into ready copy

GetWhys (16 July) generates landing pages and marketing emails directly from transcribed buyer interviews. Instead of copy invented from a blank page, it grounds the words in the language your actual buyers used about their problems, which is exactly the voice-of-customer input that makes B2B copy convert. Feed it your win/loss or discovery-call transcripts and it drafts pages that already speak your market’s language.

3. 6sense ships an MCP server for buying-intent data

6sense (16 July) released an MCP server that exposes its account buying-intent signals to external AI assistants. For ABM and demand-gen teams, that means you can ask your AI assistant which accounts are showing intent this week and have it draft the outreach, rather than logging into yet another platform. It is part of the same pattern as this week’s ad-platform connectors: your martech stack is becoming a set of endpoints your AI can call in plain language.

📊 Research and Data

1. Google says AI search sends “billions of clicks” a week. The independent data disagrees.

On 17 July, Google’s search chief Nick Fox used a LinkedIn post to push back on the fear that AI is killing web traffic, stating that Google is “now sending billions of clicks to websites every week through AI features in Search alone,” on top of the billions it sends daily through Search overall. It was the first time Google has put a number on AI-feature traffic. The catch, flagged immediately by the trade press, is that the figure came with no baseline, denominator or methodology, and cannot be compared to your own site’s numbers. Independent research points the other way. A Pew Research analysis found users clicked a traditional result on just 8% of visits where an AI summary appeared, versus 15% without one, and clicked links inside the summary on only 1% of visits. A separate field experiment from the Indian School of Business and Carnegie Mellon found AI Overviews cut organic clicks by 38% on the queries where they showed.

What to do: Both things can be true at once, a vast total number of clicks and a falling share of clicks for any individual page, so do not take Google’s reassurance at face value for your business. The only number that matters is yours. Set up a custom channel group in GA4 to separate AI-referral traffic (ChatGPT, Perplexity, Gemini, Claude) from “direct,” and track your own click-through on the queries where AI answers appear. Measure it rather than trusting either the optimistic or the doom-laden headline.

2. Where AI is actually landing in day-to-day teams

A CMSWire analysis of how customer-experience teams use AI in 2026, drawing on a Gartner survey of customer-service leaders, found 91% are under executive pressure to implement AI, and identified the four workflows where it is genuinely in production: support-ticket triage and theme analysis, tagging survey verbatims against a fixed taxonomy, detecting gaps in knowledge-base content, and rolling up cross-functional feedback for product, support and leadership. The striking thing is that these are analytical and organisational jobs, not flashy content generation.

What to do: the same pattern is the fastest win for marketing teams too. The reliable value right now is in organising and understanding your existing data, not just producing more of it. Point AI at your own qualitative pile, sales-call notes, support tickets, review text, and have it tag themes against your personas and surface what keeps coming up. That feeds straight into sharper messaging, content topics and positioning, and it is far lower-risk than publishing raw AI output.

💡 Quick Hits

1. ChatGPT Work reached Plus and Business users this week.

OpenAI also added “Company Knowledge,” which pulls context across your connected apps for business-specific answers, plus new connectors for Asana, GitLab Issues and ClickUp. Read more: https://help.openai.com/en/articles/6825453-chatgpt-release-notes

2. ChatGPT Ads added location and audience exclusions.

Advertisers can now block a campaign from specific geographies and suppress selected audience lists (for example, existing customers), with exclusions overriding inclusions. Read more: https://www.seroundtable.com/chatgpt-ads-location-audience-exclusion-41693.html

3. Google AI Overviews will start generating images inside answers.

Announced for Google Images’ 25th anniversary and built on Google’s Nano Banana model, it turns a text prompt into a custom visual within the Overview, rolling out over the coming weeks in English. Read more: https://www.searchenginejournal.com/google-adds-image-generation-to-ai-overviews-revamps-images/582242/

4. ChatGPT raised its custom-instructions limit.

Teams can now encode richer brand voice, workflow rules and guardrails once, so they carry across every conversation instead of being restated each time. Read more: https://www.seroundtable.com/recap-07-16-2026-41698.html

5. Cision built AI-search visibility into CisionOne

This allows tracking how often brands appear in AI-generated answers alongside its traditional PR and media metrics. Read more: https://martech.org/the-latest-ai-powered-martech-news-and-releases/

👀 On Our Radar

Everything is shipping an MCP server.

This week alone brought MCP connectors for Google Ads and Meta Ads (Markifact), LinkedIn Ads (AdKit), account intent data (6sense) and more, each following the same shape: your tool becomes something your AI assistant can operate in natural language, with a human approving the writes. The pattern to watch is where this goes next. Within a few months, “does it connect to my AI assistant via MCP?” is likely to become a standard question on martech buying checklists, the way “does it have an API?” or “does it integrate with our CRM?” is today. If you are evaluating new tools this half, start asking vendors that question now.

🔮 The Big Picture

Three forces converged this week, and together they redraw where a marketer’s advantage comes from.

First, the model is no longer the moat. Kimi K3 arrived as a near-frontier, open, 40%-cheaper model, and the 100-CMO study showed leaders have already stopped standardising on one platform. When capability is converging and price is falling this fast, loyalty to a single vendor is a cost, not a strategy. The durable asset is the system you build around the models, your context, your skills, your judgment about which tool does which job, which is precisely what all three of this week’s playbooks teach you to build and keep portable.

