Anything Engine Optimization The rolling record of the AI-search industry.

Original analysis · Source audit

An AI search percentage needs a denominator

We checked 20 claims in our own Numbers reference. The mistakes were often in the words around the number.

By AEO Wire · Source review completed (UTC)

What changed

Of the 20 claims selected, 7 needed a factual or source-link correction, 7 needed tighter scope, and 2 were confirmed as stated. One historical value could no longer be checked on its changing source page. Two lacked accessible primary evidence for the exact claim. One could be partly checked through an interview with its authors.

These are editorial review outcomes, not an error rate for AI search research. We chose claims likely to be reused or misunderstood. We did not draw a random sample, and we did not rerun the underlying experiments.

The denominator changes the story

People Inc.’s investor deck assigns Google Search 21% of Core Sessions. Our reference had called that total traffic. The narrower metric matters when comparing publishers: two percentages can look comparable while counting different things. The company’s Q2 presentation, page 6, supplies the definition we should have preserved.

Another example concerns attribution. Alex Galinos recorded 51,200 text-fragment events on one brand’s site. Our summary treated them as AI Overview clicks or sessions. Events do not count unique visits, and text fragments also appear outside AI Overviews. His first-party account is useful evidence of a measurement problem, but its 22.4% Direct-channel share is not a correction factor to apply to every site’s organic traffic.

Before repeating a statistic, write the denominator in ordinary language: all answers, answers with a citation, citation links, domains, events, sessions or people. If that sentence is missing, the percentage is not ready for a decision.

The publication date is not the measurement date

Ahrefs published its 58% click-through estimate in February 2026 using December 2025 data. Our reference described it as a 2026 result. The primary analysis compares observed desktop position-one CTR with an estimated result without AI Overviews. That is an observational estimate with a specific window, not a universal loss a publisher should expect now.

A source link can change too. The Ahrefs ranking behind our July YouTube figure now carries a newer monthly table. We retained the old value as historical and marked it unverified in this review. Replacing it silently would erase the history; treating today’s page as proof of July would invent evidence.

A source check is not replication

Finding the percentage in the original report answers one question: did we represent that report accurately? It does not establish that the measurement is reproducible, representative or causal. The review labels below sit beside our existing replication labels, not in place of them.

For your next report, keep six things together: the claim, the counted population, the time window, the method, the original source, and the unresolved limitation. A confident headline should never make the limitation disappear.

Method and limits

We selected 20 entries from the 203-figure local Numbers reference for editorial importance, reuse risk, changing sources, or ambiguous measurement. We compared the claim with accessible primary pages, a company presentation and an authors’ interview, recording the source, window, denominator and decision. The selected entries and previous wording are preserved in the download. This is a purposive audit of our own summaries, not a representative survey of the industry or an independent validation of vendor datasets.

“Confirmed” means the accessible source supports the wording. “Qualified” preserves the reported figure while narrowing its interpretation. “Corrected” changes a factual description or source pointer. “Historical,” “partial” and “unresolved” identify limits on this check. No label predicts search rankings or explains this site’s traffic decline.

Download the review ledger (JSON) · Download CSV · Full Numbers reference

The 20 review records

qualified · SISTRIX

74%

Weekly domain churn-in reported for ChatGPT Search by SISTRIX: 74% in Germany, versus 60% in the UK and 42% in France. Churn-in counts domains new to this week’s set; it is not a universal URL-loss rate.

Counted population
Weekly domain churn-in; domains present this week
Measurement window
Dec 17, 2025–Apr 8, 2026
Review finding
Domain churn is not the percentage of last week’s URLs that disappeared. Country rates vary.

Source checked ↗ · Wire report · Reference entry · Link to this review

Previous reference wording

Share of its cited sources ChatGPT Search replaces every week, against 56% for Google, across 82,619 prompts re-sampled weekly for 17 weeks (1,548,213 snapshots, six countries) in a SISTRIX study — with a stable core of a few domains present for 86% of prompts and everything outside it rotating at 89% per week.

corrected · SISTRIX

83%

Complement of the 0.17 Jaccard domain overlap SISTRIX reports between Google AI Overviews and AI Mode. This measures shared domains relative to their union, not the percentage of prompts on which the engines disagree.

Counted population
Union of domains in matched AIO/AI Mode citation sets
Measurement window
17 weekly observations ending Apr 8, 2026
Review finding
The 0.17 Jaccard overlap is not the fraction of prompts where the engines agree.

