Most-cited pages in AI search, not domains

The URL list that was not published

Domain lists are everywhere. Page lists are not. Louise Linehan's Ahrefs piece, reviewed by Ryan Law and published 28 October 2025, is the one that gets closest, and it stops at type. She exported the top 1,000 pages ChatGPT cited in September 2025 from Brand Radar and categorised them with Claude. The article does not print the 1,000 URLs and does not print a top 10 URL table. Partial The classifier is a model. The export is a vendor panel. The refusal to invent the missing URLs is the point of this page.

That is also why a "most-cited pages" slideshow with ten blue links and no shares is not a measurement. If the shares existed in a fixed query set, they would be in the table below. They do not. The domain lists in this series stay on domains: ChatGPT and the hub.

What the top 1,000 pages were

The rows sum to 100. Linehan calls 32.3% influenceable, which is educational plus reviews plus news plus blog. That is her sum, and it is a fair description of the pages a publisher could plausibly be. It is not evidence that those pages are easier to get cited, and it is not 32.3% of citations. Wikipedia, homepages, app stores, and dictionaries are the rest, and they are most of the thousand.

Of the 1,000 pages, 717 had organic keywords, a median of 279 keywords. The other 283 had none. Among pages that did rank, the median Domain Rating was 90, 65.3% were DR 81 or higher, and 11.7% were DR 0 to 20. URL Rating was low: 67.3% sat between 0 and 10, with a median of 6, and the median referring-domain count was 70. Just over half, 52.1%, ranked in the top 3 for a keyword. Publish dates were found for only 12.6% of the thousand. Of that dated slice, 60.5% were published within the prior two years. Update metadata existed for just over half, and of that dated set 89.7% showed a 2025 update. Dropping Wikipedia, about 82% of the remainder still did.

The article also shows individual pages with large citation counts and no current organic keywords, including one unnamed page at 4,600 ChatGPT citations. The URLs are in screenshots, not in the text. Quoting the counts without the URLs would pretend this page has a ranking the source withheld. The pattern is the finding. A page can be heavily cited and invisible in organic keywords. The reverse is the ordinary SEO chart, and it is a different chart.

A thread, a video, not a homepage

Profound looked at the URL under the domain, for social platforms only. About 700,000 social-platform citations in ChatGPT, October to December 2025, US English. Reddit was 2.4% of all citations in that window, YouTube 0.99%, LinkedIn 0.39%. The structural line is the one that matters here: 99% of Reddit citations were individual threads, not subreddit pages or profiles. When YouTube was cited, it was a specific video 85% of the time. The post does not list the threads or the videos.

That splits the work. A domain strategy that stops at reddit.com or youtube.com has not named the unit the model cites. The community and video lists carry the domain shares: community domains and video domains. This page only insists that the citation is a page.

Q&A and forum pages are 0.9% of Ahrefs' thousand, while Reddit leads several domain lists. Both can be true. A thread is one URL. A domain list rolls thousands of threads into one host. The page classification and the domain classification are not in conflict until someone adds them.

How long the page keeps the citation

Profound, Davis McCain, 30 September 2026, measured duration rather than identity. Across 883,000 pages and seven engines, including ChatGPT and Copilot, from 8 September 2025 to 8 September 2026, the median page lost half its citation share within 11 days of peaking. 78% of cited pages were at half their peak within two weeks. The sample was a deterministic 1-in-128 draw of cited URLs. No URL ranking. Partial This is their panel. The study registered on this site, AI citation half-life, is a protocol. It has not produced this number, and this page does not borrow it as if it had.

A repeated page set, rather than a median across the whole web, comes from Green Flag's personal-finance panel. Of the 250 most-cited finance pages in October 2025, 35 NerdWallet URLs were set aside because of redirects, leaving 215. By September 2026, 107 of 215 had zero ChatGPT citations. The group's citations fell from 23,251 to 3,271. The set is pages co-cited with finance anchors. It is not the ChatGPT corpus. It is the one public before-and-after on the same URLs, and it says half of them were gone.

Page concentration, which is a different question from "which URL," is on the long tail. Resonance's line that the top 1% of cited pages earn 30% of citations lives there, because it is a distribution, not a list of addresses.

What this does not prove

Open question It does not name the most-cited URL in any engine. It does not say homepages are 23.8% of citations. It does not say a page with no organic keywords will be cited, only that hundreds of heavily cited pages had none at the time of one export. The 11-day median is not a promise that your URL expires on day 12. It is the median of pages that were cited and then tracked inside one sample. Pages that are never cited do not enter that clock.

How to use it

Stop asking a domain report which page to write. The public evidence says the cited object is a specific URL, often a reference page, an educational page, a thread, or a video, and that the head of that list turns over. If you need the actual URLs for a fixed query set, they have to be collected. Nobody's published table will supply them.

The next honest step is a fixed prompt panel with the URL written down, which is the design behind the half-life protocol. Until that runs, the classification above is the page-level fact, and the domain pages remain domain pages.

Verification status