**TL;DR** — Across 23 client sites in April 2026 we measured citation persistence: once a page has been cited by Google AI Overviews on a given query, how long does it stay in that query's cited URL set before being replaced? Our cohort median was 19 days; the 25th percentile dropped out inside 7 days, and only 11% of (query, URL) pairs remained continuously cited across the full 30-day window. Three drivers explained most of the early drop-out — a fresher competitor publish on the same intent, structural drift on the previously-cited page itself, and what we call "composer topic drift" inside the engine's understanding of the query. Citation is not a stable state; it is a continuously re-evaluated one, and the operational story looks more like lease renewal than one-time acquisition.
Why persistence matters more than the first citation
Through 2025 most of the AI-search reporting we built and read was about acquisition — does the page get cited, yes or no, and how fast (we covered the first-citation half of the lifecycle last week with TTFC). That framing was useful when the question was whether AI search was real as a channel. It is not sufficient now that the channel is established. The dashboards we ship today look fine on month-over-month citation rate; the failure mode they hide is rapid churn underneath the headline number. A site can be cited on 60% of basket queries this month, cited on 58% next month, and have lost half of last month's specific cited URLs along the way. The aggregate looks stable; the underlying portfolio is not.
Persistence matters more for revenue than first citation does. The downstream effects of an AI Overview citation — branded search uplift, referral clicks, model-training exposure — accumulate over weeks rather than days. A page cited for 25 days produces more of all three than two pages each cited for 7 days. When the churn is hidden inside an aggregate citation-rate number, teams over-invest in volume (more pages, more queries) and under-invest in retention (keeping the pages already cited worth re-citing). The metric to optimise is the median duration, not the monthly rate.
How we ran the audit
23 client sites — 9 SaaS, 7 publisher, 4 DTC and 3 B2B services. Same 60-query commercial-intent basket per client we use for our other audits. Captured twice daily via Playwright at 09:00 UTC and 21:00 UTC across April 2026. For every (query, cited URL) pair we counted consecutive days the URL appeared in that query's cited set, treating any single missed capture as a non-event but two consecutive missed captures as a citation drop. The window is 30 days because shorter windows did not give the long tail a chance to declare itself, and longer windows pulled in too many editorial changes on our side that polluted the signal.
The unit of analysis is the (query, URL) pair, not the page. The same URL can be cited on five queries and drop out of three of them in a single week — that is three drop-out events, not one. We also separated "drop and return inside seven days" from "drop and stay gone." Returns inside seven days were common — the composer occasionally tries an alternate citation and reverts within a couple of crawl cycles, and we treat that as one continuous citation, not two. Drops longer than seven days are real lease expiries and they are what the median below counts.
The shape of the decay curve
The decay curve is roughly exponential with a fat tail. Median citation duration was 19 days; P25 was 7 days, P75 was 33 days. About 11% of (query, URL) pairs persisted across the entire 30-day audit window. Importantly, the per-day drop-out hazard is not constant — pairs that survive the first 14 days drop out at roughly half the per-day rate of pairs still in their first 14 days. Once a page has been continuously cited for two weeks, it becomes materially more durable than a freshly cited one. That non-linearity is the lever — early retention compounds.
Two cohorts behaved very differently from the median. SaaS product feature pages — the kind that map cleanly to a single buying-intent query — had a median of 27 days; publisher long-form articles on commodity topics had a median of 11. The composer treats canonical feature-to-query mappings as low-risk and re-cites them by default; it treats long-form publisher content as more interchangeable and rotates more freely. The implication for SaaS clients is that retention investment pays back faster; for publishers the same level of investment buys less retention because the baseline churn is higher.
Driver one: a fresher competitor publish on the same intent
The single best predictor of citation drop-out in our cohort was a competitor publishing a new page on the same query within the prior 14 days. Pairs where such a publish happened were 2.8× more likely to drop out the following week than pairs without one. The composer is biased toward freshness on commercial-intent queries — a 60-day-old page can still be excellent, but a 3-day-old well-structured page from a similar-authority domain is a credible challenger, and the composer routinely tests the swap. The swap sticks roughly 40% of the time; the other 60% the original page is re-instated within a week. Either way the disruption is measurable.
This produces a counter-intuitive operational consequence: a page that is winning citations does not need new keywords, it needs a defensive publish cadence. We now treat any cited page that has not been substantively edited in 45 days as at-risk and edit it — typically a paragraph update with fresh data, a refreshed FAQ section, and an updated `dateModified` — on a 30-to-45-day rolling basis. We do this only on commercial-intent pages, not on long-tail. The edit is small, the citation-duration delta is about 9 days in our sample, and the work fits in the weekly editorial cadence rather than needing a separate retention sprint.