Second, your tools are becoming a chat box with an approval button. The MCP wave means the interface to your ad accounts, your CRM and your intent data is shifting from dashboards you click to instructions you type, with you signing off on the actions. As that spreads, the scarce skill stops being “knowing the platform” and becomes briefing the work clearly and reviewing it well. Supervising, not operating.

Third, trust the data you own, not the narrative you’re handed. Google’s “billions of clicks” message and the independent studies that contradict it are a reminder that in an AI-mediated web, everyone with a platform has a story to sell you about it. The marketers who win will be the ones measuring their own AI-referral traffic, their own citations in AI answers, and their own results, rather than taking any vendor’s framing as fact.

Put simply: build a portable system, not a bet on one tool; learn to supervise agents, not just operate apps; and measure your own reality.

🔭 Looking Ahead

● Fable 5’s free inclusion ends tomorrow, 19 July. After this Sunday it draws on extra usage credits in standard Claude plans, so if you have a genuinely hard, high-value problem, this is the weekend to point Fable at it.

● Kimi K3’s weights drop on 27 July. That is when independent testers can finally verify the benchmarks and when the cheap-access and self-hosting options become real. Watch how the price-versus-quality picture settles before making any move.

● Gemini Notebook’s cloud computer keeps rolling out to all Pro users over the coming weeks. If it has not reached your account yet, check back, then run a real data file through it.

● ChatGPT Work continues its rollout, with Company Knowledge now in the mix. Worth watching how it lands for smaller teams over the next fortnight.

Tell me: What’s your biggest takeaway from this weeks news from the world of AI Marketing? Love to hear your thougths in the comments.

Leave a comment

Thanks for reading the Marketing with AI Weekly Roundup! If you enjoyed this read, the best compliment I could receive would be if you shared it.

Marketing with AI Weekly Roundup: ChatGPT Work Ships Finished Files, Claude Cowork Goes Mobile, and Google Ads Makes AI Your Default

Robert Gillespie — Sun, 12 Jul 2026 10:01:55 GMT

Welcome to issue #42 of the Marketing with AI Weekly Roundup. Published every Sunday, it shares the top stories, playbooks, data and quick hits from the world of Marketing with AI over the last 7 days.

Subscribe to receive the week's most important AI marketing stories direct to your inbox.

Subscribe now

The Claude for Marketers Online Masterclass

Last day to get 10% off with discount code ‘SUBSTACKJULY10’

(Now priced in $ USD + split your payments over 2-months option available)

Learn More About Claude for Marketers

This was the week the AI agent walked out of the chat window and started doing the job.

OpenAI launched ChatGPT Work - an agent that takes a goal, pulls context from your connected apps, and hands back finished decks, spreadsheets, reports and websites, working on its own for hours.

Two days earlier, Anthropic pulled Claude Cowork off the desktop and onto web and mobile, with sessions that keep running when your laptop is shut.

And in the background, Google quietly rewrote its Ads Terms of Service so that AI is now authorised by default to write, pick and format your ads - while leaving you holding the liability.

There’s a single thread running through all three: the work is shifting from *driving* AI to *supervising* AI that acts on your behalf. The agent now has write access to your inbox, your ad account and your files - which makes the review step, not the drafting step, the job. This week’s research says the same thing from the other side: the content AI can produce on its own gets cited least, and the human-checked layer is where the value now sits.

Below: the four stories to act on, two steal-this workflows with the exact prompts (an AI marketing agent build and a Claude-plus-Apify lead engine), the tools worth your time, and the data that should reframe what you publish.

Let’s get into it.

🔥 Top Stories

1. ChatGPT Work: OpenAI’s agent that hands back finished files, not answers

OpenAI launched ChatGPT Work on 9 July — a dedicated work agent that lives inside ChatGPT and, instead of replying in the chat, produces the actual deliverable: a document, a spreadsheet, a slide deck, a report, or an interactive “Site” (a hosted web page or dashboard). You give it an outcome; it gathers context from your connected apps, files and browser tabs, breaks the job into steps, and works independently for hours, pausing for approval at key moments. It runs on OpenAI’s new GPT-5.6 model (which hit general availability the same day, in Sol/Terra/Luna tiers) with Codex technology underneath, so a non-coder can point it at a folder and let it work. It ships in a merged desktop app — Chat, Work and Codex in one window — and OpenAI is retiring its standalone Atlas browser alongside. Rollout is to paid plans first (Pro, Enterprise, Edu), with Plus and Business following over the coming days.

Why it matters / what to do: This is OpenAI’s direct answer to Claude Cowork, and it changes what “using ChatGPT” means for a marketing team — from asking questions to assigning work. The practical unlock is that it connects to the tools you already run on (Google Drive, Notion, Gmail, Calendar) and returns on-brand assets saved straight to your folders. Try one real deliverable this week: give it a multi-step job you’d normally block out an afternoon for — “read these three competitor sites and build me a positioning one-pager plus a five-slide deck” — and judge the output against what a normal chat would’ve given you. (Our first Playbook below is a full ChatGPT Work marketing-agent build.)

2. Claude Cowork breaks off the desktop — now on web and mobile, running remotely

Anthropic expanded Claude Cowork to web and mobile on 7 July (it was desktop-only before), and the bigger change is what happens underneath: sessions now run remotely and are saved to your Claude account, so they follow you across devices, keep working when you close your laptop, and let scheduled tasks run with no device online at all. Chat and Cowork now share one home — a single place for your projects and artifacts across both — and the rollout starts with the Max plan, with more to follow. (Separately, Claude’s Microsoft 365 connector gained *write* tools this week — see New Tools.)

