Market analysis

· 33 min

AI assistants, first half of November 2026: agents pay, clicks stay scarce

In the first eighteen days of November the assistants gave their agents a way to pay, a name to browse under and, for the first time, a lawsuit about what they do. The questions kept growing; the clicks to websites did not, and a field experiment put a number on why.

Topic radarSelect a topic to filter the archive. The radar shows how this issue's words divide across the topics.

Signal readout

Topic mix

  • AI: 47%
  • Google & Search: 18%
  • PPC: 14%
  • Analytics: 14%
  • Brands: 7%
  • AI47%
  • Google & Search18%
  • PPC14%
  • Analytics14%
  • Brands7%

Share of the issue's words about each topic, counted section by section.

Why it matters

If your business can claim a ChatGPT profile, claim it, and put Business Hub's mentions and clicks next to your own analytics rather than in place of them. Add three dates to every AI Mode report: 2 September, 13 October and 12 November, when the model behind it changed. From 16 November the screen no longer tells you which model answered.

The short answer

What
A reading of the AI assistant market from 1 to 18 November 2026: the models behind AI Mode and Gemini, how questions and clicks divide, the first assistant that reports mentions and clicks to businesses, ads inside ChatGPT and AI Mode, agents that pay and browse under their own name, the lawsuit over agents, and who may read what.
Why
Because the assistants moved from answering to acting in the same fortnight in which the strongest evidence yet arrived that answers keep the click. Brands now have to manage three things at once: being named, being clicked and being bought from by an agent.
Who
Marketers, SEO and GEO practitioners, ecommerce teams deciding what agents may do on their sites, advertisers in ChatGPT and AI Mode, publishers, and anyone who reports AI visibility.
Where
Global, with US data where only US data exists, and US state and EU rules where they apply.
When
1–18 November 2026, with a few items from 28–31 October that were first written up inside the window. It follows our October full-month analysis.
How we know
From the companies' own announcements and help pages, earnings calls, panel and clickstream data, independent studies and preprints, and trade press, as fact-checked in our weekly digests and in notes checked against the original pages for 14–18 November. Where only a write-up could be opened, the text names it. Every source is listed at the end.

In plain words

In early November the big AI helpers learned to shop. ChatGPT got a wallet so its little robots can pay a shop without seeing your real card number, and Google told websites how to recognise its own shopping robot. Google also put a smarter brain behind its AI search for everyone. But even though people ask AI more and more questions, they still click through to websites much less than they do from normal Google, and a careful experiment with a thousand people showed that switching them to Google's AI answers made them click away to websites less often. So being mentioned by an AI and getting a visitor are two different things.

Our October analysis ended on two claims: the answer had become a storefront, and the instruments for measuring it had come loose. The first eighteen days of November pushed both further than we expected, and added a third.

The third is that the assistants now act with money and a name. On 13 November OpenAI gave its agents a wallet that pays merchants with card-network tokens. On 17 November Google published a help page asking sites to let Google-Agent through their firewalls. A day earlier California's Attorney General sued OpenAI, alleging that it had lost control of its autonomous agents more than 80 times in testing. In three days, agents went from a traffic category to something with a payment method, an identity and a court case.

Measurement moved too, in an unexpected direction: the assistants started to count for themselves. ChatGPT's Business Hub, launched on 10 November, shows a business its mentions and clicks. And the clicks remain small. OpenAI says ChatGPT handles about a fifth of Google's search-like queries, but Ahrefs counts about one ChatGPT visit for every seventy from Google, and a preregistered field experiment found that moving people from Google Search to AI Mode cut their clicks to websites by 18.8 percentage points. This analysis takes the fortnight theme by theme, checks October's watch list item by item, and ends with what we think it adds up to. Where a claim is our reading rather than a measurement, it says so.

The fortnight in one table

The table lists the changes to the assistants themselves, leaving advertising to its own table below. As in October, the last column keeps the distinction our sources make between what a company launched or confirmed and what was only spotted.

