AI citation sources

Profound analysed 680 million citations collected between August 2024 and June 2025. Wikipedia was ChatGPT's most-cited single domain, at 7.8% of its citations^[1]. Reddit led on Perplexity with 6.6% of citations^[2] and on AI Overviews with 2.2%, a much flatter distribution than ChatGPT's^[3]. Across the platforms measured, .com domains took 80.41% of all citations^[6].

A May 2026 audit by 5W put Wikipedia and Reddit together above a quarter of ChatGPT's US citations^[7], while the leading source on Google AI Mode was the wiki host Fandom at 7.16%^[8]. The Wall Street Journal, The New York Times, Bloomberg and the Financial Times appeared nowhere in that top 20^[9], despite the domain authority those titles carry in conventional search.

Retrieval and attribution are separate steps, and the gap between them is wide. In an Ahrefs sample of 1.4 million ChatGPT prompts, Reddit pages made up 67.8% of the URLs the model fetched during research and then left out of the answer^[10]. Pages that did earn a citation had a median age of roughly 500 days^[11], which sits awkwardly beside the assumption that freshness drives selection.

Domain-share tables record who gets cited, not why. A controlled experiment separates the two: 252,000 paired trials across six models, injecting exactly two candidate sources into the context and varying one content factor per trial, found topical relevance and list position the largest drivers of which source the first citation marker points at, while formatting-only edits moved little^[12]. The same run found an explicit price and a recent timestamp helping consistently, which qualifies the 500-day median age above: that median describes what engines happen to retrieve, whereas the experiment describes which of two comparable candidates wins.

Citation lists and organic rankings are only loosely coupled. Ahrefs ran 15,000 long-tail queries in July 2025 and found that 12% of the URLs cited by ChatGPT, Gemini and Copilot also ranked in Google's top 10 for the same prompt, while about 80% ranked nowhere in Google's top 100 for it^[4]. Engines differ widely inside that average: Perplexity overlapped 28.6%, against 8.6% for Gemini and 8.0% for ChatGPT in-text citations^[13].

A Semrush study of more than 150,000 citations across 5,000 keywords, published the same month, reported Perplexity domain overlap above 91% and ChatGPT the lowest of the platforms tested, tracking Bing results more closely than Google's^[14]. The two studies count different units, URLs against domains, and what they agree on is the ordering of the engines.

Per-source share moves on a timescale organic rankings do not. 5W recorded Reddit's share of ChatGPT citations falling from roughly 60% to roughly 10% inside two weeks in September 2025^[5]. Any single reading is a snapshot of one prompt sample on one date, so visibility measurement needs repeated sampling rather than a one-off audit, and a domain-level ranking published months ago should be treated as history.

A link in an answer is not the same as a correct attribution. Human evaluation across the extractive-abstractive spectrum finds that language models, unlike search engines that return snippets with links to the page they came from, do not provide reliable citation to sources, which makes the origin of an answer hard to trace^[15]. The Tow Center's audit puts a figure on one engine: DeepSeek named the wrong source for the excerpt it had been given in 115 of 200 queries^[16]. See citation accuracy.