Some of the most interesting GEO news items for me this week were that LLM bots have a word count ceiling, that schema might correlate with lower AI Overviews performance (if you're doing something for AI visibility, you should first be sure it's something that AI cares about!), the NotebookLM (or Gemini Notebook as it's now known) AI Overviews hack.
It was unsettling, but not surprising to see how much GPT 5.6 pricing information does not come directly from brand's own pricing pages, and that is also cites fewer in line citations and uses fewer sources. It wasn't surprising to see, again, that LLMs love tables!
Anyway,
Key Takeaways
- A peer-reviewed study accepted at ACM Web Conference 2026 (Allouah, Besbes, Figueroa, Kanoria, Kumar, arXiv:2508.02630) found that a "sponsored" tag lowers an AI shopping agent's probability of selecting a listing, a credibility penalty that does not apply to human shoppers.
- Meta's shopping assistant Muse never picked a single sponsored listing out of eight shown in a Walmart "AAA batteries" test, despite seeing them, according to analyst Juozas Kaziukenas of Marketplace Pulse.
- Amazon blocked Meta's Muse from accessing its shopping data entirely from 21 September 2026, days after Muse passed 500,000 users and 2 million-plus prompts in its first week (The Information).
- Hilton properties are now discoverable and bookable inside Google's AI Mode via natural-language queries, pulling in Hilton Honors points and redirecting to Hilton.com to complete the transaction (Hilton Newsroom, 28 August 2026).
- Google has begun auto-enabling native, in-answer checkout by default for merchants on Universal Commerce Protocol-connected platforms such as Shopify, inside AI Mode and Gemini, via Google Pay. It is opt-out, not opt-in.
- Since around 17 September 2026, Google's AI Overviews, AI Mode and Gemini have started citing public NotebookLM pages as sources, a technique not observed with other AI assistants (Malte Landwehr, Peec AI).
- Since GPT-5.6 shipped, more ChatGPT pricing answers pull from secondary sources such as press releases and blog posts rather than a brand's own pricing page (23%), with a further 3% now coming from competitors' pricing pages.
- An experiment hiding 15 fictional names across pages up to 705,216 words long, verified by server logs, found each AI assistant has a fixed absolute word-position reading ceiling: Meta AI and Grok read to the very end, Google AI Mode stopped around word 64,000, Claude around 16,000 to 64,000, Gemini around word 2,000, and ChatGPT declined the largest page outright (Andre Alpar, Search Engine World).
- ChatGPT citations are 2.3 times more likely to include a table than a Google Search result, and AI systems are three times more likely to cite pages with a visible last-updated date under three months old (Chris Long, Nectiv).
- Comscore's Q2 2026 AI Intelligence Report puts Google AI Overviews in 39.4% of US desktop searches in June 2026, up from 25.8% in July 2025, with Bing Copilot Search at 17.3%, up from 11.8%.
- For lodging brands, Comscore found being a source behind an AI answer and being visibly cited are very different things: one major travel brand was a source in 61% of responses but visibly cited in only 21%.
- Man of Many's 65-page free GEO guide found schema markup added across nearly 1,900 pages correlated with a 4.6% fall in AI Overview citations, and llms.txt showed no relationship to citation rates across roughly 300,000 domains.
- A SparkToro and Gumshoe study running 2,961 repeated identical prompts across ChatGPT, Claude and Google AI found less than a 1 in 100 chance that two runs return the same brand list in the same order.
- Comparing GPT-5.5 to the newer "Luna" model across 346 matched prompts, Seer Interactive found inline citations fell 12.9% and consulted sources dropped 43%, a change driven by narrower, smarter searching rather than any change in brand visibility (Bryan Gunawan, Wil Reynolds).
- The share of AI Overviews containing an external link inside the answer text itself jumped from near zero to over 26% in about ten days in mid-September 2026, according to Peec AI tracking (David Konitzny).
- YouTube's share of ChatGPT's cited sources fell 91% compared to August 2026, with Facebook down 88%, Wikipedia down 78% and Forbes down 72% (David Konitzny, Peec AI), a pattern independently corroborated by Otterly AI's September report showing YouTube down 42% in AI citations overall.
- A BrightonSEO analysis of 3.6 million-plus US ChatGPT shopping conversations found Reddit's share of shopping citations collapsed from 11.4% to 0.5% between June and September 2026, and confirmed ChatGPT truncates merchant descriptions at roughly 50 tokens, about 40 words (Metehan Yesilyurt, Peec AI).