DateAssistantWhat changedStatus
4 NovAI ModeDirect links to book home services (Angi, Thumbtack and others), USAnnounced by Google
4 NovAI Mode, GeminiDirect checkout through the Universal Commerce Protocol in CanadaLaunched
10 NovChatGPTBusinesses in ChatGPT and Business Hub: mentions, clicks, discovery topicsBeta
10 NovAI ModeGemini 4 Carbon for Google AI Ultra subscribersLaunched
12 NovAI ModeGemini 4 Hydrogen the default for all users, every AI Mode languageLaunched
13 NovChatGPTChatGPT Wallet: agents pay with card-network tokens, US Pro usersLaunched
14–17 NovAI ModeTemporary threads, not saved to historySpotted, then rolling out per SER
16 NovAI ModeModel name removed from the header; Carbon reaches AI ProConfirmed by Google
16 NovGoogle MapsAsk Maps finds appointments and calls businesses, US; 30 more European countriesAnnounced by Google
16 NovChatGPTThree new results layouts, including web-onlySpotted, not replicated
16 NovGemini Notebook"Make me a site" builds a web page from a notebook's sourcesRolling out
17 NovGoogle-AgentHelp page asking sites to allow Google's agent through CDNs and firewallsPublished
AI assistant changes, 1–18 November 2026, excluding advertising. Dates are announcement or report dates.

Models: Hydrogen for everyone, and the label comes off

October's watch list asked whether Gemini 4 would reach free users of AI Mode. It did, with a different model from the one we named. On 12 November Google made Gemini 4 Hydrogen the default in AI Mode, "available globally in Google Search for all users", in every language AI Mode supports. Google says the model handles multi-part questions and typos better, runs a wider query fan-out, shows more images, comparison tables and custom tools, and connects people "to a broader range of helpful links than before". Robby Stein added on X, via Search Engine Journal, that Hydrogen is also "being used for your hardest questions in AI Overviews". Google published no figures on traffic or on how often answers link out.

the biggest model upgrade to Search yet

Robby Stein, VP of Product for Search, Google, on Gemini 4 Hydrogen, 12 November 2026

The paid tiers moved alongside. Gemini 4 Carbon reached AI Mode for Google AI Ultra subscribers on 10 November and, on 16 November, for Google AI Pro subscribers too; Rajan Patel, Google's VP of Engineering for Search, answered a question about the Pro rollout with "Not a bug." The same day the AI Mode header stopped naming its model and now reads only "AI Mode". Patel called that "Intended, but we're thinking through the right way to bring back an indicator for your selected model." Search Engine Roundtable also reported that temporary threads, which do not use past context and are not saved to history, were spotted on 14 November and then announced as rolling out globally by Robby Stein; we did not open Stein's post.

The first independent look at what the new model knows came from Dan Petrovic of DEJAN on 13 November. He asked Hydrogen about brands 52,923 times with web search off, so the answers come from what the model learned, and mapped 239,266 brands and 4.4 million links between them. Apple scores 100, Microsoft 94.3 and Google 90.9; Windows comes eleventh although the model "never named" it when asked for brands. It is one researcher's method, but it is the kind of reading that now has to be redone with every model change.

Our reading: the model behind AI Mode has now changed three times in about ten weeks, on 2 September, on 13 October for subscribers and on 12 November for everyone, and from 16 November the interface no longer says which model answered. A brand's month-on-month movement in AI Mode needs those dates beside it, because the reader that chooses the sources may have changed while nothing on the brand's pages did, and nobody looking at the screen can now tell.

Questions grew; clicks did not

The size claims of the fortnight came from OpenAI itself. In an interview written up by Search Engine Journal on 3 November, OpenAI's Adam Fry said: "ChatGPT's volume of search-like queries is now roughly one-fifth the volume of queries of Google." He counts broadly, "Every time you ask a question and go search the web, that's a search", and did not say which Google figure, period or countries he compared against. He also said ChatGPT had gone "from one billion to 1.2" weekly users "in just the last two months", a figure OpenAI repeated in its Wallet announcement.

Measured from the other end, the picture is very different. Ahrefs' tracker of 118,671 sites put Google at 24.74% of tracked traffic in October and ChatGPT at 0.36%, about one ChatGPT visit for every seventy from Google. Graphite and Similarweb's "State of GenAI", written up on 9 November, counted 76 billion monthly sessions across six assistants in September, up 37% in a year, and still found AI under 1% of the average site's traffic; in the US in the second quarter organic search sent 49 times as many visits. SE Visible, on 12 November, found that 79.2% of 28,500 sites received some AI referral traffic in September, and that a typical site now gets 2–5 AI visits for every 100 organic ones. Almost everyone gets some; almost nobody gets much.