Why it matters / what to do: Until now, a Cowork workflow was chained to the machine you started it on. Now the useful pattern is device-free, scheduled autonomy: a weekly reporting run, a competitor sweep, or a content-repurposing job can run overnight or while you’re in back-to-backs, and be waiting for you on your phone. If you parked Cowork because it was tethered to your desktop, this is the week to switch a recurring task back on and let it run in the background. Our second Playbook — a Claude lead-generation engine — is built to run exactly this way.

3. Google Ads quietly rewrote its Terms of Service — AI can now build your ads by default

On 1 July, every Google Ads account was automatically bound to a revised Terms of Service — no login prompt, no re-acceptance, the first substantive rewrite since 2018. The key clause authorises Google “to serve ads, including through the use of automated program features to format, select, or generate targets, ads, or destinations on Customer’s behalf.” In plain terms, AI automation moved from an opt-in feature to a default authorisation, covering Performance Max, AI Max, Demand Gen and the Gemini-powered conversational campaign setup. The catch: liability stays with you. The terms are explicit that you remain responsible for reviewing, approving, editing or removing anything the AI generates — so if an AI-written headline makes a claim you can’t back or goes off-brand, that’s on your account, not Google’s.

Why it matters / what to do: If you run any Google Ads — including B2B lead-gen PMax — the “Google’s automation made it” defence no longer exists. Add a standing AI-asset review step to your workflow this week: check AI-generated headlines, descriptions, URLs and audience expansions for unsupportable claims, trademark issues and off-brand copy before they run. Also worth knowing: the inputs you type into conversational campaign setup, and the URLs you let Google crawl, are now formally part of Google’s signal pool — so treat them as such.

4. “AI poisoning”: rivals can now seed misinformation into how AI describes your brand

Digiday’s “WTF is AI poisoning?” (6 July) names a risk that’s moved from theoretical to live: because AI search answers are now a core discovery channel, they’re also a competitive arena — and a vector for bad actors. AI poisoning is the practice of planting bad-faith reviews or outright misinformation about a rival (on Reddit, YouTube, Facebook, Instagram and similar) in the hope an LLM ingests it and later repeats it as fact. It’s black-hat SEO for the AI era, and it cuts both ways: your competitors can do it to you. What makes it dangerous is trust — users tend to read AI answers as neutral, and by one figure cited in the piece, just 8% of people double-check the facts in an AI-generated answer.

Why it matters / what to do: Your brand’s reputation is now partly written by machines quoting sources you don’t control. Add one question to your monthly brand-health check: “How do ChatGPT, Perplexity and Gemini describe us — and is any of it wrong or out of date?” Run your key buyer prompts, screenshot what the models say, and document a correction path — owned content and earned media that gives the models an accurate, well-sourced version to pull from. You can’t police every forum, but you can make the truth easier to find than the poison.

📋 The Playbook

1. Build a marketing agent in ChatGPT Work (skills + connectors + one scheduled report)

Grace Leung’s walkthrough is the clearest guide yet to turning the new ChatGPT Work into an actual marketing operator rather than a smarter chatbot. Here’s her system, rebuilt as a step-by-step you can run this week. Everything happens in the desktop app — that’s where Work gets access to your local files, which is where the power is.

Step 1 — Create a project tied to a folder. Open the desktop app and start a project pointed at a real folder (say, a client or campaign folder). From now on, every asset the agent produces saves straight into that folder — no copy-pasting outputs back out. Think of the folder as the agent’s desk.

Step 2 — Install a “skill” for a job you repeat. A *skill* is a repeatable workflow packaged so the agent runs it the same way every time. Go to the plugin panel, open Skills, and install one at the user-account level so it’s available everywhere. Grace uses a brand-design skill that carries her fonts, colours and layout rules, so every deck and one-pager comes out on-brand. If you already have brand guidelines written down, that’s your first skill.

Step 3 — Brainstorm in “quick chat,” then “Attach to Task.” This is the move that separates Work from normal ChatGPT. Open the lightweight quick chat, attach a source (e.g. a deep-research report), and think out loud: *”Give me the three findings in this report that matter most to my client, then pick the single biggest opportunity we should focus on.”* Once you’re happy with the thinking, click Attach to Task — it packages the whole conversation as context, hands it to the Work agent, and you tag the folder and the brand-design skill. Then: *”Based on our discussion, build the final strategy deck using my brand-design skill.”* The agent builds an on-brand, fully editable deck and saves it to the folder.

Step 4 — Connect your real stack. In plugins, connect the tools you actually use — Google Drive, Notion, Gmail, Calendar (Asana, Canva and Figma are there too). This is what turns Work from a document generator into a workflow engine. Two flagship automations to build first:

• Content calendar → finished social visuals. With Notion and Drive connected: *”Open my Notion content calendar, import every post scheduled between [explicit start date] and [explicit end date], generate an on-brand social visual for each using my brand-design skill, save them to the ‘Social Visuals’ folder in Google Drive, and mark each Notion entry done.”* One hard-won lesson from her run: give an explicit date range, not “the next three weeks” — the agent skipped posts when the range was vague.