- From 1 September 2026 Google auto-migrated broad-match and Automatically Created Assets campaigns onto AI Max for Search with no opt-in, mixing broader-matched traffic into any concurrent test or bid change (Lukas Beeler).
- Multiple agencies report ChatGPT Ads dashboard clicks with no matching GA4 data, with one study finding 60% of ChatGPT ad conversions land after the attribution window has already closed.
- Meta is limiting Facebook Pages without a paid Meta One for Business subscription to two link posts a month; Meta's own data shows only 1.3% of viewed posts currently include an external link.
- Google's fourth spam update of 2026 began rolling out on 24 September, alongside reporting that Google has deployed an internal AI system called SAFE, built from four coordinating agents, to catch AI-generated spam that violates the spirit of its policies.
- Google Search Console now separates "Text-based" from "Multimodal" search traffic in its Performance report, covering Google Lens, Circle to Search and image uploads. Google's John Mueller confirmed to Barry Schwartz this is genuinely new data, not a relabelled existing metric.
- Lighthouse 13.5 adds an audit for Agentic Resource Discovery (ARD), a cross-vendor spec from Google, Microsoft and Hugging Face for how AI agents locate a site's services, though the spec has already moved from ai-catalog.json to ard.json while Lighthouse has not yet caught up.
This has been one of the busiest fortnights I have covered in a while, and the throughline across nearly every story is the same: AI systems are no longer just answering questions about brands, they are increasingly transacting on their behalf, and the old ways of measuring whether any of this is working are starting to look shaky.
AI agents are now doing the actual shopping and booking
Start with the most consequential story of the fortnight. A peer-reviewed paper accepted at the ACM Web Conference 2026, "What Is Your AI Agent Buying? Evaluation, Biases, Model Dependence, and Emerging Implications for Agentic E-Commerce" (Allouah, Besbes, Figueroa, Kanoria and Kumar), found that a "sponsored" tag measurably lowers an AI shopping agent's probability of selecting a listing. Human shoppers do not apply anything like this penalty. It is a genuinely new kind of bias, and it matters enormously for anyone whose paid placements are a meaningful part of how they are found.
The real-world corroboration came quickly. Analyst Juozas Kaziukenas of Marketplace Pulse tested Meta's new shopping assistant, Muse, by asking it to choose from Walmart's "AAA batteries" results, which included eight sponsored listings, more than Amazon typically shows. Muse never picked a single one, despite clearly having seen them. Days later, Amazon blocked Muse from accessing its shopping data entirely, from 21 September 2026, specifically over AI shopping access. The timing is hard to read as coincidence. Muse itself is not a niche experiment either: The Information reported it passed 500,000 users and more than 2 million prompts in its first week alone.
At the same time, Hilton has quietly built one of the clearest real-world examples yet of agentic commerce in travel. Since a Hilton Newsroom announcement on 28 August 2026, Hilton properties have been discoverable and bookable directly inside Google's AI Mode through natural-language queries, pulling in Hilton Honors points and redirecting to Hilton.com to complete the booking. Hilton has also built a ChatGPT plugin for Q&A, has a Claude connector in development, and is rolling out its own generative concierge, Hilton AI Planner, across desktop and its Honors app.
And it is not just travel brands opting in. Google itself has started auto-enabling native, in-answer checkout by default for merchants on platforms that support the Universal Commerce Protocol, with Shopify specifically named. Eligible products are switched on for purchase directly inside AI Mode and Gemini via Google Pay, and merchants have to actively opt out if they want to keep purchases on their own site. Google's own messaging to merchants stresses they "retain full ownership" of customer relationships and data, but the default-on framing is the real story here. If your paid media strategy still assumes a human is doing the browsing and deciding, all four of these developments are worth sitting with.
A parasite GEO trick, and a shift in where pricing answers come from
Two smaller but genuinely actionable findings from Malte Landwehr at Peec AI. First, since around 17 September 2026, Google's AI Overviews, AI Mode and Gemini have started citing public NotebookLM pages as sources, a technique not yet seen with other AI assistants. It is a cheap, low-effort distribution channel worth testing while it remains distinctive to Google's own surfaces. Second, since GPT-5.6 shipped, more of ChatGPT's pricing-related answers are being pulled from secondary sources, press releases, blog posts and help articles, rather than directly from a brand's own pricing page, which now accounts for a shrinking share. A further 3% of pricing answers now come from competitors' pricing pages entirely. Owning the canonical page is clearly no longer sufficient on its own.