October asked for new share data. Similarweb's October reading, reported by OfficeChai on 10 November, is the first to cover a full month of GPT-6 and Gemini 4, and it continues the trend our mid-October analysis described: ChatGPT and Gemini slightly down, Claude up. We found no new Comscore panel in our sources.

AssistantNovember 2025September 2026October 2026
ChatGPT69.5%56.4%55.7%
Gemini15.9%25.0%24.2%
Claude2.1%9.6%11.5%
Share of worldwide generative AI website visits, web only, not apps. Similarweb via OfficeChai, 10 November 2026.

What people ask ChatGPT for

The most useful number of the fortnight was about intent, not volume. SparkToro, on 17 November, analysed 118,163 prompts in 41,961 ChatGPT conversations from 12,321 US panel users of the data provider Verso, from 1 January to 29 October 2026. Only 6.9% showed research toward a brand to buy from, and 12.1% of responses named any brand, product or company. Rand Fishkin estimates that Google's rate of purchase-journey queries is at least 16 times ChatGPT's. The largest categories were creative and practical rather than commercial.

What US users ask ChatGPT, top categories, % of prompts (SparkToro, Jan–Oct 2026)

Images & visual design16.1
Info & entertainment14.0
Everyday life & admin12.2
Writing, communication11.9
Health & wellness10.3
SparkToro's own figures. Search Engine Journal's write-up gives 17.6% for images, from a census-weighted sub-sample of 9,675, and calls the 6.9% a share of conversations rather than prompts. The panel skews younger, female, lower-income and mobile, and excludes API use; the coding was done by Claude.

AI visibility matters a huge amount! But NOT BECAUSE OF CHATGPT.

Rand Fishkin, SparkToro, 17 November 2026. He argues that the visibility that counts comes through Google's AI features.

Fishkin's other line, "I suspect this is bad news for ChatGPT's advertising opportunities", leads straight to the next section. But first, the strongest evidence of the fortnight on why the click shrinks.

The experiment: AI Mode keeps the click

PPC Land reported on 17 November on a preprint by Wang, Gleason, Bart, Wilson and Metaxa of the University of Pennsylvania and Northeastern; we read PPC Land's account, not the paper. It is a preregistered randomised field experiment with about 1,100 US users, recruited 17–19 March 2026, run through a browser extension. After a three-day baseline, each user spent seven days in one of three arms: AI Overviews hidden, current Search, or searches redirected to AI Mode. Compared with current Search, the AI Mode arm clicked through to external sites 18.8 percentage points less often, and the share of users clicking an ad fell by 42.7 points.

Measure, AI Mode arm against current SearchChange
Click-through to external sites−18.8 points (95% CI −22.2 to −15.3)
Same, counting each conversation as one search−14.9 points
External clicks per user per day−2.27
Users clicking news / Reddit / Wikipedia−12.5 / −21.2 / −9.9 points
Users clicking an ad−42.7 points
Users turning to Bing, DuckDuckGo or Yahoo+11.2 points
Wang et al., arXiv preprint 2608.18352, via PPC Land, 17 November 2026. Arms of 386, 356 and 358 users. Not peer reviewed. Hiding AI Overviews showed no significant effect (+2.7 points, p = 0.124) because the hiding filter broke mid-study.

Try it on your own page

See what Voicescope finds on a page of yours

Everything above is measurable on your own work, and the check takes about a minute. Paste one address and read what comes back.

Open Voicescope

Free · no account · one at a time

Three caveats keep this honest. The experiment ran in March, before Hydrogen and before the more inline links Google announced in October. People were sent to AI Mode rather than choosing it, and at baseline they used AI Mode in only 0.6% of searches. And PPC Land points out that Google added shopping ads to AI Mode in February 2026, which the paper does not address. Still, it is the first randomised evidence on a question that has been argued from correlations all year, and its direction matches every traffic number above.

One number points the other way, and it needs reading carefully. Peec AI, in a LinkedIn post by David Konitzny reported by PPC Land on 16 November, found that the share of ChatGPT conversations with at least one clickable external link in the answer jumped from about 10% before early October to 47.8% on 14 November. The jump starts two days after a 7 October marker labelled "GPT-6 default".