• Scheduled GA4 report → branded PDF, in your inbox. Schedule a weekly GA4 email export (a CSV) to yourself. Then tell Work: *”Check Gmail for the most recent email with a subject containing ‘Google Analytics report’, confirm it covers the last 7 days, and if so create a branded weekly performance PDF summarising the numbers and what we should act on — then reply to that same email with the PDF attached.”* Save it as a scheduled task (e.g. every Monday 8am) and the report writes itself each week. The same pattern works for any data source you can email in.

Step 5 — Turn a recurring document into a reusable skill. Use the template creator: attach a document you fill in repeatedly (a campaign brief, a client-onboarding brief) and say *”Turn this into a reusable template skill.”* Next time, point Work at a messy folder of new-client materials and it drafts the brief in your standard format automatically — reusable across every future project.

Try this week: pick your single most-repeated deliverable, install one brand skill, and schedule one task (the GA4 report is the easiest win). One project, one skill, one scheduled job — that’s a working marketing agent.

Watch here:

2. The Claude + Apify lead engine: from zero to a verified outbound list (with the prompts)

Eliot Prince’s workflow builds a complete outbound lead system inside Claude Cowork — research, list-building, enrichment, offer and drafted emails — and then packages the whole thing into a one-click skill you rerun monthly. He demos it for a local-services business, but the method transfers to any outbound motion: swap the ICP and the scraper source for your target accounts. Runs best in Cowork on the desktop (it needs your local files to save the spreadsheets). Set up a project folder loaded with your business info first.

Step 1 — Build the ICP. Grab a persona to sharpen the thinking:

“You are Alex Hormozi, expert in customer acquisition. I run [your business] targeting [segment/region]. Build me an ideal customer profile: the best account/industry types, who the decision-maker is, the triggers that make them say yes, what makes them bin my email, and their pains, problems and dreams.”

Let Claude reason through *which* segments to attack first (in the demo it argued for the higher-volume segment over the higher-ticket one, on attempts-per-hour maths). Don’t let it jump to writing scripts yet — you have no list.

Step 2 — Scrape a real list with the Apify connector. One-time setup: in Claude, go to Customize → Connectors → browse connectors → Apify, install it, hit Configure, and paste your Apify API token (create a free account at apify.com — you get $5 of free usage a month, plenty for this). Then set the connector’s permissions to auto-allow so you’re not clicking “approve” on every step. Now prompt:

“You are Aaron Ross, author of Predictable Revenue. Use Apify to scrape [source — e.g. LinkedIn, Google Maps, an industry directory] for [ICP types] in [region]. Return one clean table: company name, website, decision-maker name, email, phone. Get every matching organisation — at least 50 — and put it in a spreadsheet.”

Always set a minimum (”at least 50”) or the agent gets lazy and returns ten. Claude picks the right Apify scraper itself and runs it. In the demo it pulled 158 results in about 9 minutes while he did nothing — roughly 116 usable contacts. The list looks messy inside Claude; open it in Google Sheets and it’s clean.

Step 3 — Enrich and verify. Add the Vibe Prospecting connector (on the Claude marketplace, 400 free credits to start) and ask Claude to run the list through it: fill missing emails, find the local/branch decision-maker rather than head office, and score each contact’s validity (high / verified). This is what takes a raw scrape to a sendable list.

Step 4 — Manufacture the offer angle. Don’t lead with “do you need X.” Prompt:

“You are Alex Hormozi, author of $100M Offers. I run [business] targeting [ICP]. Give me five cold-outreach offers to lead with right now, each tied to a real trigger — a season, an event, an industry deadline. For each: the hook, the offer, and why this buyer cares.”

Feed it *specific current context* (a looming deadline, a seasonal spike, a recent industry change) so the angles are real, not generic.

Step 5 — Draft in your voice, not “AI voice.” Eliot’s edge is a voice-copywriter skill trained on a large body of his own writing, with anti-AI rules, a banned-word list and spelling conventions baked in. Build your own the same way — feed Claude a big sample of your emails and posts and correct it until it sounds like you — then: *”Draft the cold email using my direct-response and voice-copy skills.”* Take two or three passes to tighten it.

Step 6 — Draft into Gmail. With the Gmail connector on: *”Draft an email to the first 10 people on the list in my Gmail.”* Claude drafts into your inbox (it won’t send by default) so you can eyeball and send.

Step 7 — Wrap it into a one-click skill. Here’s the payoff. At the top of the task, hit Turn into skill and describe the full flow — intake → ICP research → Apify scrape → enrichment → offer angles → two-or-three-pass email → Gmail draft. Save it. Now every month you open a new task, type /lead-gen-workflow, and it walks the whole engine again from scratch — or schedule it to run on its own.

Try this week: just build Steps 1–3 for one target segment — the ICP and the first Apify scrape. Getting a clean, real list out of a single prompt is the moment this clicks.

Watch here:

🛠️ New Tools & Features

1. ChatGPT Sites — publish interactive pages by prompt

Alongside ChatGPT Work, OpenAI opened up Sites: turn data or analysis into a shareable, interactive dashboard, prototype or web page from a prompt, hosted by OpenAI. As of 9 July, Business and Enterprise can publish Sites *publicly* via URL (admin-gated in Enterprise), and Sites entered public beta for Pro and Edu. For a marketer, that’s a live campaign-health page or an interactive one-pager without pulling in a designer or developer. (Public publishing isn’t available in the EEA, Switzerland or UK at launch.)

