AI citations turn over fast: in a 2026 study of 1,127 cited URLs, only about 10.6% were still being cited six weeks later. Digital Authority Partners tracked those URLs across ChatGPT, Perplexity, Gemini, Microsoft Copilot and Google AI Overviews in three waves, 28 days apart, and just 119 of them appeared in all three. The other ~89% rotated out at least once. A citation is a position you defend over weeks, not a win you bank once.
Last reviewed: 15 August 2026. All figures below come from Digital Authority Partners' 2026 AI Visibility Study (published 1 July 2026), a single-vendor, single-period read. It is directional, not definitive, so cite "Digital Authority Partners, 2026" with the date when you reuse a number, and treat the direction as firmer than any one percentage.
How long does an AI citation actually last?
On this data, a matter of weeks. Across the full six-week window, the study found only 10.6% of cited URLs (119 of 1,127) persisted across all three sampling waves. The commentary around the study coined a useful frame for it: AI citations have a short half-life — the pool of sources an engine leans on is mostly replaced within about a month, whether or not the underlying pages change.
This is a different failure mode from a page slowly aging out. It is the whole population of cited sources reshuffling. The study reports that roughly 40-60% of cited sources rotate every month, and that between the first and last wave the two waves shared only about 20% of their URLs (Wave 1 surfaced 530 unique URLs, Wave 3 surfaced 546). Nearly a quarter of the query-and-engine pairs tested saw complete turnover — zero overlap between the first and last wave.
What exactly did the six-week study measure?
A longitudinal panel of AI citations, not a one-time snapshot. The design is what makes it useful: instead of asking "who gets cited" once, it asked the same questions three times over six weeks and watched which cited URLs survived.
| Attribute | Detail |
|---|---|
| Unique URLs tracked | 1,127 |
| Engines | ChatGPT, Perplexity, Gemini, Microsoft Copilot, Google AI Overviews |
| Sampling | 3 waves, ~28 days apart, over 6 weeks |
| Persisted across all 3 waves | 119 URLs (~10.6%) |
| Monthly source rotation | ~40-60% |
| Wave 1 ↔ Wave 3 URL overlap | ~20% |
| Publisher / date | Digital Authority Partners, 1 July 2026 |
Source: Digital Authority Partners, 2026 AI Visibility Study. The one-line read: the set of pages an AI engine cites is mostly not the same set a month later. Because it is one vendor measuring one window, read the magnitudes as directional; the pattern of heavy churn is what carries.
Which engines hold their citations longest?
Retention varied widely by engine — from about 11% to 44% over four weeks. The more an engine leans on live retrieval and recency, the faster its cited set churned.
| Engine | 4-week citation retention |
|---|---|
| Perplexity | ~44% (most stable) |
| Microsoft Copilot | ~34% |
| ChatGPT | ~31% |
| Google AI Overviews | ~27% |
| Gemini | ~11% (most volatile) |
Source: Digital Authority Partners, 2026. Even the most stable engine, Perplexity, dropped well over half its cited URLs in four weeks. The takeaway is not "optimize for Perplexity because it is stable" — it is that no engine treats a citation as durable, so any single-engine, single-snapshot reading of your visibility is unreliable. This mirrors the low cross-engine overlap the same study found (78-85% of cited domains were unique to a single platform), which is consistent with our read that engines cite largely different pages.
Why do AI citations churn so fast?
Three mechanisms stack, and together they explain a ~90% monthly turnover. None of them require your page to get worse.
- Answers are probabilistic. Ask an engine the same question twice and it samples a slightly different response, so the cited set moves on its own. This is answer volatility, quantified separately: an identical question asked repeatedly in a single day returned the same brand set only 50-61% of the time. See why your AI share of voice swings week to week.
- Retrieval favours fresh pages. Live retrieval pulls recently-updated content, so a page cited today can be displaced by a fresher competitor tomorrow — the 3-month citation cliff acting on the whole field, not just your one page.
- The web underneath keeps changing. New pages get published, others get updated, and the index the engine draws from shifts continuously, so the candidate pool itself turns over.
Within-day volatility, freshness decay and an ever-changing index compound. A citation you earned in one wave has to survive all three pressures to still be there a month later — and on this evidence, most do not.
An AI citation is not a trophy you win and keep. It is a position you hold for a few weeks against a cited pool that is ~90% replaced within a month. Presence in AI answers is defended, not banked.
Does this contradict our freshness and volatility pieces?
No — it is the third leg of the same stool, and it sharpens the other two rather than fighting them. It is worth stating plainly so the corpus stays consistent.
- Answer volatility describes one moment: the same prompt returns a different answer each time it is asked. That is the within-request noise.
- The freshness cliff describes one page: your content loses citations as it ages relative to fresher rivals. That is the decay of a single asset.
- Citation turnover — this piece — describes the whole cited pool over time: aggregate across many prompts and weeks, and about 90% of the sources rotate monthly. It is what you get when within-moment volatility and per-page decay compound across the field.
All three point to the same operating conclusion: read your AI visibility as a smoothed trend across many prompts and engines, and keep re-earning the citations you care about. None asserts a fixed number the others deny.
What should you do about citation turnover?
Treat citation as a maintained position, and measure it as a distribution — not a one-off win. Concretely:
- Know which pages are cited now. You cannot defend a position you cannot see. Track citation coverage per engine so you know what to protect before it rotates out.
- Put your highest-value cited pages on a refresh cadence. Because freshness is one of the churn drivers, substantively updating (not date-bumping) the pages tied to buyer queries is how you re-win the citation, per the freshness playbook.
- Don't rely on one cited URL. With 40-60% monthly rotation, breadth is insurance: earn citations across several strong pages so losing one wave does not zero out your presence.
- Sample many prompts, repeatedly, across every engine. A single snapshot of a handful of prompts will read the noise, not the signal. Average across a large prompt set and watch the trend, the method behind measuring AI visibility and the reason share of voice is read as a distribution.
The number that matters is not whether you were cited in one snapshot — it is whether your share of voice holds up as the cited pool churns beneath it, measured across engines and over time. Continuously tracking which of your pages are cited, on which engines, and when they slip is exactly what Buffy Intel is built to do. Questions: [email protected].