What changed in AI search this fortnight

Vendor moves

Google's fortnight was about paying, and being seen to pay, the web it reads. The AI contribution pilot is the headline: a Search Console feature, offered by invitation, that shows a monthly earnings number when a site's content contributes to AI-generated responses in AI Overviews, AI Mode, and Gemini. Alongside it, Google opened Search profiles wider on 16 September: the follower bar for claiming one dropped to 10,000 on a single platform (it was 100,000 in July and 35,000 after 12 August), articles got a new layout, one login can now run up to 10 profiles, and fresh documentation shows how to place a profile badge on your own site. The EEA aggregator and supplier units, which give approved comparison services and direct businesses their own block on the results page, extended to local business queries on 18 September. Gemini 3.8 Live and Extended Thinking launched the same week for voice.

OpenAI's build-out reached its logical point: the ad itself becomes an agent. Sponsored Agents, in US testing with a small group of advertisers, open a labelled chat where a user can state what they want, ask follow-up questions, and click through when ready. HubSpot is the first CRM connection, letting advertisers build campaigns and capture leads without leaving their CRM, and Shopify handles the commerce side. The same announcement carried natural-language campaign tools and the international rollout starting 23 September. This follows a pattern from the two reports behind this issue: ChatGPT for Financial Services and an Astra for Law product both launched as answer surfaces built for one industry at a time, and an engineering post disclosed that ChatGPT now serves more than a billion users every week. The population asking an AI for a recommendation, and the number of industry-specific ChatGPT products waiting to answer them, both keep growing.

Cloudflare's Disallow AI Training setting matters because it ends a choice that used to be all-or-nothing: block the mixed-use crawlers and lose search visibility along with training access, or accept both. Fewer than 1 percent of Cloudflare's customers currently block search crawlers, while 17 percent restrict AI training, so this setting gives a lot of sites the option they actually wanted. Apple, Google, and Microsoft all accepted a new Accountable crawler designation at launch. Separately, Schema.org shipped release 30.1 on 16 September with vocabulary for EU Digital Product Passports and a set of retail feed properties, including isOftenBoughtWith and itemPopularity. No answer engine reads these fields yet, but they are built for the same product-feed surface that ChatGPT Shopping now depends on.

Money followed attention. Profound's $180 million Series D, co-led by Sequoia and Kleiner Perkins, values the AI-visibility vendor at $1.8 billion on revenue it says tripled in six months across more than 1,000 enterprise customers. Mistral raised €3 billion earlier in the fortnight, a Samsung-led Series D at a post-money valuation above €21 billion, and followed it by announcing its models will run inside Firefox, a browser-level answer surface built on a sovereign-European pitch.

How LLMs read the web

The shopping numbers are worth sitting with. Free ChatGPT answers still draw on OpenAI's own index, and paid Thinking mode still leans on scraped Google results, but product queries now resolve against merchant feeds specifically: 8.26 percent of Profound's tracked recommendations were feed-based on 9 July, 61.54 percent the next day, and near 65 percent by 3 September. That is a switch flipped overnight, not a slow drift, and it means on-page copy cannot decide a shopping-shaped answer. Feed completeness and accuracy do.

Mueller's Search Console admission closes a question worth being direct about. The position dimension in the AI reports records where the AI feature sits on the page, not whether your link is first or fifteenth inside the answer. Mueller's defence, that answers get assembled per query and a stable link position may not exist, is honest, and it also means the only first-party Google data available cannot answer the question most clients actually ask. Third-party scan data, sampling limits and all, remains the only way to see within-answer position.

Google's own ad data shows what a population that types conversationally into a search box looks like. Jason Tabeling's numbers through August 2026 show impression share for one- and two-word queries falling from 42 percent to 24 percent since January 2025, while three- and four-word queries rose from 33 percent to 48 percent, and conversion share for five-plus word queries roughly tripled. The buyer question that used to be two words is now a full sentence, and Google's own logs confirm where the conversions moved.