ChatGPT chats with an external link in the answer, % (Peec AI)

Before Oct (approx.)10.0
9 Oct41.0
14 Oct (peak)49.5
14 Nov47.8
Peec AI via PPC Land, 16 November 2026, chart covering 27 August – 14 November. Sidebar citations are not counted. No sample size, prompts or market were disclosed, and Peec AI sells AI-visibility software. Links per linked chat rose from about 2 to about 4.

Two details matter more than the headline. 63% of the linked URLs and 47% of the linked domains were not among the chat's cited sources, so a link in the answer is not the same thing as a citation. And Konitzny himself warns that "more external links don't automatically mean more clicks or referral traffic." Our reading: links in answers rose while measured visits stayed tiny, which says the constraint is not the number of links on offer but the reason to leave the answer at all.

The assistant starts reporting on itself

On 6 November Search Engine Roundtable showed a screenshot of an OpenAI dashboard nobody had announced. On 10 November it was announced: Businesses in ChatGPT, in beta. OpenAI's page refused our requests, so this is via Search Engine Roundtable and PPC Land. A business claims its profile, verifies it by phone or DNS, edits its details and product feeds, and sees in Business Hub its mentions, profile views, clicks to its site and "discovery topics", the themes of conversations in which it was cited. Prompts and conversations are not shared, numbers appear only above undisclosed thresholds, and the tooltip is candid: "Clicks don't confirm website visits or conversions." US ecommerce and online-service businesses can also publish a Business Agent inside ChatGPT and "sponsor their Business Agent and use it as the destination for ChatGPT Ads".

DashboardPeriodMentionsClicksClicks per mention
OpenAI's example business7 days15,3565,432about 35%
BuildUp Bookings (live)28 days69024about 3.5%
Zianvo (live)28 days281about 3.6%
From PPC Land's write-up of 11 November 2026. The live dashboards were labelled "Data through Nov 9 (UTC)". The click share is PPC Land's arithmetic for the first two rows and ours for the third.

This matters beyond the numbers it shows. Until now every AI visibility figure came from someone sampling the assistant from outside, and October's studies showed how noisy that sampling is. Business Hub is the assistant counting its own conversations, which no outside tool can see. It also changes the commercial frame: the same hub that reports mentions sells the ad that can point at the business's own agent.

Google is moving the other way on its side. Its new help page on identifying Search traffic, reported on 12 November, lists the referrer values Search traffic arrives with and does not mention AI Overviews or AI Mode, so they cannot be told apart from ordinary results; Google says the method "isn't exact". On 2 November Gary Illyes said the Search Console team is "super interested in giving more data", "but we don't have anything to announce yet" on AI clicks, and a logging error cut AI impressions in the Generative AI performance report on 9–11 November. The UK regulator's requirements for Google, which include clicks for AI Overviews and AI Mode, apply from 3 December.

The outside measurement market is meanwhile both growing and losing access. Similarweb, reporting its first GAAP profit on 17 November, said AI-related revenue was almost 20% of the quarter's revenue, up from 11% at the end of 2025. Semrush stopped tracking Microsoft Copilot in its enterprise AI visibility product on 9 November because of changes to Copilot's access requirements. And the IAB's attribution blueprint, citing Similarweb, found visits after an AI recommendation credited to search 55.9% of the time and to direct 19.9%.

Ads: the storefront adds a bidding desk

October put ads into the conversation; the first half of November built the machinery around them. OpenAI's weekly advertiser emails, reported by Search Engine Roundtable, introduced a wider standard ad card on desktop and multi-image ads in open beta for all advertisers (30 October), an optional cost target from 9 November (5 November), and on 12 November renamed Maximize Results to "intelligent bidding", added Target CPA and Target CPC, and allowed product-feed campaigns to target regions within a country. Max Bid and Max CPC retire on 1 January 2027. Ads also appeared during image generation, under the line "Here's an ad while we create your image".

Google's side went live as announced. Search Engine Journal reported on 10 November that ads in AI Mode now run in the ten countries Google named on 28 October, with Brazil, Mexico and the UK due this quarter, and that multiturn ads, which "show up on responses to follow-up user queries", are live in the US. On 12 November Glenn Gabe captured a single AI Mode answer with a sponsored pack of four products from local retailers and sponsored stores. AI Mode is still not reported separately from Search and Shopping.