**2. Claude’s Microsoft 365 connector gains write tools**

Claude’s M365 connector moved beyond read-only this week (7 July): with write tools enabled and admin consent, Claude can now draft, send and organise email, manage calendar events, update mailbox settings, and create or update files in OneDrive and SharePoint. Paired with Cowork’s new remote sessions, that means an agent can now *action* your Outlook and Office work end-to-end — draft the follow-ups, file the docs — not just summarise it.

3. Profound Aim — an always-on agent for AI-search visibility

Profound launched Aim (2 July), billed as “the first background agent for marketers.” Instead of another dashboard, it continuously watches your brand’s AI-search signals — citation frequency, sentiment, accuracy, prompt volumes — and when something shifts, it diagnoses what changed, writes a memo, spins up a scoped marketing project (brief plus tasks) and routes the work, with a human approving each step. Even if you don’t buy it (tiers run $99/$399/custom), it’s a useful template for structuring your own “monitor AI answers → act on them” loop — which, given this week’s AI-poisoning story, is fast becoming table stakes.

📊 Research and Data

1. The content AI can write itself is the content that doesn’t get cited

A Search Engine Land analysis of AI-search traffic across ten websites (the “SEO-GEO gap” study) found a blunt pattern: generic educational content — the how-to posts and top-of-funnel guides that fill most content calendars — earned LLM citations just 12% of the time, while trends-and-analysis posts were cited 78% of the time and data-led posts 61%. The reason is uncomfortable but logical: LLMs can generate a competent “what is [topic]” explainer themselves, so they don’t need to cite yours. What they *do* reach for is original data, proprietary research and a genuine point of view they can’t produce on their own.

What to do: audit your next content batch against one test — “could an LLM already write this?” If yes, it won’t earn citations. Shift effort toward pieces built on your own data, benchmarks, customer insight or a clear argument. In an AI-saturated feed, originality isn’t a nice-to-have — it’s the citation.

2. Where marketers are — and aren’t — actually using AI yet

New Digiday+ Research (9 July) maps how unevenly AI adoption is spread across marketing disciplines: about 49% of marketers use AI for social and 42% for retail media, but only 25% use it for influencer marketing and just 18% for connected TV (i.e. 82% aren’t). The gap isn’t random — AI has landed first where the work is text- and data-heavy, and lagged where it’s relationship- or video-led.

What to do: read the map two ways. The high-adoption areas (social, search, email) are now table stakes — if you’re not using AI there, you’re behind the median. The low-adoption areas (influencer, CTV) are where there’s still white space to build an edge before everyone else catches up. Pick one under-adopted discipline you own and run a small AI experiment this quarter.

💡 Quick Hits

1. ChatGPT Ads can now auto-generate ads for you.

OpenAI’s ad platform added a “Generated ads for you” flow that drafts an ad you then review, edit and approve.

2. Google will now label AI-made ads across Search, YouTube and Discover.

A new “How this ad was made” panel in My Ad Center (launched 9 July, global) shows whether an ad was created or edited with AI — Google auto-labels ads built with its own tools, and advertisers must self-label ads made with other AI tools.

3. Meta expanded its AI business assistant to all advertisers and agencies worldwide (in beta) — across Ads Manager, Business Suite and Business Support Home, it handles campaign analysis, benchmarking, recommendations and troubleshooting; Meta says early users cut cost per result by 12%.

4. Google says Cloudflare’s “content-signals” robots.txt directive has no effect on any crawler or LLM. John Mueller called it “bloat” that no crawler currently honours — so don’t rely on it to control how AI uses your content.

5. Google Ads is changing how it optimises budget-limited campaigns to make CPA/ROAS targets more predictable when spend is constrained. Account notifications landed this week; the change goes live 17 August — review your budget-limited campaigns’ targets now.

🔮 The Big Picture

Three launches, one direction.

The agent has left the chat window — it now runs on your phone (Cowork), ships finished files from your connected apps (ChatGPT Work), and has standing permission to write your ads (Google).

In every case it gained two things at once: write access to your real systems, and, quietly, the liability for what it does with them. Google made that trade explicit in its Terms; ChatGPT Work and Cowork make it operational by putting an autonomous agent inside your inbox and files.

That reframes the marketer’s job.

When drafting was the bottleneck, skill meant producing. Now that agents produce and act, skill means supervising - briefing well, setting guardrails, and checking the work before it ships.

This week’s data underlines it from both ends: the content AI writes on its own is the content that doesn’t get cited, and the brand reputation AI reports is now open to poisoning.

The value has migrated to the human layer around the agent: judgment, originality and review. Delegate the doing. Own the deciding.

Tell me: What’s your biggest takeaway from this weeks news from the world of AI Marketing? Love to hear your thougths in the comments.

Leave a comment

Thanks for reading the Marketing with AI Weekly Roundup! If you enjoyed this read, the best compliment I could receive would be if you shared it.

The Marketer's Guide to Context Windows

Wed, 08 Jul 2026 12:31:01 GMT

Every Wednesday, one essay on what's actually working in Marketing with AI - tools, workflows, and mindset shifts you can apply immediately.

Subscribe now

Learn More About Claude for Marketers

(Use Discount Code SUBSTACKCONTEXT to get 10% off Claude for Marketers Masterclass this week only. Substack readers only.)

Lets’s get into this weeks newsletter…this is an important one for marketers to grasp as it changes how you work with AI forever.

The Marketer’s Guide to Context Windows

It’s Tuesday afternoon. You open Claude or ChatGPT, connect your brand guidelines, your audience notes and a rough brief, and start building a campaign.

The first few replies are excellent.

