BlogFindings
AI citation tracking: 50 to 83 percent of 37,500 citations were gone by day two, and about three quarters came back
Ahrefs published a citation stability study on 29 October 2026 that followed about 37,500 citations across five AI tools every day for 20 days. The share of day one citations missing by day two ran from 50 percent in Google AI Overviews to 83 percent in Copilot. Roughly 75 percent of the citations that dropped at least once came back inside the window and about half returned within three days, so the headline finding is churn rather than loss. Lantad measured none of this: the scanner reads sites rather than answers. What it can say is that the site side of the same question does not move on that timescale.
Lantad measured none of this. The scanner fetches pages the way an AI crawler does and grades what is readable, which is the site side of the problem, and it asks no engine the same question for 20 consecutive days. So this post reports Ahrefs' study with its sample and its own caveats attached, then turns to the part that belongs here, which is what on your own site is stable enough to be worth measuring at all. Ahrefs Brand Radar and this scanner are different instruments pointed at different halves of the same question, and the comparison between them is written up separately. The gap between a volatile answer side and a stable site side is the practical content of generative engine optimization work, and it is the reason the two halves need different reporting.
In short
- AI citation tracking measures a quantity that moves daily: Ahrefs followed about 37,500 citations across five AI tools every day for 20 days and published on 29 October 2026 that the share of day one citations missing by day two ran from 50 percent in Google AI Overviews through 58 percent in AI Mode and 59 percent in ChatGPT to 81 percent in Gemini and 83 percent in Copilot.
- Most of that movement reversed in the same study: Ahrefs reported that about 75 percent of citations which dropped at least once returned inside the 20 day window, about half returned within three days, and only 20 to 27 percent of day one citations disappeared and never came back.
- Ahrefs found the two directions are not symmetric: a citation that had already lasted eight to fourteen days almost always returned, with never returned shares of 4 percent in AI Overviews, 2 percent in AI Mode and 0.7 percent in ChatGPT, against 34 to 41 percent for citations that lasted a single day.
- Google's AI features documentation, carrying Last updated 2025-12-10 UTC when Lantad read it on 29 October 2026, states that AI Overviews and AI Mode may use different models and techniques so the set of responses and links they show will vary, and that a change to preview controls has to be recrawled and processed, which it says takes several days to several months.
- Lantad measured 418 of the 1,419 hostnames in this repository's two committed corpus seed files giving a plainly identified crawler nothing to read on 16 October 2026, and 980 of the same 1,419 clearing every gate OpenAI's documentation implies on 26 October 2026, which are site side facts that do not change between two daily checks the way a citation does.
What did the citation stability study actually track?
The study is titled How Stable Are AI Citations? 10 Things We Learned Tracking 37,500 Citations, published by Ahrefs on 29 October 2026, written by Mateusz Makosiewicz with Xibeijia Guan credited as contributor and Ryan Law as reviewer. It is at ahrefs.com/blog/how-stable-are-ai-citations/, written here as plain text rather than as a link because this site does not link competitors. The method is simple enough to state in one sentence: the same questions were put to five AI tools every day for 20 days, every website cited on day one was recorded, and each of those was then checked daily to see whether it was still cited and, if it had dropped out, whether it came back.
The five tools are Google AI Mode, ChatGPT, Google AI Overviews, Gemini and Copilot. The published table carries three columns, AI tool, Questions tracked and Websites cited on day one, and the counts are 1,996 questions and 18,009 day one citations for AI Mode, 3,173 and 11,225 for ChatGPT, 563 and 4,830 for AI Overviews, 1,276 and 1,994 for Gemini, and 823 and 1,447 for Copilot. Those five day one counts sum to 37,505, which is arithmetic on the published rows rather than a figure the study states; the study says it followed about 37,500 citations. The question counts sum to 7,831 on the same basis, and no total is published for them either.
Two things about that sample matter before any percentage from it is quoted. The unit is a website cited in an answer, not a brand and not a page, so a site that keeps its citation while a different URL of it is swapped in counts as retained. And the questions are not a random draw from what people ask: they are a tracked panel, which is the same constraint that applies to every published appearance rate in this category and the reason a prevalence figure needs its sample attached or it means nothing. Ahrefs states the panel size per tool, which is the part that makes the study auditable, and it publishes its own limitations section. Both are the standard this blog holds itself to in the method behind every score, and it is worth saying plainly that the study meets it.
| AI tool | Questions tracked | Websites cited on day one | Gone by day two |
|---|---|---|---|
| Google AI Mode | 1,996 | 18,009 | 58% |
| ChatGPT | 3,173 | 11,225 | 59% |
| Google AI Overviews | 563 | 4,830 | 50% |
| Gemini | 1,276 | 1,994 | 81% |
| Copilot | 823 | 1,447 | 83% |
What does AI citation tracking actually measure day to day?