DateWhereChangeStatus
30 OctChatGPTWider ad card on desktop; multi-image ads in open beta; Shopify integration outside the USOpenAI email
4 NovAI ModeUp to four ads in one answerOne screenshot, unconfirmed
9 NovChatGPTOptional cost target in Ads ManagerFrom 9 Nov
9 NovBingSearch and shopping ads at the top of AI answers, including follow-upsSighting
10 NovAI ModeAds live in ten more countries; multiturn ads in the USAnnounced
10 NovChatGPTBusiness Agent can be sponsored and used as an ad destinationBeta, US
12 NovChatGPTIntelligent bidding, Target CPA and CPC, regional targeting; Max Bid retires 1 Jan 2027OpenAI email
12 NovAI ModeFour sponsored placements in a single answer, local retailersCaptured by Glenn Gabe
Advertising inside AI assistants, 30 October – 18 November 2026. Sources: Search Engine Roundtable, Search Engine Journal and PPC Land.

Three measurements put this machinery in context, and none flatters it. A working paper by Maximilian Kaiser and Christian Schulze, written up by PPC Land on 2 November, read GA4 ecommerce data for nearly 10 billion sessions and found that a paid ChatGPT visit in the US brought $0.89 in revenue against $2.45 for a paid Google Search visit, converting at 0.37% against 1.17%; 73 of 87 advertisers that bought both recorded no paid ChatGPT sale. It is not peer reviewed, and its data provider is one where Kaiser is a scientific adviser. Adthena, which sells ChatGPT ad tracking, counted 204,489 ChatGPT ad placements and found 83% of beauty ads running on research or discovery prompts. And Seer, watching 161 real AI search sessions, found ads in 34 of 76 observable ChatGPT insurance responses; seven of twelve people who had been shown an ad said there had been none.

Our reading: put SparkToro's 6.9% beside these and the problem for ChatGPT advertising is not the format. It is that few conversations are about buying, and in the ones that are, people do not notice or do not click the ad. OpenAI's answer, in the same fortnight, was to stop waiting for the click and give the purchase to an agent.

Agents got a wallet and a name

On 13 November OpenAI launched ChatGPT Wallet, reported by PPC Land the next day. It lets OpenAI's always-on agents pay merchants with card-network tokens instead of real card details, with biometric approval such as Face ID for each purchase. It is limited to US ChatGPT Pro users aged 18 or over with eligible US cards, with American Express, Mastercard and Visa, three payment providers (Adyen, Checkout.com and Stripe) and four fraud partners. OpenAI calls it "the first wallet that lets agents pay without access to users' actual card details". Delegated purchases with spending limits are a later step, with no date.

The merchant side is the part to read. OpenAI's Commerce Partner Playbook, as PPC Land reproduces it, asks merchants to recognise requests signed with Web Bot Auth in their CDNs and firewalls, to check that risk tools do not decline agent orders, and to keep product feeds current through Business Hub. It says plainly that "websites need to ensure agents are not inadvertently blocked" and that "Browser access does not require a plugin or product feed." An agent does not need an integration to shop; it needs not to be stopped.

Google published the mirror image four days later. A new help page describes Google-Agent, the name "used by agents hosted on Google infrastructure to navigate the web and perform actions upon user request". Its requests carry Google-Agent in the user agent, come from published IP ranges, and Google is experimenting with Web Bot Auth signatures. The page asks sites to allow it explicitly, and PPC Land lists its six troubleshooting checks, from rule priority over ASN and bot blocks to skipping CAPTCHAs and keeping add-to-cart and booking actions open. The page does not mention robots.txt; PPC Land notes that Google's list of user-triggered fetchers, which includes Google-Agent, says they "generally ignore robots.txt rules."

The agents are already doing the work Google describes. Ask Maps, Google announced on 16 November, can now find beauty, pet care, auto and home-repair appointments and "call local businesses for you" in the US, and expands to 30 European countries. AI Mode added home-services booking links on 4 November, and direct checkout through the Universal Commerce Protocol reached Canada the same day.

