Most-cited reference domains in AI search

Wikipedia, not averaged

The number people repeat is 47.9%. In Profound's window from August 2024 to June 2025, that is Wikipedia's share of ChatGPT's top 10. The same row is 7.8% of all ChatGPT citations. Reddit, in that table, is 1.8% of all citations and 11.3% of the top 10. Partial The corpus is vendor-reported. The query mix is not public. The arithmetic of the two denominators is public, and it is the part that gets dropped.

A second Profound cut, US English on ChatGPT.com from October to December 2025, is smaller and more specific: about 730,000 conversations that cited anything. Wikipedia was 5% of citations and appeared in 18% of those conversations. Reddit was 3% of citations and in 13% of conversations. The top 10 domains together were 12% of citations. A domain can be in a large minority of citing conversations and still be a small share of links, because a conversation cites several URLs.

Semrush's drop is easy to misread as Wikipedia disappearing. Both Wikipedia and Reddit stayed ChatGPT's two most-cited domains after the response rates fell. Forbes roughly doubled its ChatGPT response citations over those weeks. Medium, LinkedIn, and PR Newswire gained. TechRadar lost ground. Claude was not in that measurement. Response presence and link share answer different questions, which is why this row does not sit on the chart.

Country moves the encyclopedia more than the vendor does. On Ahrefs' German ChatGPT list for August 2026, de.wikipedia.org is 18.67% of the top-50 citation sum, and en.wikipedia.org is still there at 4.65%, rank 5. The US September list is not the German list. The engine pages keep the country cuts: ChatGPT, Perplexity, and the rest.

One vendor, one month, six engines

The chart is the only place on this page where Wikipedia is compared across engines, and only because the method matches. Ahrefs' US pages, updated 2 September 2026, each report mention share of that engine's own top sources.

Read the bars as six separate pools. The Gemini app's 8.8% is not "more Wikipedia than AI Mode" in a count of every link on the internet. It is a larger slice of the Gemini app's published head. On that list Wikipedia is third, behind Reddit at 28.5% and YouTube at 14.6%. On AI Overviews and AI Mode it is seventh and sixth, behind the platforms. On Copilot it is third, behind Amazon and Walmart, which tells you the reference layer can rank high on an engine whose head is retail.

Perplexity is the awkward one if you leave this chart. Wikipedia is absent from Profound's Perplexity top 10 through June 2025, where the head is Reddit at 6.6% of all citations. It is third in the Ahrefs US September pool, at 6.3%. Both can be true. They are not a trend line, and they are not averaged on the Perplexity list.

Dictionaries, .gov, and Britannica

The rest of the measurable reference layer, on US ChatGPT in September 2026, is dictionaries plus a handful of government and medical domains. Same denominator as Wikipedia's 7.0%: share of the top-source pool, not of all citations.

Against the prior month on that page, Merriam-Webster fell 9 places, Cambridge 13, Dictionary.com 24, and Collins 33. A one-month arrow is not a trend. Germany's equivalent is louder: duden.de is 5.19% of the August top-50 sum, rank 4, and de.wiktionary.org is rank 32 at 1.01%. A dictionary is a national habit as much as a domain strategy.

The .gov rows in the same US ChatGPT pool are PubMed Central at 2.4% (rank 9), PubMed at 1.8% (rank 17), FDA at 1.3% (rank 29), MedlinePlus at 1.1% (rank 40), and CDC at 0.9% (rank 47). Do not add them. They share a denominator, they can share answers, and anything outside the top 50 is missing, so a sum would be neither a .gov total nor a share of all citations. Similarweb, via 5W, puts nih.gov at 2.22% of about 600,000 US citation events in January and February 2026. That is a different cut again.

Britannica is the absence. It is not in the US ChatGPT top 50. On Copilot, britannica.com is rank 48 at 0.8% of that top-50 pool. WikiHow, which is a how-to site more than a reference work, is rank 13 on Copilot at 1.8%. Three reference-shaped names sit in Copilot's top 10 beside Wikipedia: scienceinsights.org at 3.5%, grokipedia.com at 3.1%, and biologyinsights.com at 2.5%. Those are rows. They are not a measured replacement for the encyclopedia, and this page does not grade the sites.

Wikidata is not in the ChatGPT top 50. The 1.93% wikimedia.org row in the Similarweb cut is a different host. No opened study isolates wikidata.org. No opened study reports a share for the .edu TLD, or for ISO, W3C, NIST, or IETF, on any engine. Open question on those names.

Claude has a class, not a name

There is still no public large-sample domain ranking for Claude. The Claude list is categories. Otterly's June 2026 file, read from indexed text after the page returned 403, puts Encyclopedia at 2.1% of 379,321 citations across 142 domains. Government and NGO is 2.4% across 596 domains. Education is 2.7% across 518 domains, and that is a category label, not the .edu TLD.

Profound's institution bucket for Claude is 7%. The definition includes Wikipedia, universities, and professional associations. It is not a government share and it is not a Wikipedia share. Quoting it as either is how the 7% escapes.

SoRank is the only model-split answer rate opened here. Wikipedia appeared in 0.03% of Claude answers. The Claude cell is 20,011 answers inside a panel that is 79% French and built from buyer questions. The authors say to read it as direction. Partial A buyer-intent set under-represents encyclopedias. A low rate on that panel does not prove Claude skips Wikipedia on a general-knowledge question.

When there is no article

Open question No study opened for this page takes topics or entities with no Wikipedia article and ranks the domains that receive the citation instead. A sentence in a synthesis that says the model "fills the gap, often incorrectly" has no substitute-domain table under it.

The closest measurement is Green Flag Digital's re-run on 21 September 2026. US personal-finance ChatGPT, via Ahrefs Brand Radar, limited to pages cited beside seven anchor domains. In the top 500 cited pages, Wikipedia and Reddit were 17.9% of citations on 23 October 2025 and 0% on 21 September 2026. Banks, card issuers, and airlines went from 1.8% to 22.4%. Government and regulators went from 0% to 5.3%. Of 41 tracked Wikipedia pages, 76% were at zero and the group kept 1% of its citations. The author says ChatGPT still cites Wikipedia heavily in general, and that part of the first-party rise is built in because two of the anchors are banks. This is Wikipedia leaving a finance set. It is not a missing-article test.

The protocol for that test is already registered, and it has no results: what ChatGPT cites when Wikipedia is silent. Treating this list as the answer would be the thing the protocol exists to avoid.

What this does not prove

A mention share is not the share of answers copied from Wikipedia. Training influence is not in any of these panels. A reference domain can shape an uncited answer and still be rare in a footnote list, or the reverse. The .gov rows are not a government total. Claude's 2.1% encyclopedia class is not a rank for en.wikipedia.org. And a finance vertical in which Wikipedia's tracked pages went quiet does not tell you what happens for a person, a product, or a local business that never had an article.

How to use the list

If you quote one Wikipedia number, quote the denominator in the same sentence. If the decision is whether a dictionary page or a government page can be cited at all, the US ChatGPT list says yes, at around 1% to 2% of the top-source pool, with Reddit still larger than Wikipedia. If the plan is to rank in the hole where Wikipedia has nothing, that experiment has not been published.

The series argument is on most-cited domains in AI search. The ChatGPT statistics page is the mechanism, not the domain table: ChatGPT citation statistics.

Verification status