Blog · 24 September 2026 · By HeardOf

Do AI answers cite the pages that rank? Every published overlap figure, and what each counted

Do AI Overviews cite top ranking pages?

By the count of citations, a minority do: 37.9% of the 4 million URLs Ahrefs found cited in AI Overviews were in the first ten result blocks for the same query, in its update of 2 March 2026, and about 17% of the sources BrightEdge tracked were in the organic top 10, in its report of 12 February 2026, both historical. By the count of answers, most do: 90% of the 362,000 AI Overviews seoClarity read on 12 October 2025 cited at least one top-10 page, and 94% at least one top-20 page, in its study dated October 2025, historical. Both are true at once: an AI Overview that cites one ranked page among several counts once for the answer and once among its citations.

The buyer's question, do AI Overviews cite top ranking pages, therefore has no single figure: the headline shares in the two tables below, from pages read on 24 September 2026, run from 6.82% to over 99%, each correct on its own unit. The two tables keep each figure beside its unit, its sample, the top-N it was compared against and the date on its page. Fourteen of the fifteen are more than 90 days old at the read and are labelled historical; hzsignal's reconciliation of 3 July 2026 is not. Surfer's page carries 17 September 2026, but the Wayback Machine's copy of 11 December 2025 carries the same figures and a published date of 30 November 2025 in its structured data, the date hzsignal's table gives, with a byline of 28 November 2025.

Why do the figures in these tables run from 6.82% to over 99%?

Because they count different things, and hzsignal's page of 3 July 2026 names four of the differences: citations against answers, all citations against the three visible ones, rankings for the typed query against rankings for its fan-out queries, and the keyword set and date. seoClarity's study of October 2025, historical, prints the first pair on one page: 56% of citations from the top 20, and 94% of AI Overviews with at least one top-20 citation. BrightEdge's page of 18 September 2025, historical, prints two thresholds on one page: 54.5% of citations ranking organically, 16.7% in the top 10.

The second difference is inside Ahrefs' own pair: the July 2025 study read the three most visible citations per AI Overview and found 76.10% in the top 10; the March 2026 update finds 37.9% of cited URLs after Ahrefs, in the page's words, "improved our parsing methodology"; Ahrefs sets the two side by side, and hzsignal's reconciliation of 3 July 2026 calls them not directly comparable. Surfer measured the same gap on one dataset: 67.82% of all cited pages ranked in none of the top 10s it pulled, against 45.86% of the top three citations. The third is which rankings: Surfer compared against the top 10 for the query and for fan-out queries generated with Gemini, and states that it pulled no positions beyond 10, so a page at position 12 counts as unranked; Ahrefs' March 2026 page compared against result blocks including features, then against blue links alone, and prints both.

The 99% is a retelling. seoClarity's impact page, last updated 23 October 2025, historical, prints ">99% of the instances of AI Overviews were sourced from the top 10 web results", a per-answer figure with no sample or date beside it on the page. hzsignal writes that its own AI Overviews guide had repeated it as "over 99% of citations" and corrected the guide in the release that published the reconciliation.

Does ChatGPT cite the pages that rank in Bing or in Google?

Bing's, in Seer Interactive's post of 6 February 2025, historical: 87%+ of ChatGPT's search citations matched Bing's top organic results when the same question was used as the query, and 56% matched Google's. That is 500+ citations from 100 queries, matched as exact URLs after lowercasing, trimming and dropping parameters, with "most of those results appearing in the top 10 positions" on Bing and "many citations ranking in positions 11-20"; on Google the matched results had a median rank of 17. Seer writes of "a limited sample of 100 queries".

Ahrefs' long-tail study of 11 August 2025, historical, also matched against both engines and counts a narrower unit, a match in the top 10 rather than anywhere in the top organic results: ChatGPT's in-text citations were at 8.00% for Google's top 10 and 8.10% for Bing's, across 15,000 prompts on data the page says was collected "in early July", with Perplexity at 28.6% for Google. Ahrefs' two later pages, of 20 August and 3 September 2025, measure Google only: 6.82% of the results ChatGPT's own 118,931 fan-out searches returned were in Google's top 10 for the same search, and 10% of ChatGPT's cited URLs for 3,311 short-tail terms were, against 31.8% of its cited domains and 65% of Perplexity's URLs. Ahrefs' summary across the three is that ChatGPT "doesn't follow Google's sources, Perplexity does".

Our own run of 4 September 2026, in who the AI actually cites, sits beside these and not among them: it counted which pages ChatGPT and Perplexity cited for 20 buyer prompts, 173 and 391 citations, and later compared the cited domains against Tranco's list of 23 September 2026, not against a search results page, so it holds no overlap figure of this kind.

Is the overlap falling?