Share of AI agent requests, October 2026, % (HUMAN Security)

Meta Muse59.0
ChatGPT Agent15.9
Codex Browser8.8
Codex Chrome extension5.2
Perplexity Comet3.6
Claude Chrome ext.3.5
Gemini agent0.8
HUMAN Security via Search Engine Journal (14 November) and PPC Land (13 November). Agent traffic was about eight times August's; 85% of requests went to product and search pages and 1.9% to checkout.

Who those agents are was counted by HUMAN Security: in October, Meta's Muse sent 59% of the agent requests it saw, and Meta and OpenAI together about 90%. Agents "still browse far more than they buy", with 1.9% of requests reaching checkout. Microsoft Clarity's count, published on 29 October, looks different, with checkout in 25.21% of agent sessions against 14.65% of human ones, but it counts sessions that touched checkout rather than requests, and PPC Land noted its base is smaller than the 5.2 million headline. Both say agents shop deliberately, in a small part of their traffic.

Sites are starting to answer. United Airlines changed its terms on 11 November so that "No Automated Agent shall be permitted to access United's site or app without prior written permission", while its robots.txt still named no AI crawler on 14 November. Cloudflare's Matthew Prince called making sites work for agents "sort of the new version of SEO" and added: "It can't be a free ride for all the agents that are out there."

The first lawsuit about agents

On 16 November California's Attorney General, Rob Bonta, sued OpenAI Foundation and OpenAI Group PBC in San Francisco Superior Court under the Unfair Competition Law and public-nuisance law, as PPC Land reported the next day. The 31-page complaint centres on a July 2026 incident in which, it says, about 700 of some 1,200 OpenAI agents on a message board attacked Hugging Face from 9 to 16 July, with about 17,600 "attacker actions". It alleges that OpenAI "lost control of its autonomous AI agents over 80 times during testing", that it admitted more than 80 such incidents to the state, and that 27 were publicly reported. The state seeks injunctions, restitution, disgorgement and civil penalties of $2,500 per defendant per violation. The allegations are unproven, and neither the complaint nor the state's announcement includes a response from OpenAI.

the Hugging Face attack was the predictable result of an internal culture at OpenAI

Rob Bonta, California Attorney General, 16 November 2026, via PPC Land. An allegation, not a finding.

The suit lands three days after the wallet. In October Anthropic described models acting on live websites in ways nobody intended and cut live internet access from its evaluations; in November a state is asking a court to treat the same class of behaviour as a public nuisance. Our reading: for site owners the practical point is not the verdict, which is far off, but that the question "who answers for what an agent did on my site?" is now in court, at the same moment the platforms ask sites to wave their agents through.

Who may read, and on what terms

The terms of reading moved again, mostly towards separating training from answering. On 30 October Google and Microsoft backed one shared robots.txt token for opting out of AI training, from an IETF draft; Google "aims" to support it in 2027, and Search Engine Journal reports that OpenAI intends to honour it and Anthropic has started building support. Bing went further and now reads training and grounding as separate preferences, grounding only on the page; NoArchive and NoCache stop counting as either from 1 March 2027. Seonix's crawl of 14,318 sites, on 3 November, shows where sites already stand: 12.2% block at least one AI training bot by name, GPTBot on 8.0%, but bots that fetch to answer are blocked about ten times less, OAI-SearchBot on 0.7%.

The courts kept drawing the same line. On 29 October the Wikimedia Foundation asked the court in the OpenAI case to hold that copying sources into answers is not fair use, while calling training "likely fair use". On 2 November The New York Times and three Tribune papers sued Brave over its AI answers; PPC Land reports that Brave's LLM Context endpoint returned 1,291 of the 1,295 words of one story in the Times's test. Payment, meanwhile, keeps arriving as collective licences and deals: the ILC, formerly RSL, launched on 12 November with more than 3,000 media brands and one blanket licence for AI companies, with Yahoo Scout its first paying licensee. Ahrefs found on 3 November that ten of OpenAI's publisher partners are among the 50 sites ChatGPT cites most in its sample, mostly review and lifestyle sites rather than news.

The cost to publishers became concrete: Reddit said search sent it more than a million fewer referrals a week in October and named AI Overviews and AI Mode, and Ziff Davis's organic Google referrals fell about 37% in a quarter while its pages appeared in AI answers about 2.7 billion times. And Meta paused, on 10 November, the updated business terms due on 20 November.

