Insight · 16 min
AI search statistics, and what each one measured
AI search statistics disagree because they count different things. “38% of AI Overview citations come from the top 10” and “54% overlap with organic rankings” can both be accurate, because one counts URLs in the top 10 and the other counts a wider ranking depth, at different dates, with different parsers. This page lists the figures people quote most, grouped by question, with the publisher, date, sample, and denominator of each, and a link to the primary source. Every figure in the main tables was checked at its primary source on November 8, 2026. Figures we could not check there are in a separate table at the end of each section, labeled as such.
Last checked 2026-11-08. We update this page when a primary study is revised or replaced.
In short
- The share of AI Overview citations that rank in Google’s top 10 ranges from 16.7% to 76.1% across studies. The spread comes mostly from the denominator, the ranking depth, and the parser, not from the engine changing its mind.
- Reddit’s “share of AI citations” ranges from under 1% to about 60%, because some studies divide by all citations, some by the top 10 domains, and some by responses.
- Every major CTR study finds fewer clicks where an AI Overview appears. Sizes range from about 15% to about 60% depending on method.
- AI referral conversion rates range from “no significant difference” to “3x and more”. The two studies with stated statistical tests or medians found modest or no lift.
- Citation volatility is high on every measure, but “half-life”, “churn”, and “gone next day” are three different metrics and should not be compared directly.
Do AI Overview citations come from top-ranking pages?
Partly, and how much depends on what the study counted. The Ahrefs series is the most quoted, and its two figures used different methods.
| Study | Publisher | Date | Sample | What was counted | Headline figure | Primary link |
|---|---|---|---|---|---|---|
| 76% of AI Overview citations pull from the top 10 | Ahrefs | Jul 21, 2025 | ~1M AI Overviews, 1.9M citations | Top 3 citations per overview, matched to the page’s Google ranking | 76.1% in top 10; 9.5% at 11–100; 14.4% below 100 | ahrefs.com |
| Update: 38% of AI Overview citations pull from the top 10 | Ahrefs | Mar 2, 2026 | 863K keyword SERPs, ~4M AI Overview URLs, Ahrefs Brand Radar | All cited URLs; a match is the same URL in the overview and the SERP for the same query | 37.9% in top 10 (all SERP blocks); 37.1% (blue links only) | ahrefs.com |
| AI Overview citations now 54% from organic rankings | BrightEdge | Sep 18, 2025 | 9 industries, May 2024–Sep 2025; counts not given | Citation overlap with organic rankings; ranking depth not stated | Overlap rose from 32.3% to 54.5%. The same page says “only 16.7% of citations come from top 10 results” | brightedge.com |
| How Google’s AI Mode compares to traditional search and other LLMs | Semrush | Jul 21, 2025 | 5,000 keywords, 150K+ citations, 4 intent groups, desktop | Domains and URLs cited, compared with Google’s top 10 for the same query | AI Mode: ~54% domain, ~35% URL overlap. AI Overviews: ~86% domain, ~67% URL. Perplexity: 91%+ domain, 82% URL | semrush.com |
How to read them:
- Ahrefs 76% to 38% is not a clean trend. The 2025 study counted only the top three citations per overview. The 2026 study counted all of them, with improved parsing that finds more citations. In the 2025 data, each lower citation slot ranked lower in Google (median position 2 for the first citation, 4 for the second), so counting every citation would be expected to lower the top-10 share on its own. Ahrefs attributes the drop to query fan-out and notes that Gemini 3 has powered AI Overviews since January 2026. Both explanations are hypotheses.
- BrightEdge’s two figures measure different things. The 54.5% appears to include rankings beyond the top 10; the page mentions pages ranking 21 to 100. The 16.7% is the top-10 share. The page does not say what period the 16.7% covers.
- Domain overlap is always higher than URL overlap. Semrush’s AI Overviews figure falls from 86% to 67% when you move from “same site” to “same page”. Quote the level.
How much of AI search cites Reddit and Wikipedia?