On brand, sharp, exactly what you wanted.

Then, somewhere around the fifteenth message, the responses start to drift. The tone flattens. It suggests an angle you explicitly ruled out earlier. It contradicts a decision you made twenty minutes ago.

By the time you notice, you’re arguing with it, re-pasting the brief, wondering why it suddenly “got dumber.”

So what’s going on?

You’ve likely been blaming the model. But the real problem is the context window. And the tools will never teach you or tell you.

Almost every marketer or marketing team I train or teach has had this exact experience, and almost none of them can name what caused it.

It isn’t a glitch, a bad prompt, or a lazy model.

It’s a hard limit sitting behind every AI conversation you have, one that quietly shapes the quality of everything the AI gives you.

And it’s the least understood, least talked-about idea in AI marketing, precisely because it isn’t a shiny new feature. It’s the invisible wiring behind every chat you have, and it shapes your results whether you notice it or not.

So this week, let’s fix that.

By the end of this you’ll understand what a context window is, why it silently degrades your work, and exactly how to manage one you can’t even see.

What a context window actually is

Let’s start with a picture.

Imagine a school whiteboard.

At the top you write your brand brief. Below it, your campaign goals. Then your questions, and the AI’s answers, filling the board top to bottom as the conversation grows.

Now the board is full.

To write anything new at the bottom, something has to be rubbed off the top.

So the oldest content goes first: your brand brief, your original instructions, the constraints you set at the very start. The AI keeps writing. But the content at the top is gone, wiped to make room.

No warning. No alarm. It simply can no longer see what it’s forgotten.

The context window is that whiteboard. It’s the total amount of information an AI can hold in its working memory at any one moment.

Everything inside the window, the AI can see and use. Everything that’s fallen off the edge simply doesn’t exist to it anymore.

Another way to picture it: every conversation is written on a single sheet of A4.

Every word you type, every reply it gives, every document you attach, all of it has to fit on that one sheet. When the sheet is full, the top lines get crossed off so the bottom ones can be written.

One thing to be clear on.

This is working memory for a single conversation, not long-term memory. It’s what the AI can hold in its head during this one conversation, not what it remembers about you forever.

Memory features are a separate thing, and we’ll come to them.

Tokens: The Currency You’re Spending

The whiteboard capacity we referenced earlier isn’t measured in words.

It’s measured in tokens, and tokens are the currency of everything you do with AI.

A token is a chunk of text, usually a piece of a word rather than a whole one. The rough conversions worth memorising:

• 1 token is about three-quarters of a word
• 1,000 tokens is about 750 words
• 1 page of A4 is about 500 words, so roughly 650 tokens

A couple of examples to make it real.

The word “ChatGPT” is actually three tokens.

The United States Declaration of Independence is 1,695.

And here’s one that catches global marketers out: the same meaning costs far more tokens in other languages. A French brief fills the window roughly twice as fast as the English version, an Arabic one around four times as fast.

Why does this matter?

Because tokens are the currency twice over. They’re what you’re billed on, and they’re what fills the window. Every token spent on an old message, a bloated PDF, or a hidden instruction is a token no longer available for the actual work.

What Actually Fills the Context Window, Faster than You Think

Most marketers assume the window fills up with their messages.

It fills with far more than that, and most of it is invisible.

The compounding problem.

This is the mechanic almost nobody knows.

AI chat has no memory between turns, so every time you hit send, the tool quietly re-sends the entire conversation so far back to the model, as if for the first time.

Your first message is cheap. But by message twenty, every new reply is re-reading all nineteen that came before it, plus every file you attached. A one-line “make it punchier” still drags the whole transcript along with it.

Your twentieth message isn’t a short message. It’s your twentieth message plus the nineteen before it, sent all over again.

The hidden overhead.

Before you type a single word, there’s a hidden system prompt loaded into the window (on some tools this runs to thousands of tokens), plus the definitions of any connectors and tools you’ve switched on.

You start every conversation already part-full.

The file trap.

This is where marketers get burned most, because we live in Word, PowerPoint, PDF and Excel.

Files are heavier than they look, and format has a hidden multiplier effect.

A raw PDF is processed page by page, like a scan, so a 10-page PDF can cost around 28,000 tokens. Copy the same text out and paste it in as plain text and it drops to around 6,500, an 80% saving for identical content. A dense spreadsheet can be tens of thousands of tokens on its own.

The takeaway: attaching a report isn’t “one file.” It’s thousands of tokens dropped into the window in a single move. Attach three or four and a smaller window is half gone before you’ve asked a question.

The Numbers: What You Actually Get

Here’s the part that will change how you work and think with AI.

The window a model advertises and the one you actually get in the chat app are usually very different, and it varies by plan, by model, and even by which app you open.

The huge numbers you see in headlines are almost always the developer API. The $20-a-month app or $25-team plan, is a different product, often with a much smaller window.

Here’s how the main tools actually stack up right now, with the entry monthly price alongside so you can weigh window against cost.

These numbers move constantly, so treat this as a mid-2026 snapshot, not gospel. Prices are the US list price for the cheapest plan that unlocks that window.

(The one figure nobody publishes is Claude Free: Anthropic doesn’t state a separate free-tier window, so free users get a standard model window but hit daily message limits well before it fills. (Grok’s app figures are reported by third parties, not published by xAI.))

Sit with the ChatGPT Free line. It has a 16,000 token window. That’s roughly 24 pages of A4 for your entire conversation, including the AI’s replies and any file you upload.