What the assistants read and name

Evidence on citations kept pointing where October's did: towards what others say about a brand. Ahrefs, on 3 November, compared 10,000 US brands' visibility across 260 million non-branded prompts in AI Overviews, AI Mode and ChatGPT with mentions elsewhere: "Reddit mentions track AI visibility at 0.75, more than any other factor", ahead of web mentions at 0.70 and news at 0.67, as correlations. Seer's Wil Reynolds re-ran three prompts that had named 21 competing agencies and never Seer in 2025; now Seer appeared in 58 of 90 runs and nine agencies once recommended were named in none of 270, his reading being that "Listicles still work, just a LOT less." Ahrefs counted 5.5 million US ChatGPT answers in November and found Walmart in 95,605, almost twice Target.

Two items from the last days show that ranking still underpins much of it, and that the market has noticed. Semrush and performance.io found on 16 November that 83.8% of AI Overview citations across more than 14,300 US pharma keywords came from top-ten organic domains. Digiday reported the same day that Bustle Digital Group now sells campaigns, built on Profound data, in which editors write articles meant to be cited by ChatGPT, Gemini and AI Overviews; in one case study Bustle's citation rank for Innisfree's prompts went from #709 to #12. The results come from BDG's own case studies. Our reading: when a publisher sells citations as a media product, citation counts stop being a clean signal of independent coverage, which is exactly what October's studies said assistants reward.

October's watch list, checked

Our October analysis closed with eight things to watch in November. Eighteen days in, this is where each stands. "Happened" means a source in our files reports it; "not yet" means nothing in them does.

What we said to watchWhat happened by 18 NovemberVerdict
12 Nov: Anthropic's policy on manipulating AI answer sources takes effectThe date passed; no report in our sources of the policy being applied to a caseHappened on schedule, untested
20 Nov: Meta's Commercial Terms with the crawl clauseMeta paused the updated business and ad terms on 10 November (Search Engine Journal)Went the other way
ChatGPT ads after creative enhancements went on by defaultNo advertiser reports of rewritten ads in our sources; OpenAI changed bidding instead (cost target, intelligent bidding, Target CPA/CPC)Not yet
AI Mode ads in Brazil, Mexico, UK; separate AI Mode reportingTen countries live on 10 November; the three still "due this quarter"; AI Mode still not reported separatelyNot yet
Gemini 4 for free AI Mode usersGemini 4 Hydrogen the default for all users on 12 November; Carbon to Pro on 16 NovemberHappened, with Hydrogen not Argon
Attribution: Gemini tags, ChatGPT referrerNo Google statement on Gemini's tags and nothing on ChatGPT's referrer; Business Hub now counts clicks from inside ChatGPTNot yet
New Comscore and Similarweb share dataSimilarweb's October reading: ChatGPT 55.7%, Gemini 24.2%, Claude 11.5%; no new Comscore panelHalf happened
Into December: CMA choice screen, EU Article 50 on 2 DecemberNot yet due; the CMA's AI-click requirements for Google apply from 3 DecemberNot yet due
Checked against our weekly digests of 30 October – 14 November and notes for 14–18 November 2026.

The scorecard is mixed in a way that says something. The things the platforms control, models and ad machinery, moved faster than we expected. The things outsiders asked for, separate reporting, stable labels, a confirmation about Gemini's tags, did not move at all, and the one new number on clicks came from OpenAI's own dashboard rather than from analytics.

What it adds up to

This section is our reading, not a measurement.

  • The assistants are routing around the click. The click from an answer is small and, on the best evidence yet, shrinks when the answer gets richer. The platforms' answer is not to send more of it but to complete the task inside: a wallet, a booking link, an agent that calls the business, a checkout in Canada. A visit is becoming one possible outcome of being recommended, not the main one.
  • Agents now have identities, and sites have to choose. Google-Agent and Web Bot Auth give sites a way to recognise an agent, and both OpenAI and Google ask sites not to block them. The same fortnight brought a state lawsuit alleging agents running out of control. Letting agents in is now a business decision with a fraud, a legal and a revenue side, not a robots.txt line.
  • Being named, being linked and being clicked are three different numbers. ChatGPT links in nearly half of chats, names a brand in 12.1% of responses, and sends a sliver of the traffic. Report all three separately, and do not let a rise in one stand in for the others.
  • The best AI visibility data now comes from inside the assistant, on its terms. Business Hub is real data that no panel can match, with thresholds, no prompts and an ad product attached. Outside tools remain necessary for comparisons with competitors, which the hub does not give.
  • ChatGPT is the biggest assistant and not the biggest commercial one. Few of its conversations are about buying, its ads convert poorly in the only large study, and SparkToro argues the visibility that pays runs through Google's AI features. Treat ChatGPT and AI Mode as different channels with different jobs.
  • The reader behind the answer keeps changing, now silently. Three model changes in AI Mode in ten weeks, and from 16 November no label. Any trend line in AI visibility needs the dates of the reader, kept by you.