It depends entirely on the denominator, and in August 2026 the ChatGPT figure for Reddit dropped sharply in one tracker’s data.
| Study | Publisher | Date | Sample | What was counted | Headline figure | Primary link |
|---|---|---|---|---|---|---|
| AI platform citation patterns | Profound | Jun 5, 2025 (updated Aug 2025) | 680M citations, ChatGPT, Google AI Overviews, Perplexity, Aug 2024–Jun 2025 | Two measures: share of all citations, and share within each platform’s top 10 cited domains | ChatGPT: Wikipedia 7.8% of all citations, 47.9% of top-10 citations; Reddit 1.8% / 11.3%. AI Overviews: Reddit 2.2% / 21.0%. Perplexity: Reddit 6.6% / 46.7% | tryprofound.com |
| The most-cited domains in AI: a 3-month study | Semrush | Nov 10, 2025 | 230K+ prompts, 100M+ citations, ChatGPT search, AI Mode, Perplexity, 13 weekly snapshots Jul 14–Oct 12, 2025 | Share of responses that cite a domain (as phrased; the article does not define the denominator) | ChatGPT cited Reddit in close to 60% of responses in early August 2025, falling to about 10% by mid-September; Wikipedia from about 55% to under 20% | semrush.com |
| Most cited domains in AI search | Peec AI | Mar 2026 | 30M cited sources, US, ChatGPT, AI Mode, Gemini, Perplexity, AI Overviews | Sources cited directly in answers; ranks only | Reddit #1, YouTube #2, LinkedIn #3, Wikipedia #4. Shares not published on the summary page | docs.peec.ai |
| Reddit citations are dropping in ChatGPT | Promptwatch | Aug 18, 2026 | Interface-collected responses with at least one citation, Jul 7–Aug 17, 2026; counts not given | Reddit’s share of all citations, per engine | ChatGPT: 3.83% (Jul 18–Aug 7) to 0.52% (Aug 14–17), an 86.4% relative drop. AI Overviews: 2.37% to 2.10%. AI Mode: 2.22% to 1.54% | promptwatch.com |
| How large language models source brand reputation across languages and markets | arXiv (Zatuchin) | Jun 24, 2026 | 167,551 URL-grounded citations, 128 brands, 12 markets, 13 languages (Rankfor.AI data) | Citations about brands, by domain and source type | 85.7% of citations point to sites the brand does not own. Wikipedia is the most-cited domain in 11 of 12 languages | arxiv.org |
How to read them:
- Profound’s Wikipedia figure is 47.9% or 7.8% of ChatGPT citations, depending on which column you read. The 47.9% is a share of the top 10 domains only. Quoting it as “half of ChatGPT’s citations” is wrong by a factor of six.
- Semrush and Promptwatch use different denominators. “Cited in 60% of responses” and “3.83% of citations” can describe the same engine in the same year. A domain can appear in most responses and still be a small share of all links.
- The August 2026 drop is one vendor’s data. Promptwatch calls it provisional and says it cannot rule out a data-collection issue. It links an August 8 change in ChatGPT’s fan-out queries, when queries using the
site:operator jumped from 0.37% to 16.8% of fan-outs, but the sharper fall came six days later. Search Engine Journal’s report noted the cause is unconfirmed. Google’s surfaces declined gradually, not in a step. - Brand-focused samples differ from general ones. The arXiv study counts citations in answers about specific brands, which is the B2B case. General prompt sets weight forums and encyclopedias differently.
The third-party citations guide covers what this means for where you earn mentions.
How much do AI Overviews reduce clicks?
All four widely quoted studies found a reduction. They measured it in four different ways.
| Study | Publisher | Date | Sample | What was counted | Headline figure | Primary link |
|---|---|---|---|---|---|---|
| Update: AI Overviews reduce clicks by 58% | Ahrefs | Feb 4, 2026 | 300K keywords (150K with an overview, 150K informational without); aggregated desktop Search Console data, Dec 2023 vs Dec 2025 | Position-one CTR for overview keywords, against a forecast built from the no-overview group | 0.016 actual vs 0.037 forecast, an estimated 58% reduction | ahrefs.com |
| Google AI Overviews study | Amsive | Apr 16, 2025 | 700K keywords, 10 sites, 5 industries; focus on 10K that triggered an overview and ranked | Average CTR change; periods and calculation not stated | -15.49% overall; -19.98% non-branded; +18.68% branded with an overview; -37.04% with both an overview and a featured snippet | amsive.com |
| Google users are less likely to click on links when an AI summary appears | Pew Research Center | Jul 22, 2025 | 900 U.S. adults, March 2025 browsing, 68,879 searches, 12,593 with a summary | Share of visits to a results page that included a click | 8% clicked a traditional result with a summary vs 15% without; about 1% clicked a link in the summary. 18% of searches produced a summary | pewresearch.org |
How to read them:
- Ahrefs’ 58% is an estimate against a counterfactual. The raw drop on overview keywords, from 0.073 to 0.016, is larger, because it includes everything else that changed in two years.