Every AI assistant has a hard reading ceiling, and it is a fixed word count
One of the more elegant experiments I have seen this year. Andre Alpar, writing on Search Engine World, hid 15 fictional names at set word-positions across a series of pages, the largest running to 705,216 words, then tested 14 AI assistants and confirmed exactly how far each one actually read using server logs (339MB of traffic across 182 fetches). The results varied enormously. Meta AI and Grok read every page in full. Google AI Mode stopped around word 64,000. Claude stopped somewhere between roughly 16,000 and 64,000 words. Gemini stopped at only around word 2,000. ChatGPT declined to process the largest page at all, hitting a content-size limit.
The key finding is that each model appears to have a fixed absolute word-position ceiling rather than a percentage of total page length, which has an immediate practical implication: front-load your most AI-extractable content within the first two thousand to sixty-four thousand words depending on which model you most care about being read by, rather than assuming a long page will simply be read proportionally less.
A companion piece from Chris Long at Nectiv updates Moz's classic "perfectly optimised page" concept for the AEO era, and pairs well with Alpar's finding. His six recommendations: allow AI bots such as GPTBot, ClaudeBot and PerplexityBot in robots.txt and make sure content is server-rendered; lead with four to six extractable "key takeaway" sentences near the top; use more tables and lists, since ChatGPT citations are 2.3 times more likely to include a table than a Google Search result; keep a visible last-updated or dateModified tag, since AI systems are three times more likely to cite pages under three months old; build FAQ sections around real customer-language questions rather than guessed ones; and optimise for the sub-queries AI models fan out from a single prompt, not just the original query itself.
A cautionary tale on entity SEO and Wikidata
Worth flagging for anyone tempted to treat Wikidata as a shortcut to entity recognition. One marketer, writing on her own blog, built an accurate Wikidata entry with proper structured data, was flagged for undisclosed paid contributions, fixed the issue, and then had the entire entry removed as "non-notable" with no appeal process available. The advice that followed from entity SEO specialist Grant Simmons was blunt: Wikidata generally requires awards or major recognition, and for a typical mid-market business it is simply not the right venue. His alternative, three-pillar approach is more useful: test what AI models already believe about your brand with web search disabled to find gaps and hallucinations, build canonical "entity hub" pages on your own site with proper JSON-LD schema rather than fighting external databases, and measure entity presence across depth, breadth and topic concentration rather than chasing a single external listing.
Two big data drops on the state of AI search visibility
Comscore's Q2 2026 AI Intelligence Report is the most comprehensive dataset I have seen on this topic all year. Google AI Overviews now appear in 39.4% of US desktop searches as of June 2026, up from 25.8% a year earlier, while Bing Copilot Search sits at 17.3%, up from 11.8%. Total US desktop search volume grew 8% over the same period, so this is genuine expansion of AI-mediated search rather than a zero-sum substitution. The most useful finding, though, is about the gap between being a source and being cited. For lodging brands specifically, Comscore found one major travel brand appeared as a source behind an AI answer in 61% of responses, but was visibly cited in only 21% of them. Being retrieved and being credited are two separate, and separately measurable, forms of visibility, and conflating them will give you a misleadingly rosy or gloomy picture depending on which one you happen to be tracking.
A widely praised, independently produced 65-page guide from Man of Many, built from around fifty named sources, adds real nuance rather than hype. Two findings stand out. Schema markup added across nearly 1,900 pages actually correlated with a 4.6% fall in AI Overview citations, the opposite of what most schema advocates would predict. And llms.txt, the proposed standard for signalling content to AI crawlers, showed no measurable relationship to citation rates at all across roughly 300,000 domains studied, corroborating an earlier HubSpot and Zyppy finding. Perhaps the most quotable line in the whole guide, from its co-founder: AI-referred traffic sits at well under one per cent of typical site sessions for most publishers studied, and "anyone selling you an AI traffic strategy is selling you a rounding error."
Why citation-count metrics are shakier than they look
This is the story I think deserves the most attention this fortnight, because it is a genuine corrective to a lot of the raw citation-tracking numbers that get shared, including some in this very newsletter. A SparkToro and Gumshoe study, running 2,961 repeated identical prompts across ChatGPT, Claude and Google AI with 600 volunteers, found less than a 1 in 100 chance that two runs of the exact same prompt return the same brand list in the same order. Rank position simply is not stable the way a search engine ranking position is. The study proposes "inclusion frequency," how often a brand appears at all across repeated runs, as a steadier alternative metric.