Source checked ↗ · Wire report · Reference entry · Link to this review

Previous reference wording

Share of the time Google AI Overviews and Google AI Mode cite different domains for the same prompt, per SISTRIX — two surfaces from one company that agree on source selection less than a fifth of the time, so a citation-share number measured on one is not a reading of the other.

corrected · SISTRIX

1.4%

Core rate for the News/Media domain category in SISTRIX’s Google AI Mode table. It is not a measured lifetime for every news article across AI search.

Counted population
News/media domain category in Google AI Mode
Measurement window
17 weeks ending Apr 8, 2026
Review finding
The table is specific to AI Mode; do not generalize it to every news URL or engine.

Source checked ↗ · Wire report · Reference entry · Link to this review

Previous reference wording

Share of news articles cited as sources that remained permanently in the citation set across 17 weeks of weekly re-sampling, per SISTRIX — for a publisher an AI citation behaves less like an asset that accrues than a placement that expires. By contrast, 43% of brand queries kept the brand's own domain present in all 17 weeks while its co-citations rotated at 70% per week.

confirmed · Sistrix Blog

±34 points

Typical deviation SISTRIX reports for one prompt measured on 20 days, versus ±9.5 percentage points for 20 different prompts and ±4.2 for 100, using 10,826 daily series. These describe its sampling analysis, not guaranteed errors for another panel.

Counted population
Typical deviation in estimated brand mention rate
Measurement window
Published Oct 28, updated Nov 3, 2026
Review finding
Breadth and repeated runs answer different questions; this is the source’s model, not a universal error bar.

Source checked ↗ · Wire report · Reference entry · Link to this review

Previous reference wording

Typical error in a brand's measured AI mention rate when one prompt is queried on 20 days, against ±9.5 points for 20 different prompts and ±4.2 for 100, per Sistrix's analysis of 10,826 daily measurement series across four AI models and eight countries — for a brand that truly appears in 25% of answers, one prompt could read anywhere from 0% to 59%. Sistrix also reports that 70% of brands appearing in a prompt's answers show up on fewer than one in five days, and that a weekly average of 20 prompts is accurate to ±1.8 points against ±4.8 for a single day.

historical · Ahrefs

21.1%

Historical July 2026 YouTube mention-share figure recorded on the wire. The source URL is a monthly-updated ranking, so its current page is not evidence of this July value. See the source-review note before reuse.

Counted population
Summed citations of the top source domains
Measurement window
July 2026 record; live source updated Nov 3
Review finding
The linked page now reports November data. The July value is retained as historical, not reverified or silently replaced.

Source checked ↗ · Wire report · Reference entry · Link to this review

Previous reference wording

YouTube's mention share in Google AI Overviews — the most-cited domain, ahead of Reddit (18.5%), Facebook (10.7%), Google itself (7.1%) and Instagram (5.8%).

corrected · Search Engine Land

76%

Share of AI Mode passage variants reused at least 100 times that came from organically first-ranked pages in Pillarbase’s study. This is an observed association in its database, not evidence that a specific passage length causes citations.

Counted population
AI Mode passage variants reused at least 100 times
Measurement window
Pillarbase report labeled June 2026
Review finding
The 117-word statistic is a median in reconstructed client passages, a separate cohort, not the mean of all recycled passages.

Source checked ↗ · Wire report · Reference entry · Link to this review

Previous reference wording

Share of AI Mode passages reused 100 or more times that come from pages already ranking #1 organically, per a Pillarbase study of 15.7 million citations; the most-recycled passages average 117 words, answer in the first sentence about 80% of the time, and are self-contained about 85% of the time.

corrected · Search Engine Land

52%

Perplexity’s rate of citations without the source brand named in the answer, in Writesonic’s analysis of about 16 million brand appearances over a recent 30-day window. The study distinguishes linked citations from textual brand mentions; it does not measure reader recall.

Counted population
Citations without source-brand names in answer text
Measurement window
Recent 30-day window; report July 29, 2026
Review finding
The original co-authored report supports the rate; the old study link pointed to a generic guide.

