AI Citations

An AI citation is a source reference, usually a link, that an AI answer engine attaches to a generated answer to show which pages it retrieved and drew on. Citations are the countable unit of visibility in AI search: they are what measurement tools tally and what generative engine optimization tries to win. The 2026 research adds an important qualification. Being cited and actually shaping what the answer says are two different outcomes, and neither is stable from one day to the next.

Selection Versus Absorption

Zhang, He and Yao (arXiv, April 2026) analysed 21,143 citations and separated two stages. Selection is getting into the list of sources. Absorption is how much of the answer's content traces back to a given page. The engines differ sharply on both, and the two move in opposite directions: the more sources an engine lists, the less each one contributes.

A citation in Perplexity is therefore easier to get and worth less of the answer than a citation in ChatGPT. The same study found that page type matters more than formatting: encyclopedic pages scored 0.21 on influence against 0.07 for news, while question-and-answer formatting on its own made no measurable difference (0.0947 versus 0.1005). A citation count alone does not say whether a page informed the answer or merely decorated it.

Where Citations Come From

Citations no longer mirror the ranked results. Ahrefs (March 2026) found that 37.9% of URLs cited in Google AI Overviews rank in the organic top 10, down from roughly 76% in July 2025; BrightEdge (February 2026), using a different method, puts the overlap near 17%. The usual explanation is query fan-out, in which the engine runs several sub-queries and cites pages that recur across them.

The distribution is heavily concentrated. Żatuchin (arXiv, June 2026; about 168,000 citations across 12 markets) found that roughly 80% of citations come from about 18% of domains, that Wikipedia is the top source in 11 of 12 languages, and that 85.7% of citations point to sites the brand does not own. Ahrefs' September 2026 rankings show the same few platforms leading: YouTube (22.9%) and Reddit (18.5%) among the top 50 sources in AI Overviews, Reddit (16.8%) and Wikipedia (7.0%) in ChatGPT, and Reddit (21.6%) and YouTube (20.8%) in Perplexity. Concentration at the platform level coexists with very low agreement at the page level. Grossman et al. (SIGIR 2026) measured cross-platform source overlap below 0.2 on the Jaccard index, and a small vendor-authored study (Tannenbaum, September 2026; 589 citations) found 96.4% of cited URLs appearing in only one of four engines.

Accuracy

A citation is not proof that the source supports the claim. Xu, Iqbal and Montgomery (arXiv, May 2026) checked about 98,000 atomic claims in AI Overviews and found 11.0% unsupported by the pages cited for them. Onweller et al. (May 2026) found deep-research agents producing more than 94% valid links while factual accuracy ranged from 39% to 77%. A working link and a faithful summary are separate properties, which is why citations reduce but do not remove the problem of AI hallucinations. For a brand, this means a citation can attach its name to a statement its page never made.

Volatility

Citations churn constantly. Schulte, Bleeker and Kaufmann (arXiv, April 2026) measured day-to-day Jaccard similarity of cited sources at 0.34 to 0.42, meaning roughly 65% of sources change daily. Kirsten et al. (ACL Findings 2026) found two AI Overview runs two months apart shared only 18% of pages, against 45% for organic results. Zhen et al. fitted citation half-lives of about 39 days for time-sensitive queries and 68 days for others.

Platform-level shifts can be abrupt and unannounced. Otterly's September 2026 index recorded YouTube's share of all Google AI Mode citations falling from 13.5% to under 4% within about two weeks. Provisional Promptwatch data reported by Semrush showed Reddit's share of all ChatGPT citations dropping from 3.8% to 0.5% in mid-August 2026 (a different denominator from the Ahrefs figure above, so the two are not comparable). The practical consequence is that any single citation check is a sample from a distribution. Every figure on this page should be read with its date.