This briefing re-checks what is known about AI search each quarter: what was newly measured, what moved, and what Clarion's own instruments read. The window for this issue is 20 July to 16 November 2026.

Four results matter most. A randomized experiment, posted in April and missed by our August ledger, found about 40% fewer organic clicks on desktop searches where an AI Overview appears. A product launch in France pointed the same way: click-through on the most exposed of 963 large sites fell 23.1%, and 5.7% at the median site. A daily tracking study found that 50% to 59% of AI citations are gone the next day. And several widely repeated statistics changed category when we went back to their sources.

What this means for a firm like yours.

Clicks to ordinary results fall where AI summaries appear, so a ranking buys less than it did. What AI engines cite changes from day to day, so one check proves little, ours included. No one has yet shown that being cited produces revenue. The one thing with any support so far is that citation tracks relevance to the question asked, in one preprint, on ChatGPT only, as an association. That is a reason to answer your clients' real questions well. It is not a promise of citation. The case for the work is that a prospect who asks an AI tool about your field hears your name, described accurately.

A note on labels. Verified means we traced the figure to its source and the source says what is claimed. It does not mean the question is settled. The full definitions are at the end.

New to the ledger this quarter.

Verified - working paper

About 40% fewer organic clicks on desktop searches where an AI Overview appears

Agarwal and Sen ran a pre-registered randomized experiment with 1,065 US participants on desktop Chrome in January and February 2026. A browser extension hid AI Overviews for one group of 374, and a control group of 396 saw them. The remaining 295 were sent to AI Mode, an arm the paper reports as exploratory. On searches that triggered an overview, people who saw it made 0.37 organic clicks per search against 0.62 for people who did not, and 73% of those searches ended without a click against 54%. Ad clicks did not differ. The April draft put the reduction at 38% and the June draft at 39.8%. It is a working paper, not peer reviewed. We read the April and June drafts and have not read the July revision. It was missing from our August ledger.

Verified - preprint

18.8 points lower click-through when every search goes to AI Mode

Wang and colleagues ran a preregistered experiment with 1,100 US Chrome users in March 2026 and posted it in August. Sending every search to AI Mode for a week cut click-through to outside sites by 18.8 percentage points, with a 95% interval of 15.3 to 22.2. AI Mode was 0.6% of these users' searches beforehand, so this measures full adoption, which has not happened. A second arm removed AI features. Counting everyone assigned to it, click-through rose 2.7 points, too small to rule out chance. Counting only the people whose AI features were actually removed, it rose 8.8 points (95% interval 2.3 to 15.3), which is unlikely to be chance. The two differ because a change to Google's page meant the extension hid only 51.1% of AI Overviews.

Verified

-23.1% click-through in nine days, where AI Overviews arrived

Google launched AI Overviews in France on 22 July 2026. Ahrefs compared 963 high-traffic French domains before and after. Median click-through on the most exposed domains went from 2.29% to 1.76%. Domains with almost no exposure rose 2.3%. A launch date and a comparison group make this a before-and-after comparison across 963 large French sites, and it points the same way as the Agarwal and Sen experiment above. The 23.1% is the raw change for the most exposed group. Adjusted for the comparison group's rise, Ahrefs puts it at 24.8%. Across all 963 domains the median fell 5.7%. Nine days, one vendor, large sites only.

Verified

Half of AI citations are gone the next day, and most come back

Ahrefs asked five AI tools the same questions daily for 20 days and followed about 37,500 day-one citations. Between 50% and 59% were missing on day two in Google AI Overviews, AI Mode and ChatGPT. Gemini and Copilot lost about 80%, a figure inflated by answers that carried no citations. About three quarters of those returned within the 20 days. A single capture of an AI answer is a sample of one. That applies to ours.

Verified

AI Overviews on 73% of branded and 78% of non-branded keywords

Ahrefs' September 2026 monthly averages, US desktop, across the keywords it crawls. That is a share of the keywords one vendor crawls, which is a different thing from a share of all searches.

Verified

14.4% of free, logged-out ChatGPT responses contained fan-out search queries

In 66.6 million free, logged-out responses from August and September, 14.4% contained search queries, by Ahrefs' count. Ahrefs first described this as the share of responses that searched the web and corrected that wording in October, so it is not a count of how often ChatGPT searches. It describes one model in one period. Paid and reasoning modes search more.

Verified as dated - a different measure

"99% of AI Overviews cite pages from the organic top 10"

The likely source is seoClarity, reported by Search Engine Land in August 2024 across 36,000 keywords: AI Overview sources "match one or more" of the top ten organic results 99.5% of the time. The report does not define its denominator. Read as written, it counts overviews with at least one top-ten source. Ahrefs' 38% counts the share of all cited pages that rank in the top ten. Both can be true at once, so the 38% does not refute it. The figure is two years old and predates the model now behind AI Overviews.

Unproven - uncontrolled

"53.7% of ChatGPT categories have no clear brand leader"

Semrush, July 2026: 1,094 categories of its own choosing, five prompts each. A fair description of that panel. It cannot say why a brand leads.