That finding landed in the same fortnight that a GEO measurement vendor's transparency was publicly challenged over undisclosed sampling methodology behind its brand rankings, with the critique that automation does not fix uncertainty in the measurement underneath it. And Wil Reynolds of Seer Interactive put the sharpest point on it. Comparing GPT-5.5 against the newer "Luna" model across 346 matched prompts, his team found inline citations fell 12.9% and the number of consulted sources dropped 43%, driven entirely by the newer model searching narrower and smarter with the site: operator, not by any actual change in brand visibility or content quality. His own words on this are worth repeating close to verbatim: citations are a bad KPI for generative engine optimisation, and it is fine to track them, but trending them out over time or looking at them in a vacuum is the issue. His practical advice is to look at what a "fan-out" browser extension actually shows happening behind the scenes, to recognise that domains are often retrieved and given a chance to influence an answer without ever being cited, and to use audience research to check whether your actual customers are on a platform before reactively abandoning it just because a raw citation count dropped.
AI Overviews suddenly started linking out again
Independent tracking from David Konitzny at Peec AI shows the share of AI Overviews containing an external link inside the answer text itself, not merely in a source card, jumped from near zero to over 26% in roughly ten days in mid-September 2026. This was a near-instant step change rather than a gradual drift, with the US, Australia and the UK leading the countries measured. In-text links matter because they sit exactly where a reader's eye already is, unlike source cards that most people scroll straight past.
This builds on a Google update from May 2026 that added more inline links, hover-preview cards and visually highlighted subscription-publication links, changes that arrived amid genuine regulatory pressure: an antitrust suit from a major publisher, a European Publishers Council complaint to the European Commission, a separate EU investigation, and independent research showing AI Overviews correlate with a 58% reduction in click-through to publisher sites. Google's own framing was about making it "easy to connect with authentic voices," but the more sceptical read is that this is a gesture toward publisher sustainability while the underlying trajectory, less traffic reaching the open web, remains largely unchanged. More visible links do not automatically mean more recovered clicks. It is worth watching closely rather than treating as a solved problem.
ChatGPT stopped citing YouTube on 18 September, and it is not the only platform swinging hard
Also from David Konitzny at Peec AI: YouTube's share of ChatGPT's cited sources fell 91% compared to August 2026, with Facebook down 88%, Wikipedia down 78% and Forbes down 72%. Crucially, this shows up in absolute citation volume too, not just as a dilution effect from a growing total citation pool, and the chart accompanying the finding is literally titled around the observation that ChatGPT stopped citing YouTube on 18 September. Konitzny is careful to stress that retrieved and cited are not the same thing: these domains have not vanished from ChatGPT's consideration process, they are simply making the final cut far less often.
A separate monthly report from Otterly AI corroborates the broader pattern using a different methodology entirely: social platforms are no longer moving together the way they did earlier in 2026. LinkedIn citations are up 39% across the AI engines tracked in September, while Reddit and YouTube citations moved in opposite directions depending on which specific engine you look at, with ChatGPT's Reddit citations down 94% from July levels while Google AI Overviews increased its Reddit citations over the same window. Two independent trackers converging on a real, sharp platform-level shift, using different methods, is a strong signal this is a genuine pattern rather than a measurement artefact of either individual tool.
Inside 3.6 million ChatGPT shopping conversations
A BrightonSEO presentation from Metehan Yesilyurt at Peec AI analysed 3.6 million-plus US ChatGPT shopping conversations and 23.7 million citations from shopping-card and product-gallery answers. The headline finding is that Reddit's share of shopping citations collapsed from 11.4% to just 0.5% between June and September 2026. The presentation also confirmed something genuinely useful for content structure: ChatGPT truncates merchant descriptions at roughly 50 tokens, about 40 words, so whatever you most want an AI shopping answer to notice about a product needs to sit in that opening line, not buried further down. Listicles remain the dominant citation format at over 45%, despite real growth in direct product and category page citations, which suggests both formats are worth continued investment rather than picking one over the other.