The same model family on the Pro model holds around 400,000 tokens, around 600 pages. A 25-fold difference.

Now the detail that really surprised even me, and it matters most if you work using Claude Cowork.
Claude Opus 4.8 gives you a full 1,000,000 tokens on the API, in Claude Code, and in Cowork. But open that exact same Opus in the Claude browser chat window and you’re capped at 500,000. Half the capacity, same model.

The only Claude model that gives you the full million tokens everywhere, including plain browser chat, is Sonnet 5 which launched last week, as of writing this newsletter.

And that leads to the comparison worth making if you’re deciding which tool to build your work around.

Set the developer API aside, since almost no marketer works there, and look only at the apps.

The most context you can get inside ChatGPT is around 400,000 tokens, and only on the Pro tier with Thinking mode switched on, which starts at $100 a month. Claude gives you a full 1,000,000 in its desktop app with Opus and Sonnet 5 both from the $20 Pro plan.

In the apps marketers actually use, Claude gives you roughly two and a half times the context window of the best ChatGPT can offer, from a plan that costs a fraction of the price.

That’s not a small gap. It’s the difference between holding a whole brand archive, a stack of competitor reports, or a full webinar transcript in one place, and having to chop it into pieces and hope nothing important falls off the edge.

This is precisely why Claude Cowork is the single most effective way to do end-to-end marketing work because it has a much higher context window than any other model. Claude Cowork allows you to build complex workflows, schedule automated tasks, and build agents to do that deep work for you. And it’s why I built the Claude for Marketers Masterclass to teach you how to use it effectively.

(Use Discount Code SUBSTACKCONTEXT to get 10% off Claude for Marketers Masterclass this week only)

The Degradation Problem: What Actually Happens Inside Your Conversations and Workflows

Here’s the twist that makes all of this urgent.

The window doesn’t just fill up and forget. Quality starts slipping long before the window is anywhere near full.

Researchers have a name for the first effect: “lost in the middle.”

Models pay the most attention to the beginning and the end of a conversation, and the least to the middle. Bury an important instruction halfway through a long chat and it may as well not be there.

Then there’s what happens over a long back-and-forth. A joint Microsoft and Salesforce study ran more than 200,000 simulated conversations across fifteen leading models, and found that when information was revealed gradually across turns, rather than all at once, average accuracy dropped by around 39%.
Their blunt summary: “when LLMs take a wrong turn in a conversation, they get lost and do not recover.”

And more context is not automatically better context.

In one study, a model given around 300 focused tokens of information outperformed the same model given 113,000 tokens of the same information. Stuffing the window with everything you have can actively make the answers worse.

What about hitting 100% of your context window?

At that point the oldest tokens fall out of the window entirely, and in most apps this happens silently, no error, no banner. Whatever was at the top of the whiteboard is simply gone.

One honest caution, because I see it claimed a lot. There’s no magic percentage where quality “collapses,” no clean 60% or 70% line.

Degradation is a gradient, not a cliff, and it tracks an absolute amount of text rather than a neat fraction of your window. The practical lesson isn’t a number. It’s that quality slips earlier than you’d expect, so you should reset sooner than you think.

(Here’s a fun explainer video I created with the new NotebookLM shorts video feature. Simply gave it this article and a few instructions on what to focus on and it came up with a nice animated video which I hope helps)

The Warning Signs: How to Know You’re Hitting the Limit

Because you can’t see the window, you have to read the symptoms. These are the tells that your whiteboard is getting crowded:

• Generic output. Answers that could have come from any AI, not one that knows your brand.
• Forgotten constraints. It recommends the exact thing you told it to avoid.
• Contradictions. It disagrees with a decision it helped you make earlier in the same chat.
• Repetition. It re-suggests ideas or copy you already covered.
• Re-asking. “Could you remind me of the brand voice?” when you pasted it eight messages ago.
• Vaguer, longer, slower. Replies get wordier but less useful, and the thread starts to lag.

Here’s a simple test to diagnose it. Take the same request into a brand-new chat with just the essential context. If the quality jumps back up, it was a context problem, so start fresh. If it’s still poor, it’s a prompt problem, so refine the prompt. Most people blame the model. The cause is almost always context.

How to Protect Yourself Day to Day

You can’t see the window, so you manage it by habit rather than by gauge.

The one exception is Claude Code on the desktop, which does show you how much of your context window you’ve used. That little circle in the bottom right of the image below, when clicked on, pops up your context window usage.

Here’s 6 rules on how to protect yourself:

1. One chat, one job.

Don’t run strategy, copywriting, competitor research and email drafts through a single thread. Mixed topics create noise that makes the model worse at all of them.

This is also how I split work across tools. I do the messy front end, the ideation, the planning, the pulling together of research, in Claude chat.

Then I move the finished assets into a folder and do the end-to-end build in Cowork, where the full one million-token Opus window has room to work across all of it at once. Thinking in one place, building in another.

2. Front-load and end-load.

Put your most important instructions at the start and restate the task at the end. The model attends most to both ends, least to the middle.

3. Give it everything up front.

A single, complete brief beats drip-feeding details across turns. That one change alone closes most of the 39% accuracy gap.

4. Convert PDFs to text before uploading.

It cuts the token cost by around 80% and leaves the window free for the actual work. A simple way to do it: paste the PDF into Claude chat, ask for the clean text back, save that as a markdown file, and use the lightweight markdown version in Cowork instead of the heavy original.