On the day one to day two step, the study reports 50 percent of AI Overviews citations gone, 58 percent for AI Mode, 59 percent for ChatGPT, 81 percent for Gemini and 83 percent for Copilot. Ahrefs qualifies the top of that range itself: excluding answers that carried no citations at all, it puts the comparable figure at 50 to 60 percent, because Gemini and Copilot frequently returned answers with no citations and an answer with nothing in it is not the same event as a citation being dropped. Only 4 percent of Gemini questions and 27 percent of Copilot questions carried citations on all 20 days, which is the detail that stops the 81 and 83 percent being read as those tools being eight times less loyal.
The second half of the finding is the half that gets lost when the number is repeated. About 75 percent of the citations that dropped at least once returned within the 20 day window, about half of them returned within three days, and only 20 to 27 percent of day one citations disappeared and never came back. So the honest summary of what a daily citation check measures is a population churning around a fairly stable core, not a population draining away. A tool that samples once a week and reports the delta as a gain or a loss is reporting, for most of its rows, the phase of an oscillation rather than a change in standing.
This is a specific instance of a general problem this blog has measured from several directions. AI rank tracking carried no rank for 15 of 31 brand mentions because the shape of the answer decided whether a position existed to report. Citation count is not answer influence because a cited source may contribute nothing to the sentence a reader acts on. And three of the eight engines in the registry this product ships cannot search the web at all, read out of the repository on 29 August 2026, so a panel's engine mix changes what a stability figure even means. None of that makes answer engine optimization unmeasurable. It makes the measurement interval, the engine list and the unit part of the result rather than a footnote to it.
Why a one day citation and an eight day citation are not the same signal
The most useful number in the study is not the headline. It is the asymmetry between short lived and established citations, which is what turns a volatility figure into something a person can act on. Of citations that had lasted a single day, the share that never returned was 34 percent in AI Overviews, 41 percent in AI Mode and 35 percent in ChatGPT, and next day return for that group ran only 14 to 25 percent. Of citations that had already lasted eight to fourteen days, the never returned share was 4 percent in AI Overviews, 2 percent in AI Mode and 0.7 percent in ChatGPT, so between 96 and 99 percent came back, and next day return for that group ran 69 to 87 percent.
Read together those two groups say that persistence predicts persistence. A citation that has held for over a week behaves like a stable fact about how an engine answers a question, and its absence on any given day is close to noise. A citation that appeared once behaves like a coin toss, and treating its arrival as a win or its departure as a regression is reading a single sample as a trend. That distinction costs nothing to apply and it changes what a dashboard should be allowed to alert on.
What the study explicitly does not do is explain why the durable ones are durable. Ahrefs joined SEO metrics, Domain Rating, URL Rating, backlinks and organic traffic, to the cited pages and found connections, and then says in its own words that these correlations show what long lasting citations have in common, not what makes them last. It describes the ranking and traffic relationship as a strong connection rather than a definitive rule, and notes that organic traffic and rankings are not independent of each other, so the two findings may be one finding seen twice. That is the right way to report it, and it is consistent with what this blog found when it looked at the entry side: four factors decided the first citation and formatting was not one of them. It also sits beside the observation that ChatGPT citations are rare and tend to land on the homepage, which is a reminder that the cited URL and the useful page are often not the same thing.
| Tool | Lasted one day | Lasted 8 to 14 days | Next day return, 8 day group |
|---|---|---|---|
| Google AI Overviews | 34% | 4% | 69 to 87% |
| Google AI Mode | 41% | 2% | 69 to 87% |
| ChatGPT | 35% | 0.7% | 69 to 87% |
What do Google and OpenAI say about how fast their own systems change?
A volatility finding is easier to size when you read what the operators publish about their own behaviour, because two of them document timescales that bear directly on it. Google's AI features documentation carried Last updated 2025-12-10 UTC when Lantad read it on 29 October 2026. It states that AI Overviews and AI Mode may use different models and techniques, so the set of responses and links they show will vary. Variation is therefore documented behaviour rather than a defect, and a stability study measuring it is measuring something Google says is there. The same page states that AI Overviews are only shown when Google's systems determine it is additive to classic Search, and that they often do not trigger.
The same documentation gives the site side timescale, and it is far slower than a day. A page has to be indexed and eligible to be shown in Google Search with a snippet before it can appear as a supporting link, and the page names nosnippet, data-nosnippet, max-snippet and noindex as the controls over how much of a page may be shown. On how long a change to those controls takes to bite, it says Google must recrawl and process the change, which it puts at several days to several months. So the lever a site owner holds operates on a scale of days at best, while the citation it is meant to influence can turn over daily.