AI Max for Search: reporting is catching up, but so is a migration headache
Google's campaign-level Keywords report now shows "AI Max expanded matches" and "AI Max landing page matches" totals, which finally gives advertisers a workable way to calculate what share of conversions AI Max is actually driving, something that was previously opaque. Google is also expanding its AI Brief context feature, where advertisers give AI Max written business and audience context, to seven additional languages, and is building a unified reporting UI that will connect search term, creative shown and landing page reached for individual AI-driven interactions, though with no confirmed launch date yet beyond "later in 2026."
The complication is timing. From 1 September 2026, Google auto-migrated broad-match campaigns and Automatically Created Assets campaigns onto AI Max with no opt-in required, both with search-term matching switched on by default. Analytics lead Lukas Beeler's warning is the practical one worth heeding: this produces more but broader-matched traffic, which is a genuine measurement problem if any landing page test, bid change or messaging change happens to overlap the migration window. His advice is to check conversion quality, not just conversion volume, since broader matching can produce cheaper leads at the cost of lead quality. Separately, the previously planned migration of Dynamic Search Ads campaigns to AI Max has been pushed back to February 2027.
ChatGPT Ads attribution has some real gaps
Multiple agencies are independently reporting the same problem with ChatGPT Ads: dashboard clicks that never show up in GA4. One agency head cited 57 dashboard clicks against fewer than 20 matching GA4 visits for the same campaign. Another reported $1,000 of spend producing zero corresponding GA4 data at all. Part of the explanation appears to be technical: OpenAI's stated 30-day attribution cookie is reportedly being silently capped at 7 days by Safari's Intelligent Tracking Prevention, and one study found 60% of ChatGPT ad conversions occur after the attribution window has already closed. A separate study found 14.35% of ChatGPT ad placements were semantically unrelated to their surrounding conversation. OpenAI has itself acknowledged it currently lacks performance benchmarks across advertisers and industries. Worth a direct gut-check against your own GA4 numbers before trusting the platform's own reported totals.
Facebook is throttling free link posts
Following the launch of Meta One for Business, Facebook Pages without a paid subscription are now limited to two link posts a month, with Publisher Pages exempted from the restriction. Meta's own Q1 2026 data shows only 1.3% of viewed posts currently include an external link at all, and the company's stated rationale is testing whether higher link-post volume is worth paying for. For anyone relying on organic Facebook Page posts to drive affiliate or referral traffic, this is a real and immediate constraint on that channel, not a hypothetical one.
A fresh Google spam update, and the AI-spam detector likely behind it
Google's fourth spam update of 2026 began rolling out on 24 September, targeting general spam-policy violations globally, with Google itself saying the rollout may take up to two weeks to complete. Notably, it followed reporting that Google has deployed a new internal system called SAFE, the Scaled Abuse Forensics Examiner, built from four coordinating AI agents, a Root Agent, a Content Understanding Agent, a Behavior Understanding Agent and a Channel Cluster Understanding Agent, designed specifically to catch "spirit of policy" violations that would otherwise slip past traditional classifiers, including coordinated AI-generated spam networks. The research paper describing SAFE is short and withholds most hard performance numbers, so this should be read as plausible context for the recent run of spam updates rather than a confirmed direct cause.
Two useful technical SEO updates
Google Search Console now separates "Text-based" from "Multimodal" search traffic in its Performance report, with the multimodal category covering Google Lens, Circle to Search on Android, image uploads to Google Search and Chrome's right-click "Search this image" feature. This is worth taking seriously rather than treating as a minor filter addition: Barry Schwartz got direct confirmation from Google's John Mueller that this is genuinely new data that was not previously included in any existing count, not a relabelled slice of web-results traffic that was always there. It is the first time site owners get direct visibility into image-driven discovery as its own measurable segment.
Separately, Lighthouse 13.5 has added an audit for Agentic Resource Discovery, a proposed cross-vendor specification from Google, Microsoft and Hugging Face for how AI agents locate the tools and services a site offers. It checks for a catalog pointer via a robots.txt Agentmap directive, an HTML link tag, or an HTTP Link header, defaulting to a check for a well-known catalog file if none of those exist. It is grouped under a new "Agent Discoverability" category in Chrome DevTools and PageSpeed Insights alongside the existing llms.txt audit. One practical wrinkle worth flagging directly to any development team implementing this now: the specification has already moved its primary file location from ai-catalog.json to ard.json, but Lighthouse 13.5's current code has not yet caught up to that change.
Meme via CanIPhish