Ahrefs followed about 37,500 citations across five AI tools every day for twenty days. Between half and four fifths disappeared by the next day, and about three quarters of those came back. We checked our own server logs over the same twenty days to see whether the crawlers were re-reading anything.
Table of contents
- What this is about
- What the source says
- What the data shows
- What it does not mean
- What to do
- The numbers
- FAQ
Key takeaways
- Across five AI tools over twenty days, 50 to 83% of citations were gone by the next day. About 75% of those returned, roughly half of them within three days.
- How long a citation had already lasted predicted whether it came back. Citations that had held for eight days or more almost always returned; citations that lasted a single day never returned in 34 to 41% of cases.
- The churn is not driven by re-reading. Ahrefs found that when ChatGPT ran exactly the same searches as the day before, the sites it cited still changed 81% of the time.
- Our own logs point the same way. Over the same twenty days, GPTBot came every day and fetched our two sitemaps every time, but fetched articles on one day out of twenty. Bingbot was the only crawler that returned to articles regularly.
- The study names its five engines and its repetition schedule. It does not say whether answers were collected through each product’s interface or an API, and it names no region. We report its figures as Ahrefs’ measurements, not as rates.
What this is about
An AI tool cited your page today and not tomorrow. Ahrefs measured how common that is across five tools, and how often the citation returns. The answer changes what a citation is worth as a metric, and it changes what you should conclude when one disappears.
What the source says
The study was published on 29 October 2026 by Mateusz Makosiewicz, with the measurements run by Xibeijia Guan and review by Ryan Law. Five tools were asked the same questions every day for twenty days: Google AI Mode, ChatGPT, Google AI Overviews, Gemini and Copilot. Every website cited on day one was followed through the window.
The question counts differ sharply by tool: 3,173 questions on ChatGPT, 1,996 on AI Mode, 1,276 on Gemini, 823 on Copilot, 563 on AI Overviews. Day-one citations ranged from 18,009 sites on AI Mode down to 1,447 on Copilot.
The headline finding is disappearance. At least half of day-one citations were gone by the next day on every tool: around 60% on AI Mode and ChatGPT, around 50% on AI Overviews, and around 80% on Gemini and Copilot. Gemini and Copilot look worse partly because their next answer often cited nobody at all; excluding those answers, all five tools land between 50 and 60%.
The second finding is recovery. About 75% of citations that dropped out returned at least once inside the twenty days. Roughly half of those returned within three days. Only 20 to 27% of day-one citations disappeared and never came back.
The third finding is the one with a mechanism attached. Before answering, these tools run several web searches, a step the industry calls fan-out. ChatGPT repeated exactly the same searches as the previous day only 26% of the time. But even when the searches were identical, the sites it cited changed 81% of the time, against 91% when the searches differed. The selection is happening after retrieval, not during it — which is the same seam we looked at when ChatGPT searched the web for one answer in seven.
Ahrefs also reports that citation age predicts survival. Citations that had lasted eight to fourteen days almost always returned after a gap; citations that lasted a single day never returned in 34 to 41% of cases.
What the data shows
First the method check we apply to any vendor research. Of the seven points we require — engine, interface or API, collection dates, region, prompt set, number of repeats, and stochastic variability — this study clears three outright.
It names all five engines, though not their model versions. It states the repetition schedule plainly: the same questions, every day, for twenty days. And variability is not a footnote here, it is the subject.
Two points are partial. The question counts are given per tool, but the questions themselves are not published, so nobody can re-run this. And the twenty-day window has no dates attached; the article appeared on 29 October 2026, and a separate cross-check is dated 28 August to 4 October 2026, which is not the same window.
Two are missing outright. The study does not say whether answers were collected through each product’s interface or through an API — the two can return different things for the same prompt. And it names no region, which matters because AI Overviews and AI Mode vary by market.
So these are Ahrefs’ measurements of Ahrefs’ question set. The direction is convincing and the sample is large. The percentages are not rates you can apply to your own site.
Our own check. We cannot measure citation stability — we have no tool that watches answers. We can measure the layer underneath it: whether the crawlers come back to the page at all. We parsed our origin logs for the twenty days from 13 October to 1 November 2026, excluding our own two IP addresses, and counted which of our URLs each crawler fetched on which day.
The pattern is lopsided. GPTBot appeared on all twenty days, and on every single one of them it fetched /sitemap_index.xml and /news-sitemap.xml. It fetched articles on one day out of the twenty, taking 34 of them at once. ClaudeBot behaved the same way at lower frequency: sitemaps on all nine days it appeared, articles on two. OAI-SearchBot appeared on five days and touched eight articles. PerplexityBot came on seven days and fetched two articles, returning to one of them six times.
Bingbot was the exception, and it is not an AI crawler: it fetched articles on eighteen of twenty days, 34 unique ones.
What this measurement cannot show. A CDN sits in front of this site and holds pages at the edge for four hours, and a scheduled job refreshes that cache every three hours. Any crawler request served from the edge never reaches our server and never enters these logs. So our article counts are a floor, not a count — the true number of article fetches is higher by an unknown amount. The sitemap figures are firmer, because we do not refresh sitemaps at the edge, so those requests do reach us.
