Ask the same question three times: 54% of the domains AI engines cite show up in only one answer.

Method, stated before the numbers

Same instrument as the six-day churn series: 12 pinned buyer questions × 4 engines (ChatGPT, Perplexity, Gemini, Claude) × 3 samples each = 144 answers in one run (2026-08-03), citations read from each engine's returned citation array, URLs collapsed to registrable domains — 506 distinct domains. The new cut: for every (question, engine) cell we hold the three answers to the identical prompt side by side and score each cited domain by how many of the three samples cite it — 1, 2, or 3. That yields 854 scored (question, engine, domain) triples. A domain scored 3-of-3 is a citation you can rely on being there when a customer asks; a 1-of-3 is a coin the engine flipped once.

The distribution: one-off citations are the majority

Per engine: “cited by ChatGPT” and “cited by Perplexity” are different claims

A concrete cell: on “What tools track whether AI recommends my business?” ChatGPT cited 9, 8 and 9 domains across its three samples — 26 citation slots — and exactly one domain appeared in all three answers. Everything else rotated. Meanwhile Perplexity's three samples of a question are usually near-identical lists: 76% of its citations held across all three asks. The engines are not equally random: a Perplexity citation is a property of the question; a ChatGPT or Gemini citation is closer to a property of the individual answer. Claude is a different failure mode again — it returned an empty citation list in 16 of its 36 samples, so for Claude the unstable thing is whether citations exist at all.

Why this matters for the weekly churn numbers

Our six-day diff found ChatGPT and Gemini retain under half of their own cited domains week to week while two-engine domains survive at 90.2%. Today's cut says much of that weekly churn is not the index changing its mind over six days — the variance is already there sixty seconds later. It also says single-screenshot AI-visibility reports — one ask, one answer, one green checkmark — are sampling a distribution and calling it a fact: on ChatGPT and Gemini, the modal citation (72%) would not have been in the next screenshot. Any honest measurement has to ask repeatedly and report frequencies, which is why our own instrument runs three samples per engine per question and scores per-sample agreement.

Limits

• Three samples is a floor, not a full distribution. A 1-of-3 domain may be a stable ~33%-propensity citation rather than a one-off; with more samples some singles would recur. The 3-of-3 figure (28.9%) is the hard number — those held every time we asked.
• Engines that cite more domains per answer mechanically carry more one-offs. Gemini contributed 323 of the 854 scored citations; its 71.8% single-sample share partly reflects long, rotating citation lists.
• One run, one panel. 12 buyer questions about AI visibility for small businesses, run 2026-08-03; the weekly runs accumulate the series from here.
• We have a commercial interest. We sell AI-visibility audits. Every number is reproducible from the dated 144-answer citation cut, shared with anyone who asks.

The practical reading: if a tool — or a competitor's report — shows you "cited by ChatGPT" off a single ask, the odds it would have said the same thing on the very next ask are about one in three. Stability, not presence, is the thing worth measuring: run the four-engine, three-sample measurement on your business in 60 seconds.