AI Overview Citations: What 15 Studies Agree On and Don't

Ask how often Google's AI Overviews cite pages that already rank in the top 10 and Ahrefs alone has published both 76.1% and 37.9%, eight months apart, and other studies on this page report figures as low as 16.7%. Between them sit thirteen other sources: nine more vendor studies, a Pew Research Center panel, two academic papers and Google's own documentation, each counting something slightly different. This page compares all fifteen, with every number tied to the study that produced it and the method that explains why it differs from the next one.

The comparison is built from the primary pages, read on November 5, 2026. Eleven of the fifteen are vendor studies, and every one of those vendors sells a tool that measures the thing it studied, which is recorded as an evidence tier. We did not average any figures across studies, because they do not share a denominator, and we flag the one number we derived ourselves. Where a page contradicts itself, that is recorded too.

Two companion pieces hold the context this page leaves out: our zero-click search data update for the traffic picture, and our Search Console guide for how Google reports AI feature traffic to site owners.

On July 21, 2025, Ahrefs published an analysis of 1.9 million citations from one million AI Overviews and reported that 76.10% of cited pages ranked in the top 10, 9.50% between positions 11 and 100, and 14.40% nowhere in the results. The study counted only the three most visible citations in each Overview. On March 2, 2026, Ahrefs published an update built on 863,000 keyword SERPs and four million cited URLs: 37.9% of cited URLs appeared within the first ten SERP blocks, 31.2% in positions 11 to 100, and 31.0% beyond the top 100. Counting organic blue links only, the top-10 figure was 37.10%.

Ahrefs reads the pair as a real change: its 2026 page says Google "is selecting far fewer pages straight from the original SERP (~76% in July 2025 vs. ~38% today)", notes that Overviews have been powered by Gemini 3 since January 2026, and suggests, as a hedge, that "fan out queries are playing a bigger role in source selection". The same page also records method changes a reader should weigh before treating 76 to 38 as a like-for-like fall, though Ahrefs does not draw that conclusion itself. The 2025 sample was the top three citations, and Surfer's data shows top-three citations rank higher than the rest (45.86% outside the top 10 against 67.82% for all citations). The 2026 test counted every SERP block, including ads, featured snippets and video packs, as positions. And Ahrefs says it had improved its parsing "so that we can see even more of the citations that appear in AI Overviews".

One vendor page shows the same problem inside a single document. BrightEdge's September 18, 2025 analysis reports that the overlap between Overview citations and organic rankings grew from 32.3% to 54.5% over sixteen months, and the same page states that only 16.7% of citations come from top-10 results. Neither the overlap definition nor the denominator is given, so both numbers may be correct and still cannot be reconciled by a reader. The example of the right use of any of these figures is to quote the study, the date, the citation scope and the rank range together. The failure case is a sentence that begins "studies show X% of AI Overview citations come from the top 10", because no single X exists.

Most of the apparent conflict between studies disappears once you ask what sits under the fraction. There are at least three denominators in use. One counts keywords: for what share of queries does the Overview cite at least one ranking page? One counts citations: of all the URLs cited, what share rank? One counts Overviews or domains. A study can report a high number on the first and a low number on the second from the same crawl, and seoClarity's October 2025 research does exactly that.

seoClarity analysed 362,000 US desktop queries that triggered an Overview on October 12, 2025, with 5.1 million citations. On the keyword denominator, 94% of queries had at least one top-20 URL among the citations and 90% had at least one top-10 URL. On the citation denominator, 44% of citations came from beyond the top 20. It also reports a position gradient: a position-one URL appears in the Overview 43% of the time, falling to 7% at position 20. The page carries an internal inconsistency on the share from the top 20, stating both 56% and a figure near 49%, and we report the gradient rather than resolve it.

Two further studies show how rank range and fan-out credit move the number. Rich Sanger's September 2024 study with Authoritas, across 11,163 queries of which 2,358 triggered an Overview, found 46.3% of linked documents in the top organic results for the exact query, rising to 67.3% once related and reformulated queries were credited. Surfer's 2026 study, on 10,000 keywords with the top 10 pulled for each query and for Gemini-generated fan-out queries, found 67.82% of cited pages outside the top 10 for any of them, and 45.86% when only the top three citations were counted. The failure case is comparing Surfer's 32% with seoClarity's 94%; the example of a fair comparison is Surfer's 54% top-three figure against Ahrefs' 76% top-three figure, same scope, different year and fan-out treatment.