OpenAI's crawler documentation is blunter about the lag and carries no publication or update date at all, which Lantad confirmed by reading it on 29 October 2026. It names four crawlers, OAI-SearchBot for search, GPTBot for training, ChatGPT-User for user initiated actions and OAI-AdsBot for checking pages submitted as ads. It states that sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers, that a webmaster can allow OAI-SearchBot in order to appear in search results while disallowing GPTBot, and that for search results it can take about 24 hours from a robots.txt update for OpenAI's systems to adjust. That 24 hours is the same unit as the day to day churn, which has a consequence worth stating: on a single daily check you cannot tell a dropped citation from your own file change still propagating. This blog has made the narrower version of that point before, that a robots.txt edit does not take effect when you save it, and the token doing the work here is the one that 16 of 140 robots.txt files named.
-
Google AI features docVariation documented States the set of responses and links shown will vary, and that AI Overviews often do not trigger. Last updated 2025-12-10 UTC. -
Google preview controlsDays to months A change to nosnippet, max-snippet or noindex must be recrawled and processed, which the page puts at several days to several months. -
OpenAI robots.txt lagAbout 24 hours For search results the page states it can take about 24 hours from a robots.txt update for OpenAI's systems to adjust. -
OpenAI doc dateNone published The crawler documentation carries no publication or update date, so a reader cannot tell when a statement on it last changed.
What can you measure on your own site that does not move daily?
The reason this blog reports a citation stability study rather than running one is that the two sides of the question have different error bars, and only one of them is under the site owner's control. A robots.txt rule either permits a named crawler at a path or it does not. A server either answers a declared crawler with HTTP 200 or it does not. A page either carries extractable words or it does not. Those are not sampled from a churning population, and a second reading on the next day returns the same answer unless somebody changed the site.
The figures are not reassuring, which is the point of reporting them next to a volatility study. Lantad measured all 1,419 hostnames in this repository's two committed corpus seed files on 26 October 2026 and found that 980 cleared every gate OpenAI's own documentation implies, while 25 were closed by a robots.txt rule and 217 permitted the crawler in the file and then answered the home page with HTTP 403. On 16 October 2026 the same frame gave a harder number: 418 of the 1,419 gave a plainly identified crawler nothing to read, 347 because they delivered no HTML at all and 71 of the 1,072 that did delivering zero words.
Put those beside the churn figures and the order of work is hard to argue with. For a site in that 418, no citation tracking tool can report anything but absence, and the absence is not volatile. It is a fixed property of the site that will read the same on every one of the 20 days, and it is fixable in an afternoon by whoever controls the template. Prose parity, the gap between what a visitor reads and what a crawler receives, is the usual cause and it is the one measurement here that is genuinely stable enough to set a target against. None of that predicts a citation. It is the condition under which a citation is possible at all, which is a different and more tractable claim.
- robots.txt permits the search crawler at the root 25 of 1,419 were closed by a rule on 26 October 2026, the smallest of the four gates measured.
- The server answers a declared crawler with HTTP 200 217 of 1,419 permitted the crawler in the file and then answered the home page with HTTP 403 on 26 October 2026.
- The HTML carries extractable words 418 of 1,419 gave a plainly identified crawler nothing to read on 16 October 2026: 347 no HTML, 71 of 1,072 zero words.
- Every gate cleared together 980 of 1,419 cleared all four conditions OpenAI's documentation implies on 26 October 2026.
What this does not tell you
Ahrefs publishes its own limitations and they are worth repeating rather than summarising away. The return rates may be understated, because a citation lost near the end of the 20 days had less time to come back than one lost on day two. Only the first disappearance of each citation was tracked, so repeat disappearances are not counted and the study cannot say how often a citation oscillates. The SEO metric and title match findings are correlations, which the authors say show what long lasting citations have in common rather than what makes them last. Brand name variants were grouped with AI assistance, and a manual check of 55 brands found that 30 to 36 percent of apparent brand losses were the same brand under a different name, so the brand level figures are described as slightly conservative. The findings about one named brand rest on 22 questions and the study says they may not apply equally to every company.
The limits on the Lantad figures beside them are different in kind and no smaller. The 1,419 hostnames are two sampling frames stratified by platform and sector for the crawlability study, not a random draw from the web, and the smaller strata hold about thirty sites each. The unit is usually one apex home page fetched once from one network location, so a site clearing every gate there may still refuse an article path, and one request on one date cannot separate a standing policy from a transient error. Every request was sent as LantadBot rather than as any vendor's AI crawler, and since the operators publish address files that edge rules commonly verify against, a site that refused this scanner may well admit OpenAI or Google. None of those runs tested that.
The honest join between the two halves is narrow, and it should be stated rather than implied. No figure in this post says that fixing a site side defect produces a citation, or a durable one. Nothing here was measured by asking an engine anything. What the two sets of numbers support together is a sequencing argument: the answer side turns over on a timescale of about a day, the controls over it take days to months to register, and the site side conditions are stable, cheap to read and currently failing on a large minority of a stratified corpus. Which of those to work on first is not a close call.
