Blog · Data and trends

Citation churn, and how fast the cited set changes

Published 8 July 2026 · Updated 12 October 2026 · 9 min read

Citation churn, and how fast the cited set changes
Photograph: Micah Bratt / unsplash

The short answer: Citation churn is the rate at which the sources attached to an answer are replaced by different sources. In our own re-runs of client prompt sets, and in the vendor panels that publish on this, the cited URLs turn over faster than the brand names in the same answers. We have not seen an independent public dataset that measures that rate across engines, so treat the direction as established and the rate as unmeasured. A citation that vanished overnight is more often turnover than a ranking event.

Most teams meet this the hard way. They screenshot a Monday answer that cites them, come back on Thursday, find the link gone, and open an investigation into a problem that does not exist.

Why the source layer moves and the brand layer does not

The two layers are plausibly produced by different parts of the machine. On the usual account of retrieval-augmented generation, a brand name can reach the answer from the model's own weights as well as from the retrieved passages, while a citation can only come from one specific retrieval pass, at one moment, against one ranked list. We have not found published internals for any of the engines we test, so this is the mechanism we infer, not one we can verify.

Google describes the retrieval side of this openly, for its own surfaces. Its documentation on AI features says both AI Overviews and AI Mode may use a "query fan-out" technique, issuing multiple related searches across subtopics and data sources, and that the two "may use different models and techniques, so the set of responses and links they show will vary." That is Google's description of Google. We have no equivalent statement from the other engines, so read the rest of this section as a model that fits Google's documented behaviour and that we assume, without proof, is broadly similar elsewhere. If each run composes its own sub-queries, each sub-query ends somewhere down a ranked list, and that cut line is where the churn lives.

Now think about what sits near that cut line. Picture two documents a rank apart, one inside the cut and one outside it, separated by very little. Google does not publish where the cut falls or how it is scored, so this is an illustration rather than a measurement. On that picture, a small perturbation upstream, a different sub-query phrasing, a freshly crawled competitor, a cache that expired, flips which of the two gets the link. Nothing about your page changed. The boundary moved past it.

The retrieval path matters as well, and here OpenAI is the engine that documents it. OpenAI publishes separate agents for separate jobs: OAI-SearchBot for the search surface, while ChatGPT-User fires on a user action and, in OpenAI's words, "is not used for crawling the web in an automatic fashion". An answer assembled from indexed content and an answer assembled from a live fetch can disagree, because they may be reading different snapshots of your site. We have not tested whether the two paths diverge in practice on a given site. That is a structural reason to expect churn, and it is related to why an AI answer changes between two identical runs.

What the public numbers support

Independent measurement of the link layer is thinner than people assume, so be careful which figures you repeat. The most useful public one comes from Pew Research Center, which analysed 68,879 Google searches made by 900 US adults during March 2025, with the result pages re-collected between 7 and 17 April 2025. An AI summary appeared on 18% of those searches. Of those summaries, 88% cited three or more sources and only 1% cited a single source, with a median summary length of 67 words.

Read that carefully before you build on it. Pew collected up to three source URLs per summary, so 88% is a floor and the study does not establish an upper bound on how many sources an AI Overview carries. It is also Google only, English language, US panel, and eighteen months old at the time of writing. What survives all of that is modest and still useful: a median answer is short, and it carries at least three sources. A short answer drawing from a wide candidate pool leaves room for the cited set to differ between runs.

Pew also found the composition differs from ordinary results: 6% of the sources linked in AI Overviews were .gov sites against 2% for standard results, in that same April 2025 collection. Composition shifts like that are slower and more informative than any single link appearing or vanishing. More figures of this kind, with their sources, sit in AI search statistics worth quoting.

Commercial monitoring panels report that cited URLs turn over faster than named brands. We are not repeating their rates here, because the ones we could find do not publish the prompt set, the engines, the window or the method behind them, and a figure we cannot open is a figure we will not quote. Take the direction from them if you like. Take the size of it from your own measurement, below.

The usual mistake is tracking "did my URL appear". Track the set instead.

Run the same prompt once a day for thirty days on each engine you care about. Record every cited URL and its domain. Then compute three things.

Day over day overlap. Take the cited set on day N and day N plus one, and divide the size of the intersection by the size of the union. That is a Jaccard score. It gives you a churn baseline per engine. In the prompt sets we have run this way, the scores spread widely between engines, from near total daily replacement at one end to most of yesterday's sources holding at the other. That is a handful of prompt sets on client work, not a survey, so use it as a reason to measure rather than as a figure to cite. You cannot interpret your own disappearance until you know which kind of engine you are standing in.

