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.
Course Update:
The Complete Guide to Claude Design Now Included Inside Claude for Marketers Masterclass
(Pro/Elite Editions Only)
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.
Read more: https://searchengineland.com/google-ai-max-billions-new-monetizable-searches-483347
📋 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.
Read more:
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.
Read more:
🛠️ New Tools & Features
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.
Read more: https://webflow.com/updates/introducing-mcp-2-0
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.
Read more: https://www.anthropic.com/news/anthropic-economic-index-connector
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.
Read more: https://www.androidauthority.com/gemini-notebook-collections-feature-3689444/
📊 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.
Read more: https://searchengineland.com/topical-authority-ai-search-482875
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.
Read more: https://claude.com/blog/the-new-rules-of-context-engineering-for-claude-5-generation-models
(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.
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.