Do lost AI citations come back? 81% stayed gone two weeks later — unless the domain had been cited in 2+ answers, then half returned

Method, stated before the numbers

The instrument is unchanged across all three runs: 12 pinned small-business buyer questions × 4 engines (ChatGPT, Perplexity, Gemini, Claude) × 3 identical samples = 144 production answers per run, citations read from each engine's returned citation payload, domains normalized hostname-minus-www — the same dated cuts behind the whole citation-economy series. The three runs sit six to seven days apart: 2026-07-28 (538 domains), 2026-08-03 (520), 2026-08-10 (540). Each domain gets a three-character presence pattern — cited/absent in run 1, 2, 3. One normalization note for anyone checking our arithmetic: the originally published 48.5% first-window churn used registrable-domain collapse; on the stored cuts the same diff reads 249 of 538 = 46.3%, and the second window's published 43.7% (227 of 520) reproduces exactly. The full cut, including all 47 returning domains by name, is on disk and dated.

The census: 969 domains, seven presence patterns

Read the pool, not just the patterns: a weekly run cites ~540 domains, but 969 distinct domains were cited at least once across three runs — the ever-cited pool is nearly twice the weekly set. 565 of 969 (58.3%) appeared in exactly one run; 225 (23.2%) appeared in all three. And of the 231 domains that entered in run 2, only 68 (29.4%) were still cited in run 3 — 163 were one-week wonders, consistent with our finding that new entrants flicker where incumbents hold.

The resurrection cut: who comes back

Depth before the loss is the discriminator: a domain that had been cited in a single answer and vanished came back 11.8% of the time; a domain that had been cited in two or more answers came back 50.0% of the time — a 4.2× gap on 203 vs 46 losses. The returns are also not random re-rolls: 37 of the 47 returning domains came back on at least one of the same engines that cited them before, and 43 of 47 returned single-engine. The picture that fits all three runs: engines draw shallow citations from a churning pool far larger than any week's answer set, while depth — being the kind of source an engine cites repeatedly — marks a domain the retrieval layer actually knows, which resurfaces even after a zero week.

Limits, named

• One resurrection window. Exactly one cohort (the run-1→2 losses) has had exactly one re-check. The 81.1%/18.9% split is a first reading, not a base rate yet; run 4 adds a second cohort and we will publish the replication either way.
• "Gone" means absent from a 144-answer weekly sample, not provably absent from the engine. Our own within-run data shows 53.9% of citations appear in only 1 of 3 identical asks — some of the 81.1% would reappear if we sampled harder. That entanglement cuts one way: it inflates apparent loss for shallow domains, which is exactly where the return rate is lowest, and it is why the depth split is the more load-bearing number.
• The breadth-2 row is five domains. 80% return on n=5 is suggestive, consistent with the multi-engine survival gradient, and proves nothing.
• One panel, one vertical. 12 buyer questions in our own market, three runs. The pool-vs-set structure may differ on other question sets.

What this means if you track AI visibility

A one-week citation loss is real news only if the citation was shallow — and then it was probably never yours to keep. If an engine cited you in one answer last week and not this week, you are watching the churning 58% of the pool, where two-thirds never return on our data. If it cited you in several answers and you drop to zero, the odds are even that you resurface within a week — panic is premature, and a single weekly snapshot cannot tell you which case you are in. That is a measurable difference, not a slogan: depth and breadth of citation, per question, is what the free 60-second check reads directly from the engines' own citation payloads.