Source checked ↗ · Wire report · Reference entry · Link to this review

Previous reference wording

Perplexity's rate of "ghost citations" — pulling from content without naming the source brand — the highest of seven engines Writesonic measured, followed by Google AI Mode (49%), AI Overviews (41%), ChatGPT (37%), Gemini (25%), Grok (22%) and Microsoft Copilot (19%), across roughly 16 million brand appearances.

qualified · Seer Interactive

72% vs. 42%

Shares of 4,124 cited pages classified as less than a year old using last-update dates versus original publication dates, in Seer’s study. The same pages were measured two ways; this does not establish that refreshing content beats publishing new content.

Counted population
4,124 cited pages with both dates readable
Measurement window
Mar–Jun 2026
Review finding
This is a comparison within cited pages, not a test of whether updating a page causes citation.

Source checked ↗ · Wire report · Reference entry · Link to this review

Previous reference wording

Share of AI-cited pages that looked fresh by last-modified date versus by original publish date, among 4,124 pages with both dates in a Seer Interactive analysis of 47,097 citations across ChatGPT, Gemini and Perplexity — refreshed old pages outcited newly published ones. Overall, 75% of cited pages had been updated within the past year, 88% within two years; recency requirements varied by engine (Gemini 78%, ChatGPT 73%, Perplexity 65%).

confirmed · BrightEdge Blog

33x

Growth in Reddit's tracked search footprint between June 2025 and February 2026, with its AI citations up another 84% from February to July 2026 before plateauing at a level several times higher than a year earlier — Google AI Overviews drives roughly 88% of Reddit's citation volume, per a BrightEdge analysis using its AI Hyper Cube, Data Cube X and Expert Services datasets.

Counted population
BrightEdge tracked keyword footprint, paid and organic
Measurement window
June 2025–February 2026
Review finding
Keyword footprint, citation volume and citation share are different metrics.

Source checked ↗ · Wire report · Reference entry · Link to this review

Previous reference wording

Growth in Reddit's tracked search footprint between June 2025 and February 2026, with its AI citations up another 84% from February to July 2026 before plateauing at a level several times higher than a year earlier — Google AI Overviews drives roughly 88% of Reddit's citation volume, per a BrightEdge analysis using its AI Hyper Cube, Data Cube X and Expert Services datasets.

qualified · Promptwatch

86.4%

Relative decline in Reddit’s share of ChatGPT Search citations, from 3.83% to 0.52%, in Promptwatch’s two date windows. Its proposed connection to changing search operators is an explanation to test, not a demonstrated cause.

Counted population
Reddit share of citations in answers with citations
Measurement window
Jul 18–Aug 7 vs Aug 14–17, 2026
Review finding
The site-operator explanation is the vendor’s hypothesis, not a causal result.

Source checked ↗ · Wire report · Reference entry · Link to this review

Previous reference wording

Relative drop in Reddit's share of ChatGPT Search citations — an average 3.83% (July 18-Aug 7) down to 0.52% (Aug 14-17) — against just an 11.3% decline on Google AI Overviews (2.37% to 2.10%) and 30.5% on AI Mode over the same window, per Promptwatch. It attributes the ChatGPT-only break to an August 8 change that pushed use of the site: operator in ChatGPT's background search queries from 0.37% to 16.8%, and calls the decline provisional pending confirmation it isn't a data-collection issue.

qualified · Ahrefs Blog

1 in 5 to 1 in 43

Change in the proportion of US ChatGPT answers citing Reddit in Ahrefs’ separate Brand Radar citation dataset, from March to late September 2026. This is answer incidence, not Reddit’s share of all citation links.

Counted population
ChatGPT answers citing Reddit
Measurement window
Mar–Sep 2026; 2.4–3.8m answers/month
Review finding
Answer incidence is not a share of all citation links.

Source checked ↗ · Wire report · Reference entry · Link to this review

Previous reference wording

Reddit's share of ChatGPT citations, which fell from about 1 in 5 answers in March 2026 to 1 in 43 by late September — an 88% drop — while Wikipedia fell from 1 in 7 to 1 in 35, down 80%, per Ahrefs' Brand Radar data tracking 2.4 to 3.8 million ChatGPT answers a month. Over the same stretch the combined citation share of OpenAI's 17 licensed content partners more than doubled, from 7.5% in March to a July peak of 15.8%, easing to 12.5% by late September. Reddit still ranks ninth among domains ChatGPT searches directly with the site: operator, which Ahrefs reads as the model still fetching Reddit for opinions while citing it less.

unresolved · Search Engine Journal

40%

Share of U.S. Google searches that sent a click to the open web in June 2026 (40.7% in the EU/UK) — a record low — while Google AI Mode held at just 0.13% of U.S. search-related visits, per Datos and SparkToro's Q2 State of Search report.

