How many sites compete for one AI answer? 35 to 93 domains per question — and the width sticks to the question

Method, stated before the numbers

The instrument is the one behind our whole citation-economy series, 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 — never from answer prose — and domains normalized hostname-minus-www. The runs are dated 2026-07-28, 2026-08-03 and 2026-08-10, six to seven days apart. The new cut counts, for each question in each run, the distinct domains cited across that question's 12 answers. Width rank-stability across runs is Spearman rank correlation on the 12 questions: 0.78 (run 1 vs 2), 0.69 (run 2 vs 3), 0.79 (run 1 vs 3). The full per-question cut — per-run counts, three-run unions, persistence splits, top cited domains — is on disk, dated 2026-08-13.

The per-question census: every question, every run

Two structural facts. First, the spread is wide and real: within every single run, the widest question fields 2.3–2.6× the domains of the narrowest (93 vs 40, 87 vs 37, 91 vs 35) — same engines, same days, same sampling. Second, the ordering barely moves: "What tools track whether AI recommends my business?" is the widest field in all three runs (tied with the audit question in the third), the one-time-audit question never fields more than 45 domains in any run, and rank correlation across the full panel never drops below 0.69. Width is not sampling noise — it is a property of the question. A per-run field of ~90 domains across just 12 answers also means almost every answer draws a substantially different source set; the three-run union runs 72–177 domains per question against a mean per-run width of ~62.

Narrow questions keep their sources; wide questions re-draw them

The persistence split is the finding we did not expect: narrow questions hold 32.3% of their ever-cited domains across all three runs; wide questions hold 14.4% — a 2.3× gap. A wide field is not a bigger pond, it is a faster-churning one: the widest question ("What tools track whether AI recommends my business?") cited 177 distinct domains across three runs and kept only 24 of them in all three. This refines what our 81%-stayed-gone census measured at corpus level: churn is not uniform — it concentrates on wide questions. One first-party data point sits exactly on this gradient: the only question on which the engines have ever cited askedabout.com — Perplexity, 3 of 3 samples, on the Profound-alternative question — is the panel's third-narrowest field (78 union domains, 32.1% persistent). One citation on one question proves nothing by itself; we note where it landed because the gradient predicts exactly that kind of placement.

Limits, named

• 12 questions, one vertical, three runs. The width ordering is stable on our panel of small-business AI-visibility questions; the 35–93 range is a property of this panel, not a universal constant.
• Width is bounded above by sampling. A question's field is counted across 12 answers per run; harder sampling would widen every field. Our own within-run instability number (53.9% of citations appear in 1 of 3 identical asks) says the true pools are larger everywhere — but that cuts against the wide questions' persistence too, so it does not rescue them.
• Persistence groups are 3 questions each. The 32.3%-vs-14.4% split aggregates 223 vs 508 domains, which is a real sample — but the grouping (top/bottom 3 by union) is one reasonable choice among several. The full 12-question table is above so any other grouping can be checked.
• One engine family per question could drive width. We did not split width by engine here; per-engine field sizes are in the per-engine cut and a per-(question × engine) cut is a natural follow-up.

What this means if you want AI engines to cite you

"How visible is my business in AI search" is not one question — it is a portfolio of differently-sized contests. On a wide question, ~90 domains rotate through 12 answers and 86% of them will not be there in three weeks; landing there is easy and keeping it is the hard part. On a narrow question, a third of the field is fixed — harder to enter, but entries stick. If you are choosing where to compete, the narrow, high-intent questions ("alternative to X", "one-time audit") are where a citation, once won, is most likely to still be there next week. Which questions are wide and which are narrow for your market is measurable, not guessable — the free 60-second check reads the engines' actual citation sets on the questions buyers ask about businesses like yours.