Driver two: structural drift on the previously-cited page
The second driver is changes to the page itself that the composer reads as a low-confidence signal. Examples we caught in audits this quarter: a canonical tag flipping (often during a CMS migration), `dateModified` stale-dating because the team believed leaving it alone was the safer choice, schema markup drifting from valid to malformed after a plugin update, a CDN-edge cache returning a different rendering to Googlebot than to humans, and an experimental A/B test variant being served to the bot. None of these changes the content the user reads, but each shifts the composer's confidence — and a small confidence drop is enough to lose a citation when a competitor is one swap away.
We instituted what we call a "cited-page change log" in our retainer engagements: any deploy that touches a page currently in the cited URL set gets a pre-deploy schema validation, a canonical check and a render-diff against the prior Googlebot capture. The cost is roughly five minutes per affected page per deploy; the citation-retention delta is about 6 days. The mistake is treating cited pages as ordinary pages in the deploy pipeline — they are load-bearing structurally, and small unintentional changes are disproportionately expensive. The pages doing the most revenue work need the most deploy discipline, not the least.
Driver three: composer topic drift
The third driver is the subtlest. The composer's understanding of what a query "means" drifts over weeks. The query string is constant; the implied intent the composer reaches for shifts as it absorbs new web content, new freshness signals, and new user-side behaviour. We saw this most clearly on "what is X" queries where, over 30 days, the composer migrated from a definition-heavy answer (which cited explainer-style pages) toward an example-heavy answer (which cited product-comparison pages). The originally cited URL did not get worse; the question itself quietly became a different question, and the page that answered the old framing was no longer the best answer for the new one.
Topic drift is impossible to prevent and partially possible to track. We added a weekly "answer-text diff" to our audits — for each basket query we record the first 800 characters of the AI Overview answer and diff it week over week. When the diff exceeds a threshold (we use 40% of the prose changed) we flag the query and review whether the previously-cited page still answers the drifted version. The answer is often "partially" — the page still answers the original framing but is no longer the best answer for the new one. That is the editorial decision point: rewrite to follow the drift, or accept the citation loss and recover somewhere else in the basket.
What changed in our renewal checklist
Four additions. We require monthly review of every page currently in any client basket's cited URL set, with a refresh action — even minor — on any cited page not edited in the prior 45 days. We require pre-deploy structural validation (schema, canonical, render-diff) on any deploy touching cited pages. We require a weekly competitive-publish scan against the basket queries, flagging any competitor publish on the same intent within the prior 14 days. And we run the answer-text diff weekly to catch composer topic drift before it becomes a citation loss the client reads as inexplicable.
We dropped one habit. Through 2025 we reported "monthly citation rate" as our flagship retention metric. In 2026 we report median citation duration as the headline and the rate as a secondary stat. The two move differently — and the second-derivative story (durations falling while rates hold steady) is the most common early-warning pattern we now flag to clients, typically two to three weeks before it shows up as a rate drop on the rolled-up monthly chart everyone else is watching.
- 01Measure citation persistence per (query, URL) pair, not just citation rate. Our cohort median was 19 days; P25 7 days; only 11% of pairs persisted continuously across the full 30-day window.
- 02Treat cited pages as load-bearing. A refresh every 30–45 days adds roughly 9 days of citation duration in our sample; a pre-deploy structural check on cited pages adds roughly 6.
- 03Track competitor publishes on basket queries weekly. The presence of a fresh same-intent competitor publish doubles drop-out hazard the following week.
- 04Diff the AI Overview answer text week over week. Composer topic drift accounts for a meaningful slice of "we lost the citation but our page did not change" cases.
Where this argument breaks
For sites cited on fewer than about 20 (query, URL) pairs in any given month, the per-pair statistics are too sparse to read and the retention investment is hard to justify until the cited portfolio is larger. For programmatic sites with thousands of cited templated pages, the median is dominated by template-level patterns rather than individual page behaviour, and the audit needs a per-template view rather than a per-page one. In Chinese-language search, 文心 and 通义 use different rotation rhythms — 通义 in particular rotates citations on commercial queries roughly weekly, and our 19-day median does not transfer. Outside those carve-outs, citation persistence is the quietest variable on most teams' AI-search dashboards, and the operational ground available to recover from it is wider than the time-to-first-citation work everyone is rushing to do this quarter.