Are new AI citations stable? Domains that survived a week held 3-of-3 identical asks 36.3% of the time — new entrants 12.4%, and ChatGPT gave them zero

Method, stated before the numbers

Same instrument as the rest of the citation-economy series: 12 pinned small-business buyer questions × 4 engines (ChatGPT, Gemini, Claude, Perplexity) × 3 identical samples = 144 production answers per run, citations read from each engine's returned citation payload. Two dated runs: 2026-07-28 (538 distinct cited domains) and 2026-08-03 (520). Domains are normalized hostname-minus-www — the same rule as our published dated cuts; the within-run stability post used registrable-domain collapse, so its 854 citation triples and 28.9% read here as 861 and 29.2% — same data, stated normalization difference. The join is mechanical: every (question, engine, domain) citation in the 08-03 run is scored by how many of the 3 identical samples it appeared in, then labeled incumbent (the domain was cited anywhere in the 07-28 run) or new entrant (it was not).

The split: surviving a week predicts holding three asks

Three-quarters of new-entrant citations (74.8%) appeared in exactly one of three identical asks. A domain the engines were already citing a week earlier is 2.9× more likely to hold every sample. The week-to-week churn we measured before — 48.5% of cited domains gone in six days — is not a separate fact from within-run flicker; the flickers are overwhelmingly the newcomers, and the newcomers are what churns.

It is not an engine-mix artifact

Perplexity is by far the most stable engine (76.4% of its citations hold 3-of-3) and also cites many incumbents, so the headline gap could in principle be composition. It is not — the gap holds inside every engine:

ChatGPT and Claude between them handed new entrants 89 citations in this run and did not repeat a single one across all three asks of the same question. Engine breadth shows the same gradient inside the run: citations from single-engine domains held 3-of-3 20.1% of the time, two-engine domains 36.6%, three-plus-engine domains 44.2% — consistent with the six-day survival gradient (42.1% single-engine vs 90.2% two-engine).

What this means if you chase AI citations

A first-time citation is weak evidence by itself: on these numbers it is a one-in-eight shot of being reproducible even minutes later, and roughly a coin flip to exist at all next week. What actually distinguishes durable visibility is repetition across time and engines — the incumbents' 36.3% and the multi-engine 44.2% are the same signal read on two axes. Practically: measure a citation twice before celebrating it once. That is the design rationale for running the same question three times in our own audits.

Limits

Four, named. One run pair: incumbency is defined against a single earlier run (2026-07-28); a longer citation history would separate true incumbents from twice-lucky flickers. Three samples is a floor: a 1-of-3 citation may be a stable ~33%-propensity citation rather than noise; the 3-of-3 figures are the hard numbers. Domain-level incumbency: a domain counts as incumbent even if its 07-28 citation was on a different question or engine than its 08-03 one; question-level incumbency would be stricter. Gemini's payload is bare domains, so its rows join on domain like the others but cannot be checked at URL level.