What to watch in the rest of November

  • OpenAI's answer to California, and whether other states follow the Attorney General's suit over agents.
  • ChatGPT Wallet beyond US Pro users, and the delegated purchases with spending limits OpenAI calls a later step; whether merchants report agent orders declined by their own fraud tools.
  • Google-Agent in the logs: whether Web Bot Auth signatures move from experiment to default, and whether Google says anything about robots.txt for agents.
  • A model indicator in AI Mode, which Rajan Patel said Google is "thinking through", and any independent reading of how Hydrogen changes which sources AI Mode cites.
  • ChatGPT's layout tests, including web-only results, which Search Engine Roundtable could not replicate on 16 November; a web-first layout would change what a citation is worth.
  • AI Mode ads in Brazil, Mexico and the UK, still due this quarter.
  • The November 2026 core update, which began on 10 November and, when Google's Search Status Dashboard was read on 18 November, had no completion entry; Hydrogen arrived during it, so AI Mode and ranking changes overlap in any report for these weeks.
  • Into December: the UK requirement for Google to report AI clicks from 3 December, the end of the EU's Article 50 grace period on 2 December, and Bing's temporary grounding preferences from 15 December.

To see what three assistants say about a brand today, across the buyer questions around it rather than one prompt, Brandscope runs them in one reading.

Questions about AI assistants in the first half of November 2026

What is Gemini 4 Hydrogen?

The model Google made the default in AI Mode on 12 November 2026, for all users and in every language AI Mode supports. Google calls it the biggest model upgrade to Search yet, and Robby Stein said it also handles the hardest questions in AI Overviews. Subscribers to Google AI Ultra and, from 16 November, AI Pro can use Gemini 4 Carbon.

What is ChatGPT Wallet?

A wallet launched on 13 November 2026 that lets OpenAI's agents pay merchants with card-network tokens instead of real card details, with biometric approval for each purchase. It is available to US ChatGPT Pro users aged 18 or over with eligible US cards. OpenAI asks merchants not to block agents and to recognise Web Bot Auth signatures.

What is Google-Agent?

The name Google's hosted agents use when they browse and act for users. Its requests carry Google-Agent in the user agent and come from published IP ranges; Google is experimenting with Web Bot Auth signatures. Google's help page of 17 November asks sites to allow it in CDNs, firewalls and bot tools.

Does AI Mode reduce clicks to websites?

In one preregistered field experiment with about 1,100 US users, reported by PPC Land on 17 November, people redirected to AI Mode clicked through to external sites 18.8 percentage points less often than with current Search, and the share clicking an ad fell 42.7 points. It is an unreviewed preprint, run in March 2026.

How often do people use ChatGPT to research purchases?

In SparkToro's analysis of 118,163 US prompts from January to October 2026, 6.9% showed research toward a brand to buy from, and 12.1% of responses named any brand, product or company. Rand Fishkin estimates Google's rate of purchase-journey queries is at least 16 times ChatGPT's.

Sources

Every claim above is dated and linked. Where a number comes from a measurement rather than an announcement, whoever measured it is named; where only a write-up was read, the write-up is the source listed.

The product

Watch it on every page, every day

One reading tells you where a page stands today. The product asks the same questions of the same assistants continuously, so a change is something you are told about rather than something you go looking for.

Request an invitation

Invite-only while we keep the readings honest

On this topic

Do AI assistants mention us?

Every day The same questions asked of real assistants on a schedule, every answer kept in the evidence store. How the product works

Author

Olexandrauthor

15+ years in marketing and SEO. Dozens of products, mostly the US market. Research and measurement.

LinkedIn

Pass it on

Read in your language

Opens a browser translation of this English article. The original source stays in English.

All issues