5. Reset with a summary.

After roughly fifteen to twenty exchanges, or the moment quality dips, ask the AI to summarise the key decisions, constraints and next steps, then paste that into a fresh chat. You keep the substance and drop the dead weight.

6. Store standing context in a Project.

Your brand voice, audience and rules belong in a Claude or ChatGPT Project, not re-pasted into every chat where they can scroll off the top.

Choosing The Right Model: Five Marketing Workflows

This is where it gets practical.

Model choice isn’t really about which one is “smartest.” It’s about matching the window to the job in front of you. Here’s how that plays out across the work marketers actually do.

1. Research, then content creation.

You’re building a long-form piece from a research brief, your brand guidelines and a couple of source documents, then drafting from all of it. You need everything visible at once. On a 16K or 32K window it runs out partway through and your early style guidance quietly drops off. This is a job for a big-window surface: Claude Sonnet 5, Cowork with Opus, Gemini on AI Pro, or the current GPT flagship. Give it the source material first, your question last.

2. Prospecting and competitor research.

You’re dropping in multiple competitor reports, landing pages and market research to find the gaps. This is the most window-hungry task there is, because each document is thousands of tokens and you want them all in view together. At 32K you’re comparing two or three at a time and deciding on partial evidence. At a million on Opus you can hold the whole set at once. Or maybe use Perplexity, which retrieves and cites live sources instead of making you paste everything in.

3. Cold email sequences.

A ten-email sequence built from your ICP, a messaging playbook and a few winning past emails is easily 20,000 to 40,000 tokens once it’s all loaded. Comfortable on most paid models. But run it as one long thread and by email seven the voice has drifted, because your original playbook is fading off the top of the board. Better: keep the playbook in a Cowork Project so it never scrolls away.

4. Working with spreadsheets.

A CSV export is deceptively heavy. A thousand-row spreadsheet can run anywhere from 50,000 to 150,000 tokens depending on how text-heavy it is, enough to blow a small window on its own before you’ve asked a single question. For real data work, reach for a genuine large-window model, and if the file is huge, feed it in chunks and ask it to quote the rows it’s using so you can catch any drift.

5. PowerPoint and long documents.

Analysing a full annual report (around 112,000 tokens on its own) or a large deck simply won’t fit inside a ChatGPT Plus window. Trying anyway means silent truncation and confident, wrong answers. This is Sonnet 5, Gemini or Cowork Opus territory.

The pattern underneath all five: small, one-off jobs work fine on anything, so don’t overthink those. But the moment you’re holding a lot of material at once, or working across a long session, the window becomes the deciding factor in whether the output is any good.

The Payoff: What You Get When You Understand Context Windows

Learning one invisible concept can sound abstract, so here’s what actually changes.

You stop losing your brief. The most common failure in AI marketing work, the brand guidelines you pasted in message one being ignored by message twelve, simply stops happening once you manage context deliberately.

You stop blaming the model. “The AI was brilliant last week and rubbish today” is nearly always a context issue. Diagnosing it correctly means you fix it in seconds instead of fighting the tool.

You get better output, for less. Leaner, more focused conversations don’t just cost fewer tokens and run faster, they produce sharper answers, because you’re not drowning the signal in noise.

You make better model choices. Knowing that a Plus user has around 32K to play with, not a million, changes how you structure a task before you even start, and tells you when to trust a long session and when to reset.

The Honest Caveats

A few things worth naming plainly.

The numbers move monthly. Context windows and model names change constantly. The figures here are current as I write, but check the model you’re actually using rather than trusting any table for long.

Memory softens this, but doesn’t remove it. Every major tool now has memory features that carry facts across chats. That helps between conversations. Inside a single long chat, the window limit still bites exactly as described.

There’s no perfect gauge. The “effective working window” rules of thumb, like staying in the first quarter or so of the window for important work, are sensible habits, not guarantees. Treat them as a nudge to reset, not a precise science.

The Bigger Picture

For most of the last two years, we obsessed over prompts, the perfect wording to squeeze a good answer out of a blank-slate AI.

The skill that actually separates strong AI marketers now is quieter than that. It’s managing context: what’s in the window, what’s fallen out, and when to start again.

Understanding the context window is what turns AI from a slot machine, sometimes brilliant, sometimes baffling, with no idea why, into a tool you can actually direct.

The marketers who struggle treat the chat box as infinite and wonder why it lets them down. The marketers who win know the whiteboard is finite. They can sense it filling even when the screen won’t show them, and they clear it before it clears itself.

You stop hoping the AI remembers. You start managing what it can see.

This week’s challenge: next time you’re deep in a long AI session and you feel the quality slipping, don’t push on and don’t re-paste your brief. Instead, ask the AI to summarise the key decisions, constraints and next steps so far, open a fresh chat, and paste that summary in as your starting point. Notice how much sharper the next answer is. Reply and tell me what you were working on, I’d love to know if it made the difference.

Thanks for reading Marketing with AI. If you found this useful, sharing it with one other marketer is the best compliment I could receive.

Need 1:1 Help to Embed Claude into Your Marketing?

Let’s talk →

Want a behind the scenes look at Claude for Marketers Online Masterclass?

Remember, you get lifetime access, along with full course updates and new modules being added until the end of 2026. New modules coming soon include Claude Code, Claude Design and How to Build an AI Marketing Team.

Substack Newsletter readers get 10% off this week only with code SUBSTACKCONTEXT