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.
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
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.
Read more: https://www.cnbc.com/2026/07/17/moonshot-ai-kimi-k3-model-openai-anthropic-china.html
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.
Read more: https://blog.google/innovation-and-ai/products/gemini-notebook/notebooklm-gemini-notebook/
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.
Read more: https://www.tryprofound.com/resources/white-papers/how-cmos-are-using-ai-to-change-marketing
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.
Read more: https://www.markifact.com/google-ads-mcp
📋 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.)
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.)
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.
🛠️ 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.
Read more: https://agilebrandguide.com/marketing-technology-ai-news-july-15-2026/
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.
Read more: https://martech.org/the-latest-ai-powered-martech-news-and-releases/
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.
Read more: https://martech.org/the-latest-ai-powered-martech-news-and-releases/
📊 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.
Read more: https://www.cmswire.com/customer-experience/how-cx-teams-are-actually-using-ai-in-2026/
💡 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.
Read more: https://martech.org/the-latest-ai-powered-martech-news-and-releases/
🔮 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.
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.
What’s the smallest portable context file you’d keep model-agnostic from day one?