Reported, not measured here
- About 37,500 citations, five tools, 20 consecutive days, published 29 October 2026
- 50 to 83 percent of day one citations gone by day two, 50 to 60 percent excluding empty answers
- About 75 percent of dropped citations returned inside the window
- Cannot say why durable citations are durable: the authors call it correlation
- Cannot say how often a citation oscillates: only first disappearance tracked
Measured by Lantad, on the dates given
- 1,419 hostnames from two committed corpus seed files, stratified not random
- 980 cleared every gate, 25 closed by rule, 217 permitted then 403, on 26 October 2026
- 418 gave a declared crawler nothing to read, on 16 October 2026
- One apex home page, one request, one network location, sent as LantadBot
- No engine asked anything, so no figure here reports a citation
Lantad
Published .
A buyer evaluating AI visibility measurement usually arrives with one question about the numbers on the dashboard: are they stable enough to act on. Ahrefs published a study on 29 October 2026 that puts a figure on it, and the figure is large. Tracking about 37,500 citations across five AI tools every day for 20 days, it found that between 50 and 83 percent of the websites cited on day one were not cited the next day, depending on which tool was asked. That is the clearest published answer so far to what AI citation tracking is actually measuring from one day to the next, and anyone selling or buying a number in this category has to deal with it.
Common questions
What does AI citation tracking measure if citations change this often?
It measures the state of a churning population on the day it is sampled. Ahrefs' study of 29 October 2026 found 50 to 83 percent of day one citations missing by day two across five tools, but about 75 percent of the citations that dropped came back inside 20 days and about half returned within three. So a single check reports a position in an oscillation, and only a run of checks separates a trend from noise.
How many daily checks does it take before a citation means something?
Ahrefs' figures suggest about eight. Of citations that had already lasted eight to fourteen days, the share that never returned was 4 percent in AI Overviews, 2 percent in AI Mode and 0.7 percent in ChatGPT. For citations that had lasted a single day the same share was 34, 41 and 35 percent. A citation seen once is close to a coin toss; one that has held for over a week behaves like a stable property of how the engine answers.
Did Lantad measure citation stability?
No. Lantad reads sites rather than answers: it fetches pages as an AI crawler would and grades what is readable, and it does not put the same question to an engine for 20 days. Every stability figure in this post belongs to Ahrefs and is attributed to it. The Lantad figures quoted are earlier corpus measurements of site side conditions, each carrying the date it was measured.
How long does a change on my own site take to show up in an AI answer?
Longer than the churn. OpenAI's crawler documentation, which carries no date and was read on 29 October 2026, states it can take about 24 hours from a robots.txt update for its systems to adjust for search results. Google's AI features documentation, last updated 2025-12-10 UTC, says a change to preview controls must be recrawled and processed, which it puts at several days to several months.
See what AI can read on your site
Run a free scan and get a graded report of exactly what AI crawlers can and cannot read, with ranked fixes.
AI visibility
AI visibility is whether AI systems can access, read, and faithfully represent your web pages, the prerequisite for being cited in an AI answer.
Generative Engine Optimization (GEO)
Generative Engine Optimization (GEO) is the practice of structuring and publishing content so that generative AI engines can fetch it, read it in full, and cite it when they answer a question.
Answer Engine Optimization (AEO)
Answer Engine Optimization (AEO) is the practice of making a page readable, allowed, and structured enough that AI answer engines can use it as a source when they compose an answer.
AI rank tracking: 15 of 31 brand mentions carried no rank, and the answer shape decided
Lantad read position and answer shape out of the 55 answers committed at core/fixtures/tracker-golden/goldens.json on 24 October 2026, using the stored labels and the recorded judge replies. 31 of the 55 name the brand. 16 of those 31 carry a position and 15 carry none. All 12 of the list answers that named the brand carried a rank, against 4 of the 17 prose answers, so the shape of the answer rather than the brand's standing decided whether there was a rank to report at all.
More AI citations did not mean more of your page in the answer
A measurement framework posted to arXiv on 29 April 2026 scores how much of each cited page reaches the answer text. Across 21,143 valid search-layer citations, Perplexity averaged 16.35 citations per prompt to ChatGPT's 6.88, and mean influence among fetched pages ran the other way, 0.0646 against 0.2713.
AI visibility tracking: 3 of the 8 engines we ask cannot search the web
Read out of this repository on 29 August 2026, the answer engine registry Lantad ships carries 8 engines. Five are recorded as performing live retrieval and returning the sources they grounded in. Three are recorded as having no web access and no sources at all, and one of those three is the only engine a free run asks.