The register below is the asset this page exists to hold. Each row is one study, its publication date as printed on the page, the sample it states, the figure most often quoted from it, and an evidence tier: vendor for a study by a company selling a related tool, independent for Pew, preprint for arXiv papers not shown as peer reviewed, and vendor documentation for Google. Eleven of the fifteen are vendor studies, and most of them leave at least one of market, device, period or denominator unstated.

Four studies put a number on YouTube's share of Overview citations and three of them agree. Authoritas in September 2024 found 6.46%. Ahrefs in March 2026 found YouTube was 5.6% of all cited URLs, and 18.2% of the cited pages that ranked nowhere in the top 100. The arXiv measurement across 55,393 trending queries in spring 2026 found 5.49%, with Facebook at 3.68%, Instagram at 3.65%, English Wikipedia at 4.39% and Reddit at 0.87%. Surfer's 2026 chart puts YouTube at 23.29%, ahead of wikipedia.org at 18.41% and google.com at 16.38%, among its twenty most-cited domains.

Surfer's figure is four times the others and the page does not explain the gap. The likely contributors are its keyword set, a chart built over the top twenty domains only, and the Gemini fan-out context of a 2026 crawl, but those are our inferences, not Surfer's statements. Until another study reproduces a share above 20%, it is an outlier to report, not a consensus to repeat. Pew's panel gives the broadest anchor: Wikipedia, YouTube and Reddit together accounted for 15% of the sources in the summaries it examined, against 17% of standard results.

Two studies describe the distribution rather than the head. The arXiv paper's May v1 reports that 56.3% of cited hostnames were cited exactly once (56.2% in the October v2), a long tail by any definition, and that user-generated content platforms made up 14.2% of citations against 41.4% of first-page URLs (49.9% in v2). Wellows, weighting by a link-based authority score over 3.62 million Overview citations, found a median cited domain authority of 47, with 40.8% of citations from domains under 40 and 23.8% from domains at 90 or above (across all five engines, just 301 DA 90+ domains took 17.0% of citations). "Concentrated" and "long-tailed" are both defensible descriptions of the same behaviour, depending on whether you count the head or the tail. The failure case is quoting Surfer's 23% as "YouTube gets a quarter of all AI Overview citations".

The schema question has more opinion attached to it than any other on this page and exactly one controlled test among these fifteen studies. Ahrefs' May 11, 2026 study tracked 1,885 pages that added JSON-LD between August 2025 and March 2026, each with more than 100 Overview citations in February 2025, before any schema was added, against matched control pages on other domains that never added schema. Citations were counted in the 30 days before and after each page's schema date. The matched difference-in-differences result: AI Overview citations fell 4.6% relative to controls, which Ahrefs calls small but statistically significant, while AI Mode rose 2.4% and ChatGPT 2.2%, both statistically indistinguishable from zero.

Ahrefs is careful about what that does and does not show. The decline is described as real but unexplained; the treated pages were already declining; the test covers only heavily cited pages, pools all schema types, uses a 30-day window and cannot separate schema from other changes made at the same time. The same study reports the correlation that drives the opinion: pages cited by AI systems are almost three times more likely to carry JSON-LD than uncited pages. Correlation and the controlled test point in different directions, which is the finding.

Google's documentation is shorter. Its page on AI features states that a page must be indexed and eligible to appear in Search with a snippet to be a supporting link in AI Overviews or AI Mode, that "There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary", and that "There's also no special schema.org structured data that you need to add". The example of the right reading: schema remains worth having for the rich results and machine-readability it was built for, which our schema markup guide covers. The failure case is selling it as an AI Overview citation lever when the one test found none.

Pew Research Center's July 22, 2025 analysis is the only independent behavioural measurement in this set. It used the browsing data of 900 US adults in March 2025, 68,879 unique Google searches of which 12,593 produced an AI summary, with the searches re-run in April by a third party to capture what the summaries cited. Users who saw a summary clicked a traditional result link in 8% of visits; users who did not clicked in 15%. About 1% of visits with a summary produced a click on a link inside the summary, and the browsing session ended on 26% of pages with a summary, against 16% of pages without.