- Amsive’s figures cannot be recomputed. The study does not say whether they are relative or percentage-point changes, or which periods were compared.
- Pew measures people, not sites. It is the only one of the three not built from site CTR data: it records what a panel of people did on results pages.
A diagnosis method for your own site is in rankings held, clicks fell.
Does AI search traffic convert better?
The evidence is mixed, and the larger the stated multiple, the narrower or less tested the sample tends to be.
| Study | Publisher | Date | Sample | What was counted | Headline figure | Primary link |
|---|---|---|---|---|---|---|
| LLM traffic conversion rate study | Amsive | Sep 3, 2025 | 54 sites, first-party GA4, 6 months, last-touch; conversions manually audited | Session conversion rate: B2B demo or form fills, ecommerce purchases | Organic 4.60% vs LLM 4.87%; paired t-test p = 0.794, not significant. LLM was 0.24% of sessions | amsive.com |
| AI traffic conversion rates: are they really 7x better? | Siege Media | Jun 30, 2026 | 78 of 120+ screened GA4 properties, Jan 1–May 31, 2026; minimum 250 AI sessions | GA4 session key event rate, AI vs organic, median ratio | Median 1.26x. AI matched or beat organic on 72% of sites. B2B SaaS median 1.05x | siegemedia.com |
| AI traffic converts at 3x the rate of other channels | Microsoft Clarity | Nov 6, 2025 | 1,277 publisher and news domains, one month of conversion data | Sign-up and subscription rates by channel, from Clarity smart events | LLM sign-ups 1.66% vs search 0.15%; subscriptions 1.34% vs 0.55%. AI under 1% of traffic | clarity.microsoft.com |
How to read them:
- Amsive is the only one with a reported significance test on the main result. Its B2B subset showed LLM 2.17% vs organic 1.16% on the full dataset, and the difference was not significant in the thresholded 33-site sample (p = 0.705).
- Siege reports a median, not a mean. It also grouped ratios above 3x so that one site could not dominate a category. Siege sells GEO services, so a larger multiple would have suited its sales case.
- Clarity’s sample is publishers. Sign-ups and subscriptions on news sites are not B2B demo requests, and the post does not show how the “3x” headline was derived from its table.
- All three rely on referral data. Visits from assistants that strip the referrer land in “direct”, so AI traffic is undercounted and its conversion rate is computed on the visits that can be identified. Attributing AI search to pipeline covers the setup.
How stable are AI citations?
Not very, on every measure. But each study defines stability differently.
| Study | Publisher | Date | Sample | What was counted | Headline figure | Primary link |
|---|---|---|---|---|---|---|
| The half-life of an AI citation is 11 days | Profound | Sep 30, 2026 | ~883K pages, ~1.19M page-engine lifecycles, 7 engines, Sep 8, 2025–Sep 8, 2026 | Days from a page’s peak citation share until it falls to half and stays there 14 days | Median half-life 11 days after peak; 42 days from first citation. About 78% fall to half within two weeks | tryprofound.com |
| How stable are AI citations? 10 things we learned tracking 37,500 citations | Ahrefs | Oct 29, 2026 | ~37,500 day-one citations, 5 tools (AI Mode, ChatGPT, AI Overviews, Gemini, Copilot), same questions daily for 20 days | Whether each day-one citation appeared the next day, and whether it returned | Gone next day: AI Overviews 50%, AI Mode 58%, ChatGPT 59%, Gemini 81%, Copilot 83%. About 75% of lost citations returned within 20 days | ahrefs.com |
| What 4,599 ChatGPT responses reveal | Go Fish Digital | Oct 5, 2026 | One client, 73 prompts asked daily for 63 days, one region | Brand mentions, compared day over day (4,108 comparisons) | In 60.0% of comparisons, at least half of the previous day’s mentions were gone | gofishdigital.com |
| AI citation drift | SISTRIX | May 1, 2026 (modified May 12) | 82,619 prompts, 1.55M snapshots, 6 countries, 17 weekly dates Dec 17, 2025–Apr 8, 2026 | Weekly churn-in: share of a week’s cited domains not cited the previous week | ChatGPT Search 74%; AI Mode 56% | sistrix.com |
How to read them:
- These are four metrics. Half-life after peak, next-day loss, day-over-day mention loss, and weekly domain churn cannot be lined up as one number. Ahrefs’ data shows many lost citations come back, which a half-life metric does not capture.
- Gemini and Copilot look unstable partly because they often cite nothing. Ahrefs found 64% of Gemini answers and 53% of Copilot answers had no citations.