Unproven - self-reported

"170% more leads and 82% more deals from AEO"

HubSpot's figures for customers it describes as actively optimizing for AI search, against "comparable" customers it does not define. The leads are marketing-qualified leads. No published method. The first evidence tying AI visibility to revenue will need a comparison group somebody else can inspect.

Unproven - depends on a definition

"Earned media drives 84% of AI citations"

Muck Rack's count is real under its own definition of earned, which includes encyclopedias, government sites, academic research and third-party corporate content. Journalism was about 27% in its May edition and 22% in October.

Refuted

"Schema markup increases AI citations by 30-40%"

Still circulating, with no method attached. Ahrefs' matched study of 1,885 pages found no lift.

One academic result deserves a line. A September preprint, CiteChoice, replayed real search transcripts with one document rewritten in a structured form. The rewritten document picked up about half a citation more per answer. Whether it was cited at all did not change by a detectable amount. Structure moved credit among sources already in the answer. It did not get a source into the answer.

Platform changes.

Google began a core update on 10 November and began rolling a new model, Gemini 4 Carbon, into AI Mode for paid subscribers from 10 November. OpenAI opened ChatGPT Business Hub in beta on 10 November, which shows verified businesses a count of mentions and clicks. Any comparison of an October reading with a November one spans at least two of those changes.

Our data.

These are Clarion's own readings from October and November. The October readings were taken about two months after we rewrote how the site describes the studio.

Search. Google Search Console showed 174 impressions and 9 clicks for clarion.studio in the 28 days to 10 October, and 260 impressions and 4 clicks in the 28 days to 1 November. Nearly all of it is searches for our own name. The first non-brand impressions arrived in September, on queries about the AI visibility gap. They have produced no clicks.

Citations. On 10 October we asked ChatGPT with web search and Perplexity what Clarion Studio is. Both cited clarion.studio. On a general question about the AI visibility gap, ChatGPT cited one of our pages first. Each was a single run, which the stability finding above says to treat as an anecdote.

Fidelity. Both engines described the studio the way the site described it at the time. Perplexity also drew on sources about unrelated businesses called Clarion, and on 1 October it had confused us with one of them.

Absences. We found no Clarion citation in a sampled set of Google AI Overviews, and on 15 October Bing showed no sign of having indexed the site.

November. On 16 November we put the general question about the AI visibility gap to ChatGPT and Perplexity through their APIs, once each. ChatGPT answered without searching the web and cited nothing. Perplexity cited ten sources, and none was ours. These were single runs, so they do not show a change.

Demand. Monthly US searches for "generative engine optimization" read about 4,400 in October, the same as in January. May read higher, at about 5,400, and April is missing from the series. "Answer engine optimization" read about 2,400 from April to July and rose to about 3,600 by October. These are one vendor's rounded estimates, pulled on 16 November.

In October a studio that sells this work was cited by two engines for its own name and by one for one topic, and was absent from the Google AI Overviews we sampled. In November one engine answered without searching the web and the other cited ten sources, none of them ours. The November runs used the APIs and the October runs used the apps, so the two are not a like-for-like comparison. Three months in, we have no repeatable citation on our own topic. One engine cited us once in October. Neither did in November.

Verdicts that moved.

We re-check every standing verdict against its primary source each quarter. This time nine entries in our register and statistics ledger changed category: the eight below, each showing its old label, and the 99% entry earlier in this issue. Both pages carry the current verdicts.

Verified as dated (was refuted)

"Wikipedia is 29.7% of ChatGPT's top citations"

The figure is in Ahrefs' study of the 1,000 pages ChatGPT cited most in September 2025. It comes from the same category table as the 32.3% influenceable share. It is a share of those 1,000 pages, and the snapshot is fourteen months old.

Verified as dated (was refuted)

"Perplexity overlaps Google's top 10 by 28.6%"

Printed in the Ahrefs overlap study, July 2025 data. It is one of the five values averaged to produce the 12% figure in our ledger.

Verified as dated (was refuted)

"80% of AI citations don't rank anywhere in Google"

Same Ahrefs study, same data. It counts cited pages that do not rank for the original prompt. The 12% counts cited pages in the top ten. The two figures describe different parts of one chart.

Unproven (was refuted)

"GEO tactics boost visibility by up to 40%"

The sentence is in the abstract of the 2024 GEO paper. It reports a benchmark result for sources already placed in the engine's context. A July 2026 survey of 45 GEO studies found no technique with a stable, cross-platform causal effect on being found in the first place. Accurate as a lab result. Unproven as a promise to a client.

Unproven (was refuted)

"Organic CTR dropped 61% on AI Overview queries" (41% without one)

Both figures are Seer Interactive's, published November 2025 from 3,119 informational queries across 42 client organizations. Seer's own April 2026 update reported click-through on AI Overview queries, across all query types, rising from 1.3% in December to 2.4% in February. Ahrefs' comparison ends in December 2025, the month Seer's series bottomed, so it neither confirms nor contradicts the rebound. One agency's client set, since revised by the agency.