Counted population
Reported US desktop Google searches
Measurement window
June 2026
Review finding
The report landing page is readable but gates the data. Exact percentage and denominator not reverified; do not use as a current benchmark.

Source checked ↗ · Wire report · Reference entry · Link to this review

Previous reference wording

Share of U.S. Google searches that sent a click to the open web in June 2026 (40.7% in the EU/UK) — a record low — while Google AI Mode held at just 0.13% of U.S. search-related visits, per Datos and SparkToro's Q2 State of Search report.

corrected · Digiday

40%

Year-over-year decline in People Inc.’s Google Search Core Sessions in Q2 2026. Google Search represented 21% of Core Sessions, while total Core Sessions declined 22%, according to the company’s investor presentation. Core Sessions is not all traffic.

Counted population
People Inc. Google Search Core Sessions
Measurement window
Q2 2026 versus Q2 2025
Review finding
The company’s page 6 says 21% of Core Sessions, not all traffic.

Source checked ↗ · Wire report · Reference entry · Link to this review

Previous reference wording

Year-over-year decline in People Inc.'s Google search traffic — now 21% of its total traffic, down from 25% the prior quarter, with core sessions off 22% YoY — even as the publisher's CEO says it won't block Google's AI crawlers because that "would turn off search."

corrected · Search Engine Journal

58%

Lower average desktop position-one CTR associated with AI Overviews in Ahrefs’ December 2025 analysis, relative to its estimated no-AIO counterfactual. It compared 150,000 AIO keywords with 150,000 informational keywords without AIOs; the report was published in February 2026.

Counted population
Desktop position-one CTR versus modeled counterfactual
Measurement window
December 2025; published February 4, 2026
Review finding
300,000 keywords, not the 15-million-point analysis cited elsewhere; observational estimate.

Source checked ↗ · Wire report · Reference entry · Link to this review

Previous reference wording

Drop in Google AI Overviews' position-1 click-through rate in 2026, up from a 34.5% drop in 2025, per an Ahrefs analysis spanning roughly 15 million data points.

corrected · Search Engine Land

22.4%

Share of 51,200 tracked text-fragment events assigned to Direct in Alex Galinos’s nine-month, single-brand GA4 study. The signal is not exclusive to AI Overviews, and the figure is not a universal correction factor for organic traffic.

Counted population
51,200 tracked fragment events at one transport brand
Measurement window
Sep 2025–Jun 2026
Review finding
Events are not unique clicks or sessions. Text fragments also occur outside AI Overviews.

Source checked ↗ · Wire report · Reference entry · Link to this review

Previous reference wording

Average share of Google AI Overview citation-click traffic misattributed to the Direct channel instead of Organic Search in GA4, per a nine-month tracking study of 51,200 clicks across 1,661 AI Overview-cited snippets — worst in May 2026 at 29.3%. The same tracking found AI Overviews drove 7.53% of organic sessions overall, peaking at 16-17% in February-March 2026 before falling to 2-4% by August.

qualified · Ahrefs Blog

23.1%

CTR lost by French sites most exposed to AI Overviews (over 20% of their queries triggering one), comparing the 28 days before to the 9 days after AI Overviews launched in France on July 22, 2026; the median site across 963 tracked domains lost 5.7% of CTR and 8.8% of clicks, and nearly 48% of sites lost more than 10% of daily clicks within nine days. Health queries lost the most — a 22.2% AI Overview trigger rate and a 20.4% CTR decline. Ahrefs' Juliette Begue, using Google Search Console data, found roughly "one point of CTR lost per point of exposure."

Counted population
High-exposure domains’ CTR before/after rollout
Measurement window
28 days before vs 9 after July 22, 2026
Review finding
963 domains; short observational comparison, not an individual-site forecast.

Source checked ↗ · Wire report · Reference entry · Link to this review

Previous reference wording

CTR lost by French sites most exposed to AI Overviews (over 20% of their queries triggering one), comparing the 28 days before to the 9 days after AI Overviews launched in France on July 22, 2026; the median site across 963 tracked domains lost 5.7% of CTR and 8.8% of clicks, and nearly 48% of sites lost more than 10% of daily clicks within nine days. Health queries lost the most — a 22.2% AI Overview trigger rate and a 20.4% CTR decline. Ahrefs' Juliette Begue, using Google Search Console data, found roughly "one point of CTR lost per point of exposure."

partial · Digital Content Next

-40%

Approximate reduction in outbound clicks per search reported by Saharsh Agarwal and Ananya Sen in their randomized AI Overview experiment, confirmed in their interview. The paper’s exact sample and secondary figures were not reverified in this review.