- Go Fish’s result is one client. The study itself says it does not generalize and that its 60% threshold was chosen after collection.
- SISTRIX’s AI Overviews figure is not quoted here because the page’s chart and text describe it differently.
The practical lesson is the same across all four: one check is not a measurement. Run each query repeatedly and report a rate, as described in measuring AI search visibility and share of answer.
How often does ChatGPT search the web?
About a third of the time in early 2026, by one clickstream estimate, and the share moves a lot from month to month.
| Study | Publisher | Date | Sample | What was counted | Headline figure | Primary link |
|---|---|---|---|---|---|---|
| ChatGPT traffic analysis: insights from 17 months of clickstream data | Semrush | Apr 7, 2026 | Semrush U.S. clickstream panel, mobile and desktop, Oct 2024–Feb 2026 | Share of queries (the article also says sessions) with web search enabled | 34.5% in Feb 2026, down from 46% in late 2024; monthly range 15% to 66.3% | semrush.com |
This figure is often attributed to “Semrush and Datos”. The article uses Semrush’s own clickstream panel. Datos appears only as the employer of a quoted executive. The unit is also unclear: the article switches between queries and sessions.
Which widely repeated figures could we not verify?
The figures below circulate widely. We could not confirm them at a primary source on November 8, 2026, so they are not in the tables above. Quote them only with this caveat, or not at all.
| Claim as usually repeated | Attributed to | What we found |
|---|---|---|
| Organic CTR fell 61% and paid CTR 68% on informational queries with an AI Overview | Seer Interactive, Nov 2025 | Seer’s page blocked automated access. PPC Land’s report gives the method: 3,119 informational queries, 42 clients, June 2024–Sept 2025; organic 1.76% to 0.61%, paid 19.7% to 6.34%; queries without an overview also fell about 41% |
| About 17% of AI Overview citations rank in the top 10 (February 2026) | BrightEdge | Reported by secondary sources. We found no February 2026 BrightEdge page. BrightEdge’s September 2025 page has a 16.7% top-10 figure, listed above |
| 43% of AI Overview citations come from the top 10 | Conductor | No Conductor source found. A 43% figure circulates for a different metric, how often position-one URLs are cited |
| AI traffic converts at 7x, 4.4x, or 23x organic | Various | Siege Media attributes a 7.1% AI conversion rate to Similarweb. We did not verify the original sources for these multiples |
| LinkedIn is cited in 14.3% of ChatGPT Search responses | Semrush 2026 (325K prompts) | Reported by secondary sources; we did not reach the primary page |
Why do AI search studies disagree?
They disagree for six recurring reasons, and most disagreements trace back to the first two.
- The denominator. Share of all citations, share of the top 10 domains, share of responses, and share of prompts give very different numbers for the same data. Profound’s Wikipedia figure for ChatGPT is 7.8% or 47.9% depending on which one you divide by.
- The ranking depth and match rule. Top 10 versus top 100, blue links versus all SERP blocks, and domain versus URL. Semrush’s AI Overviews overlap is 86% at domain level and 67% at URL level.
- The parser. Vendors improve how they detect citations, and each improvement changes the count. Ahrefs’ 2026 parser found more citations than its 2025 one.
- The engine and version. AI Overviews moved to Gemini 3 in January 2026. ChatGPT’s default model and its search behavior change often, and Semrush reports the web-search share swinging between 15% and 66.3%.
- The date. Many studies were collected in 2025 and are still quoted in late 2026. Citation sources shift within weeks.
- The vendor. Most of these studies come from companies that sell tracking tools or services. That does not make them wrong, and several state their limits clearly. It does mean the sample is usually their customers or their own prompt set, not a random draw.
How should you read a statistic before quoting it?
Check six things, and if you cannot answer one, say so when you quote it.
| Check | Question to ask | Red flag |
|---|---|---|
| Source | Can you open the study itself, not a roundup? | The only link is a list of statistics |
| Denominator | Share of what: citations, responses, prompts, sessions? | The article never says |
| Sample | How many queries, sites, or people, and how chosen? | One client, or an unstated count |
| Date | When was the data collected, not published? | Collected before the engine’s last major model change |
| Method | Is it observed, or a forecast or model? | A counterfactual reported as a measured drop |
| Interest | Does the publisher sell what the number supports? | No limitations section |
Then quote it with its frame: “In Ahrefs’ March 2026 study of 863,000 keyword results, 38% of AI Overview citations ranked in the top 10.” That sentence carries the publisher, date, sample, and denominator, and it will still be true when the next study disagrees.
