Ahrefs: AI citations drop the next day, and most return within 20 days
In a study dated October 29, 2026, Ahrefs followed about 37,500 citations on five AI tools with a daily check over 20 days. At least half were gone by the next day, yet roughly 75 percent of the dropped ones came back during the study. Brand mentions proved steadier than citations. Ahrefs presents the findings as correlations from one month of data.
Ahrefs checked five AI tools daily across a 20-day window
Ahrefs published a study on October 29, 2026 that measures how durable AI citations are. The article is bylined to Mateusz Makosiewicz and reviewed by Ryan Law, while the data work belongs to Xibeijia Guan. Guan asked five AI tools the same set of questions daily over a 20-day window, followed each website that earned a citation on the first day, and checked whether it stayed cited the next day or, after dropping out, ever returned. Ahrefs says it followed about 37,500 citations in total.
Day-one counts of cited websites by tool: Google AI Mode 18,009 across 1,996 questions, ChatGPT 11,225 across 3,173, Google AI Overviews 4,830 across 563, Gemini 1,994 across 1,276 and Copilot 1,447 across 823. Ahrefs also re-ran the measurements on two sets of questions it tracks daily in its Brand Radar product, 25 and 69 in size. Ahrefs sells SEO tools, so it is the author of the study as well as a vendor in this field. Every figure below is Ahrefs' own.
More than half of citations drop by the next day, and most come back
According to Ahrefs, at least half of citations were gone by the next day: roughly 60 percent on Google AI Mode and ChatGPT, 50 percent on AI Overviews and about 80 percent on Gemini and Copilot. Over half of the answers from Gemini and Copilot listed no sources whatsoever (64 percent and 53 percent), which makes those two tools look more volatile than they are. With empty answers set aside, losses on all five tools moved toward the 50 to 60 percent range. About 75 percent of the citations that dropped came back within the 20 days, around half of them within three. Only 20 to 27 percent of the websites cited on day one disappeared for good during the study.
How long a citation had lasted matters. Among citations that appeared for a single day, 34 to 41 percent never returned on AI Overviews, AI Mode and ChatGPT; for citations that had lasted 8 to 14 days the share was 0.7 to 4 percent. Brand names also outlasted links: 54 to 84 percent of cited websites were not named in that answer, and while 50 to 70 percent of cited sites showed up on only one day, the figure for brand names was about one in five. That is why Ahrefs recommends tracking brand mentions separately from citations.
A single check misleads, and Google rankings travel with durability
The takeaway is that earning or losing a citation on one day says little. Ahrefs suggests tallying, for each tracked question, the days of the past week on which a page earned a citation, and reading a drop against how long the citation had held beforehand. Pages that rank in Google for the same question were cited for at least seven days at a clearly higher rate: 29 percent versus 4 on AI Overviews, 28 versus 4 on AI Mode and 11 versus 4 on ChatGPT. On AI Overviews, the median page that held its citation for 15 days or longer drew 88 estimated organic visits a month, against 8 for pages cited on a single day, while Domain Rating and backlinks made little difference. Ahrefs scored how closely a page's title and description match the question; the median was 0.54 for single-day citations and 0.65 to 0.68 for those lasting over 15 days on AI Overviews and AI Mode, with a much weaker link on ChatGPT.
Question type matters too. For questions that name a brand, the brand's own site stayed cited for all 20 days in 28 to 33 percent of cases on Google's tools and ChatGPT, versus 4 to 6.8 percent for a typical cited site. In Ahrefs' own test, factual questions such as pricing almost always cited its site, while on reviews and comparisons Google's tools leaned on third parties. On ChatGPT and AI Mode, 64 to 76 percent of cited pages were swapped out daily whatever the question type; questions asking for the "best" of a category were among the least stable and brand-named questions among the most stable. The authors stress that these are correlations from a month of data, that the facts-versus-reviews result rests on 22 questions about one brand, and that the ranking analysis covers only 230 questions.
The page says nothing about Türkiye, so this section is commentary
We found no information about Türkiye on the Ahrefs page, and no statement on the language or country in which the questions were asked. This section is therefore commentary from UNALSOFT, not a fact drawn from the source. Whether the figures would hold for Turkish-language questions, for user behavior in Türkiye or for local results is unknown, and the study does not show it.
Our reading: for a business reporting its AI visibility in Türkiye, the practical lesson here is about method, not about the numbers. Looking at a few weeks of trend instead of a single screenshot, and tracking brand mentions separately from citations, are habits that hold even if the figures differ on local data. We covered ChatGPT citations in a separate October 18 story, which is independent of this study.
The UNALSOFT view
We read this study as a lesson in measurement. Reporting AI visibility from one query on one day can mislead; Ahrefs recommends watching the trend over time rather than a single check, and tracking brand mentions apart from citations. The same study links longer-lasting citations to pages that rank in Google and match the question, and clear, accurate pricing, plan and feature pages fit with a brand's own site being cited on factual questions. On the web design side we treat aligning each page title and description with the question people ask as its own task. Keep in mind that these are relationships in a vendor's own data, not proof of cause.
Sources
Ahrefs, How Stable Are AI Citations? 10 Things We Learned Tracking 37,500 Citations
Are you tracking citations as a single day or as a trend?
A short conversation is enough to review your pages, titles and measurement approach together.