How many sources does each AI engine cite per answer? Perplexity: exactly 20, almost every time. ChatGPT: 7. Claude: none in 44% of answers.

Method, stated before the numbers

Same standing instrument as the rest of the citation-economy series: 12 pinned small-business buyer questions × 4 engines (ChatGPT, Perplexity, Gemini, Claude) × 3 identical samples = 144 production answers, run 2026-08-03 through each engine's production API. Citations are read from each engine's returned citation payload — the structured citation array the API hands back — not scraped from answer prose, and not the consumer web UI, which may render a subset. Counts below are raw citation URLs per answer; deduplicating to distinct domains per answer moves the means by less than half a point (ChatGPT 7.0 → 6.7, Perplexity 19.9 → 18.3), so the shape survives normalization.

The table: four engines, four different citation economies

The Perplexity row is the strangest and the most consequential. 34 of its 36 answers carried exactly 20 citations — a hard-sided retrieval list, the same size whether the question was easy or obscure. That is not an editorial judgment about which sources earned a mention; it is a fixed-capacity container being filled. ChatGPT's count actually varies with the question (4 to 11), which looks like selection. Claude sits at the other extreme: when it cites at all it averages ~9, but in 44% of answers it recommends businesses with no citations whatsoever — visibility on Claude is mostly uncited prose, and a citation tracker pointed at Claude under-reads it structurally.

Why the denominator changes what every citation stat means

• A 'citation' is worth different amounts on different engines. Being one of ~7 sources in a ChatGPT answer and one of exactly 20 in a Perplexity answer are different placements, but every per-engine 'citation share' metric counts them as one unit each. Perplexity alone emitted 717 of the 1,587 citation URLs in this run — 45.2% — so any pooled most-cited ranking is nearly half a Perplexity ranking before the other three engines vote.
• This mechanically explains engine-skewed 'winners.' When we found reddit.com to be the single most-cited domain in this corpus, 32 of its 46 citing answers were Perplexity's — an engine that deals 20 slots per answer has far more room for a community domain to ride along than one that deals 7.
• More slots, more churn. A 20-slot list has a long tail of marginal members, which is consistent with what we measured about instability: the multi-engine citation core replaced a third of its members in six days, and new citations flicker 2.9× more than incumbents. Getting into a 20-slot list is easier — and means less — than getting into a 7-slot one.

The six-day comparison: Perplexity's list grew ~30%

The 2026-07-28 run — same 12 questions, same engines, same 3 samples — read differently on exactly one engine. ChatGPT returned 252 total citations in both runs (mean 7.0 in each — the most stable volume of the four). Gemini read 445 then, 440 now. But Perplexity went from 546 citations (mean 15.2, range 10–18) to 717 (mean 19.9, range 18–20) — its per-answer list grew roughly 30% in six days and pinned to the 20 ceiling. Claude's zero-citation share also moved, from 30.6% (11 of 36) to 44.4%. Two dated runs are a comparison, not a trend — but if Perplexity's slot count is a product parameter, it changed recently, and every week-over-week citation metric on Perplexity moved with it, no matter what any brand did.

Limits

• n = 36 answers per engine, 12 questions, one run date (plus one comparison run). Our panel skews to small-business buyer questions; the slot counts may differ on other query classes.
• API citation payloads, not the consumer UI. The UI may display fewer citations than the API returns; what we measure is the retrieval layer, which is also what most AI-visibility trackers measure.
• We sell this measurement. AskedAbout sells AI-visibility audits, so check the method — which is why it is stated in full.

The practical read: before comparing your citation counts across engines — or paying anyone who does — know each engine's denominator, because which engine recommends your business matters more than a pooled number. A free 60-second check reads ChatGPT and Perplexity on your own buyer questions ($79 audit adds Gemini and Claude, 25 questions × 3 samples), and reports what each engine cited — with the answer, not a share.