Counted population
Outbound clicks conditional on a search
Measurement window
Experiment described by authors; exact paper not retrieved
Review finding
Authors confirm an approximately 40% effect in their interview. SSRN returned 403; exact sample/date/submetrics remain unreverified.

Source checked ↗ · Wire report · Reference entry · Link to this review

Previous reference wording

Reduction in outbound clicks per search when Google shows an AI Overview — from 0.62 down to 0.37 — in the first randomized field experiment on the question: 1,065 U.S. desktop searchers randomly assigned to see results with or without AI Overviews over two weeks in early 2026. Self-reported search satisfaction held flat at 4.0 out of 5 either way, and just 7.1% of outbound clicks on AI Overview-bearing pages came from links inside the AI-generated answer itself, with the rest going to ordinary organic results on the same page.

qualified · Sistrix Blog

2.7x

Citation-odds ratio associated with a recent year in the title in SISTRIX’s matched-page study. The feature was added retrospectively; the result is correlational and does not establish that changing a title’s year increases citations.

Counted population
Matched same-domain citation odds
Measurement window
Q2 2026, DACH market
Review finding
The title-year feature was added retrospectively. An association does not justify changing dates without content changes.

Source checked ↗ · Wire report · Reference entry · Link to this review

Previous reference wording

Citation-odds multiplier for a page carrying the current year in its title, versus a matched, uncited page on the same domain — the strongest of 43 page-level features tested, per Sistrix's conditional logistic regression across 48,527 pages cited by ChatGPT, Google AI Mode and AI Overviews in Q2 2026, matched against uncited same-domain pages (96,418 URLs, 17,380 domains). Other factors: a table of contents, used by only 14% of pages, carried a 1.7x factor; Article schema markup 1.6x; five or more H2 headings 1.6x; FAQPage schema 1.5x. Pages carrying many of these features were cited 2.1 times as often as minimally optimized ones; on ChatGPT specifically, a question-phrased title carried a separate 1.9x factor. The dataset covers the German-speaking (DACH) market.

qualified · Search Engine Roundtable / AlsoAsked

97%

Share of People Also Ask answers classified as AI Overviews in AlsoAsked’s first-week-of-September sample, reported by Mark Williams-Cook. The source describes 19.2 million English-language queries; this is not the share of all Google searches.

Counted population
People Also Ask answers in AlsoAsked’s sample
Measurement window
First week of September 2026
Review finding
Author’s post supports the reported rate, not a claim that 97% of all Google searches have AI Overviews.

Source checked ↗ · Wire report · Reference entry · Link to this review

Previous reference wording

Share of Google's People Also Ask boxes answered by an AI Overview in the first week of September 2026, per AlsoAsked's analysis of 19.2 million English-language queries — up from 86% in August, the last major SERP feature AI Overviews hadn't fully absorbed.

unresolved · Search Engine Roundtable

93%

Share of roughly 100 large brand names that triggered a Google AI Overview when searched directly — including Reddit, Salesforce, Amazon, Adobe, Wikipedia, Samsung, Apple and Yahoo — well above the 43% aggregate rate across all Google searches, per researcher Chris Long. Most placements sat mid-page or lower rather than at the top; Google showed none for searches on its own name or on news publishers' names. Long found AI Overviews far less common on the same kind of brand search when he ran a similar test roughly two years earlier.

Counted population
Approximately 100 selected large-brand queries
Measurement window
September 2026
Review finding
Original X post could not be retrieved. Secondary reporting repeats the claim; retain as a labeled historical report, not independent confirmation.

Source checked ↗ · Wire report · Reference entry · Link to this review

Previous reference wording

Share of roughly 100 large brand names that triggered a Google AI Overview when searched directly — including Reddit, Salesforce, Amazon, Adobe, Wikipedia, Samsung, Apple and Yahoo — well above the 43% aggregate rate across all Google searches, per researcher Chris Long. Most placements sat mid-page or lower rather than at the top; Google showed none for searches on its own name or on news publishers' names. Long found AI Overviews far less common on the same kind of brand search when he ran a similar test roughly two years earlier.