What survives that caveat is the shape: the daily visit from an AI crawler is spent on the sitemap, checking what is new. Re-reading the article is not what it comes for. Put next to Ahrefs’ finding that identical searches still produced different citations 81% of the time, both measurements point away from the page and toward the moment of selection.
What it does not mean
A disappeared citation is not a penalty, and usually not a signal at all. With half or more of citations gone by the next day across every tool measured, a single absence carries almost no information. Ahrefs’ own framing is that the volatility looks like a built-in trait of these systems.
It does not mean your edit worked. This is the trap the study is most useful for. If roughly 75% of lost citations return on their own, and about half of those within three days, then a page you edited on Monday reappearing on Wednesday tells you nothing. The base rate of spontaneous return is high enough to explain almost any single comeback.
It does not mean citations are worthless as a measurement. It means a citation is a poor unit. The study’s own suggestion is to count how many of the last seven days you were cited for a given question rather than whether you are cited today. That is a different metric built from the same observations, and it survives the churn — though it still says nothing about whether anyone arrives, which is where citation counts and traffic parted company in the 2026 data.
And our crawler numbers do not prove crawlers ignore articles. They prove our server did not see those requests. The CDN in front of it is the likeliest reason, and we have said what we cannot separate.
What to do
Stop reading single checks. If you are watching a prompt, record whether you were cited on each of the last seven days and use that count. One screenshot on one day is not evidence of anything in a system where half the citations turn over nightly.
Judge a drop by what preceded it. A citation that had held for a week or more and then vanished is very likely to return by itself; the study puts recovery for those in the high nineties. A citation that appeared once and vanished may well be gone for good — that is the one worth examining. Age of the citation is the cheapest signal you have.
Separate brand questions from discovery questions. Ahrefs reports that for questions naming a brand, that brand’s own site was cited on 60 to 68% of days on Google’s tools and ChatGPT, against 16 to 20% for other sources in the same answers. Tracking “best X” and “is X any good” in the same bucket averages two behaviours that are nothing alike.
If a drop persists for two or three weeks, then look at the page. Not before. The study found that pages which rank in Google, receive search traffic and answer the question in the title held their citations longer — ordinary search performance, not a separate discipline. That is the mechanism the section pillar sets out.
The numbers
| Measure | Reported figure |
|---|---|
| Citations followed, five tools, twenty days | ~37,500 |
| Gone by day two, AI Mode and ChatGPT | ~60% |
| Gone by day two, AI Overviews | ~50% |
| Gone by day two, Gemini and Copilot | ~80% |
| Dropped citations that returned within the study | ~75% |
| Of those, returned within three days | ~half |
| Disappeared and never returned | 20–27% |
| One-day citations that never returned | 34–41% |
| ChatGPT repeating identical fan-out searches | 26% of days |
| Cited sites changing despite identical searches | 81% |
| Brand’s own site cited, questions naming the brand | 60–68% of days |
| Other sources in those same answers | 16–20% of days |
Ahrefs, published 29 October 2026; measurements by Xibeijia Guan across Google AI Mode, ChatGPT, Google AI Overviews, Gemini and Copilot, asked daily for twenty days. The study does not state whether answers came from each product’s interface or an API, names no region, and does not publish its questions.
| Our logs, 13 October to 1 November 2026 | Days seen | Days it fetched an article | Unique articles |
|---|---|---|---|
| GPTBot | 20 | 1 | 34 |
| Bingbot | 20 | 18 | 34 |
| ClaudeBot | 9 | 2 | 43 |
| PerplexityBot | 7 | 7 | 2 |
| OAI-SearchBot | 5 | 2 | 8 |
Origin access logs for seoafterai.com, successful responses only, with the site owner’s two IP addresses excluded. A CDN holds pages at the edge for four hours and a scheduled job refreshes it every three, so requests served from the edge are absent here: article counts are a lower bound. GPTBot and ClaudeBot fetched both sitemaps on every day they appeared.
FAQ
My page lost its AI citation. Should I change it?
Not on that evidence. In this study between half and four fifths of citations were gone by the next day, and about 75% of the ones that dropped out came back inside twenty days, roughly half of those within three days. Changing the page on day two means you will probably see it return and credit the edit. Wait two or three weeks of sustained absence before treating the page as the problem, and then look at ordinary search performance first.
If the citations change every day, is the crawler re-reading my page every day?
Our logs say no, with a caveat. Over twenty days GPTBot fetched our sitemaps every single day and our articles on one of them; ClaudeBot did the same on a slower schedule. A CDN in front of the site hides some repeat requests from us, so those article counts are a floor. But Ahrefs’ finding points the same way from the other side: when ChatGPT ran identical searches on consecutive days, the sites it cited still changed 81% of the time. The page is not what changed.
Can I use the 75% recovery rate as a planning number?
No. It comes from one company’s question set, over one undated twenty-day window, with no region stated and no indication of whether answers were collected through each product’s interface or an API. Ahrefs’ own cross-check against a second dataset returned a lower figure, closer to two in three. Treat it as evidence that spontaneous recovery is common, which is the part that should change your behaviour, and not as a number to forecast with.
