How often do AI citations change? 17 days of data

“Surely this is just SEO” is a reasonable objection. If search engines already find, rank and revisit your pages, why treat AI citations as a separate problem? A useful page still needs access, clear language and evidence that answers a buyer’s question.

But the available data points to a distinction worth tracking. Search rankings can stay put while the underlying cited page changes. The domain may remain familiar to the system while the exact page supporting the answer rotates underneath it.

How often do the domains that cite you actually change?

Of the 40 domains cited at least three times in the first scan, 27 were still cited 25 days later, according to how CiteGraph measures. The churn series records that as 68%.

Exact pages were much less stable. Of the 179 pages cited in that first scan, 14 appeared again after 25 days. The churn series records that as 8%.

Here is the arithmetic and the decision it supports. Start with the 40 heavily cited domains. Subtract the 27 that returned, and 13 did not return in that comparison. That’s roughly a third of the heavily cited set gone, smaller than the page-level churn, but not trivial.

Now start with the 179 exact pages. Only 14 returned, so most of that page set did not return in the comparison. If one of your pages disappeared, check whether your domain is among the 27 that held. If it is, investigate a page-level rewrite or replacement before treating the result as a domain-level SEO problem.

The domain result looks durable; the page result looks highly changeable.

The broader count is less concentrated. Across every cited domain in the comparison, 56 of 102 came back after 25 days. The project notes that the long tail is mostly sampling, not change, so the heavily cited set gives the clearer signal.

What should you change in your content plan?

Treat the domain as an asset, but treat the page as a replaceable component. With only 8% of exact pages surviving 25 days but 68% of domains holding, betting on one URL is betting against the data.

One page may explain the category. Another may compare options. Another may answer implementation or fit questions. If the system changes its source page, you still have relevant material available, provided those pages state distinct facts rather than repeating a brand claim.

The data does not support a plan built around a single evergreen winner. It supports a maintained group of answerable pages on a domain that remains easy to recognise and access.

A fall in page citations is not automatically a domain failure. Check whether another page on the same domain has taken the slot, whether the buyer question stayed the same, and whether the old page still contains the clearest answer.

How should you measure AI citation volatility?

Start with repeated questions, not a single answer. CiteGraph’s live research data covers 77 scans of 44 sites in 48 categories, with 507 distinct buyer questions. Those scans produced 64,667 citations across 9,791 AI answers.

The weekly research records the wider sample, while the measurement method guide explains how to connect answer-level results to cited pages.

Record the cited domain, exact cited page, buyer question and scan date for each observation. Without the page field, you cannot tell whether a stable domain is carrying a stable source or a succession of different sources.

Then group results by question. A page can disappear from one answer because the query changed slightly while remaining useful for another. Comparing totals without keeping the question fixed turns ordinary sampling variation into apparent volatility.

The sample also shows why the long tail needs care. Across the broader research set, engines cited 9,192 distinct domains and 21,465 distinct pages. Many of those sources appear infrequently, so a later absence does not prove that the engine rejected the page.

“The long tail is mostly sampling, not change.”

CiteGraph’s methodology makes that claim about the 56 of 102 cited domains that returned after 25 days. It is a useful warning against reading every missing domain as a rejection. It does not explain away the page result: only 14 of 179 exact pages returned, or 8%. Sampling may account for some page-level absence, but this comparison still shows far more movement at page level than among heavily cited domains.

Which pages should you protect when citations rotate?

Protect pages that answer a buyer question in plain, quotable language. The relevant comparison, constraint, definition or recommendation should be explicit rather than buried across several sections.

The page should carry enough context to stand alone. If a citation is extracted without the surrounding navigation, the cited passage still needs to identify the product, category or decision it addresses.

Do not respond to page churn by publishing endless variants. That creates more room for conflicting facts and makes it harder to tell which page deserves maintenance. Build pages with distinct purposes, then keep their claims aligned.

A useful review checks whether the page answers a tracked buyer question, states that answer directly, adds evidence or a comparison that other pages do not, and has a relevant fallback on the same domain.

You are not trying to make one URL immortal. You are building enough useful coverage that a page rotation does not erase your position in the answer.

How can you tell a real citation change from a false alarm?

Read a citation dashboard as a sequence, not a verdict. A missing page citation on one question may be ordinary sampling, a changed answer or another page taking the source position.

First hold the buyer question fixed. Then compare the domain and exact page separately. A stable domain with a changed page is a page-level event. A missing domain may justify a wider review, but it still needs repeated observations before you treat it as a trend.

As noted above, the wider sample contains 9,192 distinct domains and 21,465 distinct pages. That scale helps you watch movement, but it does not turn every absence into a diagnosis.

This is one comparison window from one project’s own site, not a survey of every engine or category. We measure samples, print their margins and do not know how citation behaviour looks outside the measured set. That is the caveat that matters most.