Not every research claim from the fortnight holds up the same way, and the gaps matter as much as the numbers. A single-source study from Sales.co reports that 87 percent of 26,000-plus ChatGPT citations point to pages ranking in Bing's top 20 for the query, which sits in tension with RESONEO's measurement, the strongest evidence this newsletter has seen, that paid ChatGPT Thinking mode draws 75.3 percent of its retrieval from scraped Google results. Bing and Google overlap on plenty of queries, so both studies could be reading different slices of the same market, but the retrieval-source question is not settled yet, and any single figure deserves a second source before it goes in a deck. BuzzStream adds a distinction worth keeping separate: only about 1 in 4 AI brand mentions comes with a citation attached (vendor-sourced), so mention tracking and citation tracking are two different jobs.

One free check is worth running this week. Paste a distinctive sentence from your own page into a chatbot and ask where it came from. If the URL comes back, retrieval is not what is holding you back, something else is. Two practitioners published tools pointed at the same question from the crawl side: Glenn Gabe built a bookmarklet that exposes ChatGPT's actual query fan-outs per prompt, and Lily Ray described how AI Overviews assemble passages from across the live web index with sources attached for long-tail questions.

Community signal

Cyrus Shepard's ranking-factors survey, run through his Zyppy newsletter, published its AI section this fortnight: more than 130 practitioners rated brand recognition and trust, and what a model already knows about a company from training, above every AI-specific tactic on the list. llms.txt scored 2.0 out of 10. Near-consensus formed around one idea: original data and citable facts beat generic content, in an AI answer just as much as in a search result. Aleyda Solis's read of the same data made the same point in fewer words: chunking and llms.txt sit near the bottom, brand and trust sit at the top.

Glenn Gabe made the sharpest point of the fortnight about what AEO actually is. AI systems judge a brand across a source universe of PR coverage, social posts, and reviews that reaches well past the company's own domain, so a plan built only around the company website is judging the wrong surface. Marie Haynes kept her running list of Google AI changes updated with more international AI Overviews rollouts, local Knowledge Panels gaining AI Overviews, and ChatGPT stock monitoring, none independently confirmed yet but each worth a note.

The practitioner video crowd is where the specific tactics show up. One explainer drew a distinction worth repeating to anyone on your team who still treats the two as the same thing: a ranked link and a cited source are different wins, and a page can rank well while barely appearing in AI answers. A ten-years-in practitioner opened with a warning rather than a tip: running all of your content production through AI, with no human step, is the fastest way to lose visibility. A roundup from another channel collected tactics that are working right now: answer-first intros written to stand alone outside their page, a fixed ten-question buyer scoring routine, and direct targeting of best-of list pages, where a lot of AI citations trace back to.

A smaller, stranger note from the same two weeks: an account run by Grok itself answered AEO questions inside the open community discussion twice running, machine commentary about how to be visible to other machines. Worth remembering when weighing an opinion post about AI visibility: check who, or what, wrote it.

What it means for your brand

Check Search Console for the AI contribution panel. The pilot is invitation-only and aimed at smaller publishers right now, so most brand sites will not see it yet, but it costs minutes to look, and if AI citation turns into a real revenue line, you want to notice on day one rather than month six.

If you sell anything, treat your Merchant Center feed as an AI visibility asset, not just a shopping-ads input. Audit completeness, accuracy, and freshness the same way Profound and Lily Ray are watching it. For product queries, this now matters more than the product page copy.

Look at your crawler settings wherever you sit behind Cloudflare or a similar service. If you are currently blocking AI crawlers wholesale to keep your content out of model training, check whether that same rule is also blocking the retrieval that keeps you visible in answers. Split the two decisions on purpose instead of by default.

Change how you talk about the number in client and leadership reporting. Search Console cannot tell anyone where a brand sits inside an AI answer, only that it showed up somewhere in the block. Say that limit out loud before someone builds a target around a number the data cannot support.

Check whether you clear the 10,000-follower bar for a Search profile on any single platform, and claim it if so. It is a cheap way to get a followable presence inside Google search results and a badge for your own site.

Use the survey data the next time a vendor pitches an llms.txt package or a chunking service. More than 130 practitioners rank brand recognition, trust, and original data above the tactic menu, and llms.txt scoring 2.0 out of 10 is a number you can put in front of a client. Spend the budget on content and reputation work first.