The authors, Athena Chapekis and Anna Lieb, put the central result in one sentence: "Users who encountered an AI summary clicked on a traditional search result link in 8% of all visits. Those who did not encounter an AI summary clicked on a search result nearly twice as often (15% of visits)." The rest of the analysis is paraphrased above.

Ahrefs measured the same question with Search Console data rather than a panel. Its February 4, 2026 update compared 150,000 keywords with an Overview against 150,000 informational keywords without one, December 2023 against December 2025, desktop only. Position-one click-through on Overview keywords fell from 0.073 to 0.016, where a forecast from the no-Overview trend predicted 0.037; Ahrefs states the presence of an Overview correlates with a 58% lower click-through rate for the top result, with smaller gaps at lower positions. Semrush's December 2025 study, following a matched set of keywords before and after Overviews appeared, found zero-click share moved from 33.75% to 31.53% and concludes Overviews do not automatically raise zero-click behaviour.

The three agree more than they appear to. Pew and Ahrefs compare with and without an Overview and both find fewer clicks; Semrush follows the same keywords through time and finds no rise in zero-click share. A query that gains an Overview may lose clicks to the summary and still send the same share of users somewhere, because zero-click and click-through-to-position-one are different quantities. All three authors state their results as correlation. The example of a fair summary: where a with-and-without comparison has been run, clicks on results are lower when a summary is present. The failure case is "AI Overviews cut traffic by 58%", which turns one desktop CTR gap at position one into a general law.

Strip away the headline numbers and the studies agree on more than their authors' framing suggests. First, a large minority of cited pages do not rank in the top 10 for the exact query. Every study that measures it finds at least about a quarter outside, from 24% in Ahrefs' top-three 2025 sample to 67.82% in Surfer's all-citation 2026 sample, and the arXiv measurement finds 29.8% of cited domains absent from the entire first page. Second, ranking and citation are related: seoClarity's gradient runs from 43% at position one to 7% at position 20, Sanger's from 53% to 36.9% at position 10, and Ahrefs' 2025 median cited rank was 2 to 3.

Third, a typical Overview cites several sources: Pew found 88% of summaries cited three or more, Authoritas averaged 8.9 links, and our own division of the arXiv counts, 61,212 reference URLs across 7,583 Overviews, gives about 8.1 per Overview, a derived figure the paper does not state. Fourth, YouTube and Wikipedia appear prominently in every study that lists domains, though the share varies widely. Fifth, Google says nothing beyond indexing and snippet eligibility is required, and the one controlled schema test found no positive effect. Sixth, clicks on traditional results are lower where a with-and-without comparison has been run.

Seventh, and least quoted: every vendor study is a single snapshot with a commercial interest, most have at least one undefined or inconsistent method detail, and of the two academic pieces, one (2509.14436) is still a preprint while the other was accepted to IMC 2026 in October. That is not a reason to ignore them. It is the reason to quote them with their dates and denominators attached, which is what the register above is for. Teams building an AI-search content programme can take these constraints into our agentic SEO practice, where the measurement is done on the client's own queries rather than inferred from someone else's sample.

The true share of Overview citations from top-10 pages today is unknown. Published figures run from BrightEdge's 16.7% and Surfer's 32% to Ahrefs' 76.1%, with BrightEdge's own page also offering 54.5% overlap, and no two studies share a definition. Only seoClarity and Pew state market, device and date together. Which fan-out queries Google actually runs is inference in every study that invokes them; Surfer's page calls that a key limitation of all AI Overview studies, and it is right.

Nobody has explained why Surfer's YouTube share is four times everyone else's. None of these studies has tested whether schema affects pages that were not already being cited, or why Ahrefs' treated pages fell 4.6%. The freshness finding, that Overview top-three citations average about 1,432 days old and are 16 days older than organic results, rests on one Ahrefs study from July 2025 with internal text-and-table mismatches and has not been replicated. No study has established a causal link between Overviews and click loss; Pew and Ahrefs report correlations, Semrush finds no zero-click rise, and Google has disputed the methods, according to trade coverage rather than the Pew page itself.

AI Mode is almost unmeasured: Wellows' domain-authority split and Ahrefs' schema result are the only AI Mode rows in the set, and no ranking-overlap study of AI Mode was found. Nearly everything is US data; Wellows is 76.5% US and most vendors do not state a market. Our AI Overviews optimisation guide carries the practical advice that survives these gaps; this page carries the gaps.