Three in four AI citation drops were temporary during a 20-day study by Ahrefs involving five different AI products. Mateusz Makosiewicz, an SEO and marketing educator at Ahrefs, and his colleague Xibeijia Guan collected all websites that had been cited on day one, then traced each reference through to day 20. They studied about 37,500 first-day citations made by Google AI Mode, AI Overviews, ChatGPT, Gemini, and Copilot. The study included only one organization’s view based on its own group of questions answered daily by each tool for 20 days. About half of the citations that dropped across the five tools came back within three days. I find the volatility of these tools makes immediate reaction to a lost link a mistake for any SEO.
The difference between empty answers and competitor takeovers
A lost citation does not always mean a competitor stole your spot. In Gemini, responses lacked citations 64% of the time in the data from this study. Copilot showed answers without citations around 53% of the time. When these tools stop citing, they often cite no one at all. In contrast, AI Overviews, AI Mode, and ChatGPT almost always cited sources, with fewer than 1% of their answers having no citations. Makosiewicz wrote that a lost citation on those tools almost always means another site took the spot. You should check if the AI tool provides a link to a competitor or if it simply provides no link.
I would prioritize tracking the frequency of citations over single instances. The study showed that about half of the references in AI Overviews disappeared by day two. Most of these returned before the study ended. One measure from Brand Radar, which Ahrefs owns, showed a different pattern where about two-thirds of the disappeared citations ultimately came back.
| AI Tool | Frequency of No Citations |
|---|---|
| Gemini | 64% |
| Copilot | 53% |
| AI Overviews | < 1% |
| AI Mode | < 1% |
| ChatGPT | < 1% |
How citation history and brand mentions dictate edits
The length of time a citation remained active before it disappeared determines if a site needs to change its content. The study ranked all citations on three of the tools in order of how long they had been there until they went away the first time. Citations that stayed active for eight to 14 days before going away had a 4% or lower chance of not returning. For citations with a one-day lifespan, 34% to 41% of them did not return during the study period. A site with a long history of being cited is more stable than a site that only appeared once. A site referenced once that then disappears is a different circumstance than a site that has maintained its position for 10 days and disappeared yesterday.
You should check if the brand name still appears in the text even if the link is gone. A citation is a link, but a mention is a brand name in the answer text. Each tool saw 50% to 70% of all cited sites appear on only one day, while only about one in five brand names appeared on only one day. AI tools use different versions of the same brand name from answer to answer. A mention that seems to be gone could be the same company referred to by a different name. Because brand names do not move as much as citations, Makosiewicz suggests that mentions may be the better metric to follow.
I recommend waiting two or three weeks before you decide to edit a page because a single comeback does not prove an edit worked. On AI Overviews and AI Mode, cited pages that ranked in Google Search for the same question kept their citations longer. This finding came from an examination of 230 questions, although most pages cited by ChatGPT and AI Mode did not rank in Google for the question. While appearing in Google results correlates with longer-lasting citations, ranking position does not have the same impact. The researchers measured the disappearance of citations, but they did not test if changing a page brings back a citation that has gone missing, which means one comeback alone does not prove an edit worked.
Manage expectations with frequency counts
An agency should follow a strict order before proposing new content to a client. A team should check the text of the response, the history of the citation, and the ranking of the page prior to making a proposal. Management learns more from a count of how many of the last seven days a site was cited than from a yes or no report. One comeback after an edit provides weak evidence that the change worked, especially since half or more of citations disappeared by the next day in the study.
A content team must recognize the difference between correlation and causation. If a citation disappears the very next day after you edit a page, the edit did not cause the loss. You should keep a daily log of which sites are cited for the questions that matter most to your business. This log helps your team provide a reliable internal tally of days cited before a manager asks why performance is down. If a citation remains missing after three weeks, you should examine the page and how it ranks for the question.