Two publishers print a fall and one prints a flat line, on three different counts. Ahrefs' pair went from 76.10% in July 2025 to 37.9% in March 2026. The March page reads the drop as AI Overviews, on Gemini 3 since January 2026, "relying less on the direct search results and more on the sources showing up in fan out query SERPs", and separately says Ahrefs has "improved our parsing methodology" since the first study; hzsignal's reconciliation of 3 July 2026 adds that how much of the drop is parsing cannot be separated. seoClarity's research page, comparing October 2025 with May 2025, prints an average of 3 top-20 URLs per AI Overview, down from 5, and 6% of keywords with full top-10 overlap, down from 19%. BrightEdge's report of 12 February 2026 prints a top-10 overlap that "has been remarkably flat" at about 16.1% to 16.9% a month from February to July 2025 while its top-100 overlap rose from 48.7% to 53.1%; its page of 18 September 2025 prints the wider overlap rising from 32.3% in May 2024 to 54.5%, with 16.7% from the top 10. Each of the three pages names its publisher's own AI search tool: Ahrefs Brand Radar, Clarity ArcAI on seoClarity's research page, and BrightEdge AI Catalyst on BrightEdge's February 2026 page.

What do the pages that relay these figures print?

Two of the four print a figure or a unit that differs from the page they relay. Ahrefs' statistics roundup, "Updated: November 10, 2025" on the page and modified 31 May 2026 in its structured data, historical, prints "76% of AI Overview citations are pulled from the pages, ranking in Google's top 10 organic results" and links the July 2025 study; the update of 2 March 2026, on the same blog, is not on the roundup as read. The roundup also prints "20.7% of citations for ChatGPT short-tail prompts are in Google's top 100 results" against the September 2025 study, whose text prints the 10% and the 31.8% and not the 20.7%.

ALM Corp's post of 3 March 2026 on the drop from 76% to 38%, historical, relays Ahrefs' 38% and 37.1% and BrightEdge's approximately 17% as their pages print them, and writes of "October 2025 (BrightEdge): BrightEdge reported a 54% overlap between AI Overview citations and organic top-10 rankings"; BrightEdge's page of 18 September 2025 prints 54.5% as citations that "rank organically" and 16.7% as the share from the top 10. ALM Corp's other post of the same day, on the surge across nine industries, relays BrightEdge's about 17% and 49% to 53% as the February 2026 page prints them. The first post also relays "SeoClarity's analysis of 432,000 keywords found that 97% of AI Overviews cite at least one source from the top 20", which is on seoClarity's impact page; that page points readers to "our full study on the overlap between AI Overviews and organic rankings", and the research page of that title prints 362,000 keywords and 94% for the same unit. What that second ALM Corp post says about how often AI Overviews appear by industry is checked against BrightEdge's own pages, beside the other published presence figures, in how often Google shows an AI Overview.

hzsignal's table of 3 July 2026 was checked figure by figure against the nine primaries it links that were opened here, Ahrefs' three, BrightEdge's, seoClarity's three, Originality.AI's and Surfer's, and each figure matched its page. So did every date: Surfer's page now carries 17 September 2026, but its archived copy of 11 December 2025 carries hzsignal's 30 November 2025 in its structured data. Its note that BrightEdge "doesn't publish whether it matches at URL or domain level" holds for the two BrightEdge pages as read, which say "sources" and "citations" and neither.

How was this compiled?

The nineteen pages, the fifteen in the two tables, Originality.AI's companion methodology page and the three relaying pages that are not in a table, were fetched with curl and a browser user-agent between 22:47 and 22:51 UTC on 24 September 2026 and rendered in Chrome 153 headless with JavaScript, a nine-second wait and two scrolls, between 22:50 and 22:56 UTC. Every figure printed here was found in both the served and the rendered text of its page. Dates are the page's own; where a structured-data date is used, it is named as such. Historical means dated, or shown by an archived copy to have been published, before 26 June 2026.

What was not done: no study was re-run and no figure was recomputed from anything but its page. hzsignal's one derived number, that Originality.AI's 52% of the 48% is about 25% of all citations, is hzsignal's arithmetic and is reported as such. ALM Corp's "40% of sources shown in AI Overviews would rank in positions 11 to 20", attributed to Ahrefs, was not traced to an Ahrefs page and is not in the tables. Search Engine Land's report on Surfer's fan-out data, which hzsignal cites for 161%, answered 403 to this read and is not used. No study page links a spreadsheet or dataset of its citations, by a sweep of every link on the fifteen pages; BrightEdge's two pages each link a PDF report, not opened here. Whether any of these pages ranks for anything is not on record. Everything this post says about our own run comes from us, carries its date in the sentence that says it and cannot be checked from outside; everything else carries a page, a byline and a date you can follow.

How often does Google show an AI Overview? The published figures, with what each measured