Half the domains AI engines cite were gone six days later — unless more than one engine cited them.

Method, stated before the numbers

Both runs are production runs of the same instrument: 12 pinned buyer questions × 4 engines × 3 samples = 144 answers per run, citations read from each engine's returned citation array (not a substring test on prose), URLs collapsed to registrable domains, one count per (domain, answer). Run one (2026-07-28) cited 538 distinct domains; run two (2026-08-03) cited 506. A domain "survived" if any engine cited it anywhere in run two. Nothing about the questions, engines, sampling, or extraction changed between runs — the only variable is six days of engine behaviour.

Survival by how many engines cited you

The overall number — half the citation layer replaced in under a week, with 229 domains entering that were cited nowhere in run one — is the one that should reframe how a business reads any single AI-visibility report. But the split is the finding: the single-engine tail, which is 81.6% of everything the engines cite, churns at nearly 6 in 10, while the multi-engine core barely moves. Depth predicts durability the same way: domains cited on one answer survived at 28.3%, on 2–4 answers at 70.8%, on 5+ answers at 94.6%.

Engines differ in how much they forget

ChatGPT and Gemini kept fewer than half of their own citation sets across six days; Perplexity kept 71.1% of its. If your AI-visibility evidence is "ChatGPT cited this source last week," the base rate says that observation has roughly a coin-flip's chance of still being true this week — which is worth knowing before anyone builds a strategy, or sells one, on a single reading.

The series since: ten weekly pairs, 2026-07-28 to 2026-10-05 (added 2026-10-08)

The instrument has run every Monday since this post. Each run is the same 144 answers; each pair below diffs one run against the next. A domain in the earlier run survived if any engine cited it anywhere in the later run. The rows come from the committed weekly cuts (growth/citations/cited-domain-cut-<date>.json) through content/news/survival_series_2026-10-08.mjs; the artifact is content/news/survival-series-2026-10-08.json, 19 of 19 self-checks against the cuts' own diff fields.

One instrument change sits inside the series. On 2026-09-20 (commit 4ceb08f) the citation reader was narrowed to the sources an answer actually cites, instead of every source the engine returned. Perplexity's cited URLs fell from 720 to 285 per run on that change alone (Claude 222 to 127; Gemini 498 to 431; ChatGPT 265 to 259), so the 09-14 to 09-21 pair (marked *) mixes engine behaviour with the reader change and is not like-for-like. The two pairs after it are like-for-like under the new rule. The seven pairs before it are like-for-like under the old one.

The gradient held in all ten pairs. Single-engine survival stayed between 44% and 53% in the seven pairs before the instrument change and read 46% and 37% in the two after it. Domains cited by two engines survived at 84–92% in every pair but the newest (70.9%, 39 of 55, the lowest of the series). Three-engine domains survived at 100% in seven pairs and never below 91%; four-engine domains at 100% in nine pairs of ten. Depth says the same thing: domains cited on a single answer survived at 23–33% per pair, on 2–4 answers at 54–76%, on 5 or more answers at 86–99%. The 08-03 to 08-10 row is the pair reported the week after this post (43.7% weekly churn).

The same engine, two runs apart. The table below counts domains an engine cited in the earlier run that the same engine cited again in the later one. Perplexity's drop after the asterisked row is the reader change: its retrieved-but-uncited sources no longer count, so its cited set shrank from 182 domains (09-14) to 95, 89 and 66 in the three runs since, and the share it keeps week to week fell with it. ChatGPT, Gemini and Claude each kept between a third and seven tenths of their own citation sets in every pair.

Two stagings of the 08-03 run. The figures in the sections above come from the cut staged 2026-08-04 (506 distinct domains on 08-03; 277 of 538 survived, 51.5%). The weekly series file for the same run, staged 2026-08-08 under the series' domain rule (hostname minus www), holds 520 domains, and the 07-28 to 08-03 row in the tables here is diffed against that file: 289 of 538 survived (53.7%), single-engine 44.4%, two-engine 91.8%, three-engine 100%. The twelve extra survivors are domains the two rules split differently. Both readings are kept; the gradient is the same in each.

What the series changes about the reading. The post's claim was one pair. Ten pairs later the shape has not moved: being cited by one engine on one answer is a coin flip with worse-than-even odds of being there next week, and being cited by three or more engines has been close to permanent through the whole series. The multi-engine core itself shrank after the reader change (98–101 domains per run before it, 72, 76 and 62 since) because each answer now contributes fewer domains, not because engines agree less; a comparison of core size across 2026-09-20 is not valid.

Limits

• Sampling and churn are entangled at the thin end. Each run samples each engine 3 times per question, so a domain cited once in run one can miss run two by sampling luck rather than any change in the engine. That is precisely the operational point — a domain seen once in 144 answers has no stable claim on the answer surface — but it means the 42.1% tail survival mixes "dropped" with "was never reliably there." The multi-engine rates, measured on domains too widely cited for sampling luck, are the clean signal.
• Two runs, six days, when published; ten pairs by 2026-10-08. The weekly series is tabled above. One instrument change (2026-09-20) splits it into seven like-for-like pairs under the old citation reader and two under the new; the pair that crosses the change is marked and not scored.
• One question panel. 12 buyer questions about AI visibility for small businesses; other topics may churn differently.
• We have a commercial interest. We sell AI-visibility audits. Every number above is reproducible from the two dated citation cuts, which we share with anyone who asks.

The practical reading for a business: a screenshot of one AI answer naming you is weather, not climate. What persists is being cited by more than one engine, on more than one answer — the 99-domain core that survived at 90–100% in the first pair and at 79–95% in every pair since. To see which engines currently cite anything about your business — and whether it is core or tail — run the measurement on your business in 60 seconds.