Unproven (was refuted)

"9 in 10 ChatGPT-cited pages don't rank in Google's top 20"

The wording matches Semrush, July 2025: pages cited by ChatGPT search ranked 21st or lower "almost 90% of the time" for related queries. The study as a whole covered more than 500 marketing topics, and Semrush did not say how this figure was measured. Ahrefs measured something close in August 2025: 9.85% of the URLs ChatGPT's own searches returned were in Google's top 20 for the same search. AirOps, counting differently in March 2026, found 55.8% of cited pages in a top 20 for the prompt or one of its sub-queries. The answer depends on which query you rank against.

Verified as dated (was refuted)

"AI Overviews pull 76% of citations from the top 10"

This was Ahrefs' real finding in July 2025, counted on the three most visible citations in each overview. Ahrefs measured again in March 2026, across about four million cited URLs, and got 38%. The two differ in method as well as date. The 76% was accurate for the three most visible citations in July 2025 and does not describe AI Overviews now. It also measures Google's AI Overviews, where the 80% figure above measures the chat assistants, so the two are not in conflict.

Unproven (was refuted)

"AI engines prioritize claims repeated across many publishers"

We find no measurement of this mechanism, for or against. Google's guidance for its AI features says seeking inauthentic mentions is less useful than it looks. That is advice, not a test, so under the rule at the end of this issue the claim is unproven. Until someone measures it, it is not worth paying for.

Verdicts that held.

Pew's measurements. A traditional result was clicked in 8% of visits with an AI summary against 15% without, and the browsing session ended after 26% of pages with a summary against 16% without (March 2025 data, published July 2025). No newer behavioral measurement from Pew has appeared. Its survey published in June 2026 (fielded in February), in which 60% of adults said they read AI summaries, is self-report and answers a different question.

The schema null result. No new experiment on schema surfaced this quarter. Three things line up behind the Ahrefs study. An October preprint from the University of Toronto, covering 158,099 AI answers, found no significant difference in citations between pages with schema or FAQ blocks and other pages on the same site. Sites that use them were cited more, which is the correlation usually quoted. Google's guide for its AI features states that structured data is not required for them. Our own law-firm study found no detectable link between a firm's structured-data score and being named. One limit belongs on the record: every page in the Ahrefs study already had more than 100 AI Overview citations, so it says nothing about pages the engines have not yet found.

New support for the relevance entry. In the same preprint's replay of 6,178 pages on 362 sites, whether ChatGPT's own search retrieved a page accounted for 32% of the variation in citations within a site (95% interval 24 to 42). The page properties the authors measured accounted for 2.8%. It is one preprint, ChatGPT only, built on replayed searches, and it reports associations. It is consistent with the ledger entry that citation tracks relevance to the question.

The 12% overlap. It stands as published by Ahrefs, with two notes. The data is from July 2025. The average includes Perplexity at 28.6%, and the other four values sit between 6% and 9%.

The winnable third. The 32.3% stands as a September 2025 measurement. It has not been re-measured, and the ground under it has moved. Ahrefs reported in October that Wikipedia fell from about one in seven US ChatGPT answers in March to about one in 35 by late September, and Reddit from one in five to one in 43. Treat "about a third" as a dated estimate.

The hold list: claims we will not repeat as fact yet.

"AI-cited brands see 35% more clicks." The source is Seer Interactive's November 2025 study: 0.70% organic click-through when cited against 0.52% when not. Seer says in the study that it cannot show the citation causes the difference. Its April 2026 update, on a larger client set, puts the gap at 120% and repeats the caution. It stays on hold, now as a correlation with a name.

"91% of marketers name freshness." An NP Digital survey of 100 content marketers, April 2026. It measures belief.

"Top pages lose 58% of clicks." Ahrefs has published its method. It remains one vendor's correlation. Seer's 2026 data shows click-through on AI Overview queries rising over the following months, which is a different measure. The Agarwal and Sen experiment above also measured a different thing, all organic clicks per search on desktop, and found about 40% fewer where an AI Overview appears.

Forecasts. None survived again. Seer's own September 2025 projection of continued decline missed within four months.

How verdicts are assigned.

These definitions apply from this issue on. Verified means the figure is in its primary source and measures what the claim says. It does not mean the question is settled. Verified as dated means it was accurate when measured and should not be read as current. Refuted means a source we can cite contradicts it. Unproven means we cannot confirm it, or its source cannot support the general claim built on it. Every verdict records the address of its primary source and what that source measured, so each re-check starts from the source.

Issue 2 is due in February 2027. After it we review the cadence. If an issue finds nothing new to verify or refute, the field has settled and the briefing ends.

Updated 16 November 2026 with two randomized click experiments, the University of Toronto preprint and our November readings.

Sources.


Mark Nelson

Founder, Clarion Studio

Mark builds machine-readable identity infrastructure for expert practices - entity graphs, structured data, and editorial systems, working at the intersection of strategy, design, and front-end engineering. Based in Lynden, Washington.

Published November 16, 2026 · State of AI Search, Issue 1

Read more about Mark·Start a conversation