How do we use these numbers?
We use them for direction and use a client’s own data for size. Every study above measures someone else’s queries. For a B2B site, the numbers that decide a plan come from its own query universe, run repeatedly across a fixed engine panel, with mention, citation, link, and accuracy scored separately. That is the method behind the measurement platform, and it is why we treat SEO, AEO, and GEO as one system scored on one query set.
Update policy
Last checked 2026-11-08. We update this page when a primary study is revised, withdrawn, or replaced by a newer edition, and when we can verify a figure now in the unverified table. Each update changes the “last checked” date and the affected rows. If you find an error or a newer primary source, tell us.
Common mistakes
- Quoting a top-10 share and a top-100 share as if they disagree. Fix: state the ranking depth with the number.
- Quoting Profound’s 47.9% as “half of ChatGPT’s citations”. Fix: it is a share of the top 10 domains; the all-citation share is 7.8%.
- Treating Ahrefs’ 76% to 38% as a trend line. Fix: say the method changed between the two studies.
- Citing a roundup instead of the study. Fix: open the primary page and copy its sample and date.
- Comparing half-life, churn, and next-day loss. Fix: name the metric, or do not compare.
- Using a vendor’s conversion multiple as a forecast. Fix: measure your own conversion rate, with enough AI sessions to mean something.
Sources
- 76% of AI Overview citations pull from the top 10 — Ahrefs (Louise Linehan), July 21, 2025, accessed 2026-11-08
- Update: 38% of AI Overview citations pull from the top 10 — Ahrefs (Louise Linehan), March 2, 2026, accessed 2026-11-08
- AI Overview citations now 54% from organic rankings — BrightEdge, September 18, 2025, accessed 2026-11-08
- How Google’s AI Mode compares to traditional search and other LLMs — Semrush (Eugene Levin), July 21, 2025, accessed 2026-11-08
- AI platform citation patterns — Profound (Nick Lafferty), June 5, 2025, updated August 2025, accessed 2026-11-08
- The most-cited domains in AI: a 3-month study — Semrush (Luke Harsel), November 10, 2025, accessed 2026-11-08
- Most cited domains in AI search — Peec AI, March 2026, accessed 2026-11-08
- Reddit citations are dropping in ChatGPT — Promptwatch (Klaas Foppen), August 18, 2026, accessed 2026-11-08
- Why Reddit’s ChatGPT citation drop isn’t fully explained — Search Engine Journal (Matt G. Southern, secondary), August 19, 2026, accessed 2026-11-08
- How large language models source brand reputation across languages and markets — arXiv (Dmitrij Zatuchin), June 24, 2026, accessed 2026-11-08
- Update: AI Overviews reduce clicks by 58% — Ahrefs (Ryan Law), February 4, 2026, accessed 2026-11-08
- Google AI Overviews study — Amsive (Will Guevara), April 16, 2025, accessed 2026-11-08
- Google users are less likely to click on links when an AI summary appears in the results — Pew Research Center, July 22, 2025, accessed 2026-11-08
- AIO impact on Google CTR: September 2025 update — Seer Interactive, November 4, 2025 (blocked automated access; not verified at source), accessed 2026-11-08
- Google AI Overviews reduce organic CTR 61%, paid traffic 68% — PPC Land (secondary, reporting Seer Interactive), November 5, 2025, accessed 2026-11-08
- LLM traffic conversion rate study — Amsive (Will Guevara), September 3, 2025, accessed 2026-11-08
- AI traffic conversion rates: are they really 7x better? — Siege Media (Marianna Mankle), June 30, 2026, accessed 2026-11-08
- AI traffic converts at 3x the rate of other channels — Microsoft Clarity (Ihab Rizk), November 6, 2025, accessed 2026-11-08
- The half-life of an AI citation is 11 days — Profound (Davis McCain), September 30, 2026, accessed 2026-11-08
- How stable are AI citations? 10 things we learned tracking 37,500 citations — Ahrefs (Mateusz Makosiewicz), October 29, 2026, accessed 2026-11-08
- Go Fish research: what 4,599 ChatGPT responses reveal — Go Fish Digital (Ashish Jacob), October 5, 2026, accessed 2026-11-08
- AI citation drift: how stable are sources in AI search results? — SISTRIX (Johannes Beus), May 1, 2026, accessed 2026-11-08
- ChatGPT traffic analysis: insights from 17 months of clickstream data — Semrush (Luke Harsel), April 7, 2026, accessed 2026-11-08