Ahrefs tracked 37,500 AI citations for 20 days and half vanished each day

Most AI citation tracking is measuring noise. Ahrefs says it tracked 37,500 citations across five AI engines for 20 days. On most engines, half or more of the cited pages were gone the next day.

Ahrefs published the study on 29 October 2026. Mateusz Makosiewicz wrote it, with Xibeijia Guan. The team queried each engine daily with the same prompt set. They recorded which pages were cited each run.

The disappearance rates by engine

Ahrefs reports that AI Overviews dropped about 50 percent of its cited pages between runs. Google AI Mode and ChatGPT dropped 58 to 60 percent. Gemini and Copilot dropped 80 to 83 percent. Ahrefs says Gemini returned no citations at all on 64 percent of its answers. Copilot did so on 53 percent.

Most lost citations came back

Ahrefs says about 75 percent of lost citations returned at some point in the 20 days. On the two Google tools, 40 percent returned the next day. Older citations were steadier. A page cited for 8 to 14 days before vanishing came back 96 to 99 percent of the time. A page cited only once had a 34 to 41 percent chance of never coming back.

What held a citation in place

Ahrefs says ranking in classical Google search is the strongest signal. Pages that rank in Google kept their AI Overviews citation seven times longer than pages that did not. On Google tools, long-lasting citations matched the query in the page title more often. The title-match score was 0.65 to 0.68 for them, against 0.54 for one-day citations.

Brand-named queries behave differently. On about 30 percent of these queries, the brand's own site stayed cited every day. The brand's site appeared on 60 to 68 percent of days overall. Competitors on the same queries appeared on 16 to 20 percent of days. Factual questions cited official sources 97 to 100 percent of the time. Comparison and "best X" prompts changed their cited pages on 75 percent of days.

What the study tells a reader to do

Ahrefs recommends looking at the last 7 days of citations rather than today. It also recommends waiting 2 to 3 weeks before deciding that a page edit changed anything. Brand-factual queries should be reported as a separate group from reviews and comparisons.

The study is Ahrefs' own and uses its own tooling. Reveneau has not audited the sample or the engines' answer logs. Read every figure as what Ahrefs says it saw this October.

A single-day check reports today's noise. For a team showing AI visibility to a client, that is below the floor the data supports. The weekly rolling window Ahrefs proposes is the smallest window this study backs.

Source

How stable are AI citations? 10 things we learned tracking 37,500 citations, Ahrefs.