The persistent core. Count how many days in thirty each URL appears. Sort descending. There is usually a short head of sources present on most days and a very long tail present once. The head is the answer's actual spine. If you are in the tail, you were sampled, not selected.

Domain versus URL churn. Recompute both numbers at domain level. In the prompt sets we have measured, domain level churn came out lower than URL level churn, which is what you would expect if engines swap between pages on the same site, though we cannot see the swap itself. A brand that reads one deep article in and out of the citations each week may have a domain that stayed put throughout.

That third check is the one that resolves most arguments in our experience. Losing a URL while holding the domain is the common case and rarely worth acting on. Losing the domain for a week on a prompt where you previously held the core is more likely to be a signal, and it is worth tracing back to a crawl error, a redirect, or a change in who the engine trusts on that subtopic.

The failure modes this creates

Reacting to noise. A single run is an anecdote. If your reporting cadence is a weekly screenshot, every report will show movement, and none of it will mean anything. Report the core set and its overlap, with the run count attached.

Over-fitting to a lucky day. A page that was cited once gets rewritten to look like the day it was cited. You are optimising against one sample from a distribution you never measured.

Mistaking an engine shift for a site problem. Engines change their own composition. Google's AI Mode is a different surface with different fan-out behaviour from AI Overviews, and what that means for a publisher is covered in Google AI Mode: what changes for a publisher. When your whole peer group drops out of the citations on the same date, the cause is upstream.

Measuring the wrong prompts entirely. Churn on a prompt no buyer types is a statistic about nothing. Build the prompt list from how people actually research, which is the subject of how buyers are actually using AI search.

What to do about it

Reduce your dependence on any single page holding a slot. Three practical moves.

Give each subtopic its own clean, answerable page rather than one long article that has to win every sub-query. Fan-out asks narrow questions, and narrow pages are easier to retrieve for them.

Make the same claim true in more than one place the engine already trusts. Corroboration survives churn better than a single well written page, because it is not competing for one slot.

Keep the facts identical across your own pages. When two of your pages disagree, each run can pick a different one, and you have added churn of your own making.

If you want a worked version of this on a real prompt set, the sample audit report shows the shape, including how evidence is attached to a finding rather than asserted.

Honest limits

This post establishes that citation churn exists and gives you a way to measure it. It does not establish a rate, and the figures we would have needed for one do not exist in a form we can open: there is no independent public dataset on URL level citation persistence across engines, and the commercial panels that publish rates do not publish their prompt sets or methods. The mechanism in the second section is documented by Google for Google's own surfaces and inferred for every other engine, so a careful reader should push back on any sentence here that generalises fan-out beyond AI Overviews and AI Mode. Our own observations come from re-running client prompt sets, which is a sample of commercial work rather than a census, with no fixed cadence and no published prompt list, so they are evidence that we saw something and not evidence of how often it happens. We also cannot separate retrieval randomness from genuine re-ranking, because none of the engines we test exposes the sub-queries behind a given answer, and we have found no documentation saying any of them does.

FAQ

How fast does the cited set actually change? We cannot give you a rate, and nobody we can cite has published one from an open method. What we see on the prompt sets we re-run is heavy day to day replacement on some engines and a short head of sources that persists on others. Measure your own baseline per engine before you judge a drop.

Is losing a citation the same as losing visibility? No. Brand mentions and citations move at different speeds. Check whether your brand is still named in the answer before you investigate a missing link.

How many runs do I need before the number means anything? More than one, and more than one prompt. We have not published a curve for this, and the reasoning is ordinary sampling rather than a measurement of ours: a few dozen prompts run repeatedly tells you more about a distribution than a hundred prompts run once.

Should I report citations or mentions to leadership? Report whether you are named as the headline, and keep citations underneath it as the diagnostic that explains why the headline moved.

Does more publishing reduce churn? It can widen the number of pages eligible for a slot, which smooths your exposure. It does not make any individual slot stable, and volume without accuracy adds contradictions.

Sources

All three were opened and checked against the sentences that cite them on 12 October 2026.

No source is cited here for a churn rate, because we did not find one with an open method. See the limits section above.

Related reading

AI search statistics worth quotingHow buyers are actually using AI searchGoogle AI Mode: what changes for a publisherWhy an AI answer changes between two identical runsDo AI referrals convert betterHow much traffic AI search actually sends