The “durable core” of AI citations kept its size — and replaced a third of its members in six days.

Method, stated before the numbers

Same instrument, both runs: 12 pinned buyer questions × 4 engines (ChatGPT, Perplexity, Gemini, Claude) × 3 samples = 144 answers per run, citations read from each engine's returned citation array, URLs collapsed to registrable domains, one count per (domain, answer). A domain is “multi-engine” if ≥2 distinct engines cited it anywhere in the run. Run one (2026-07-28) cited 538 distinct domains, 99 of them multi-engine; run two (2026-08-03) cited 506, again 99 multi-engine. Today's cut diffs the two 99-domain membership lists. The raw diff, with per-domain engines and answer counts, is the dated cut we publish and share on request.

Where today's 99-domain core came from

And the outflow that made room, computed from the same two lists: of the 99 domains that were multi-engine on 07-28, 65 stayed multi-engine (65.7%), 27 were demoted to a single engine (27.3%), and 7 were cited nowhere in run two (7.1%). The turnover happens to balance exactly — 34 in, 34 out — which is why the core's size looks stable while a third of the names on the list changed. A stable aggregate hiding heavy churn underneath is exactly the kind of thing a one-off AI-visibility screenshot cannot show.

The entry rates: the door is open, and measured

Two admission rates fall out of the diff. From the 439-domain single-engine tail, 26 domains — 5.9% — climbed to multi-engine within six days. And of the 229 domains that entered the corpus fresh in run two, 8 — 3.5% — landed multi-engine immediately, skipping the tail entirely. Put next to yesterday's survival split (single-engine citations survive at 42.1%, two-engine at 90.2%), the ladder is now measured at every rung: uncited → cited-once is a lottery with weekly drawings, cited-once → multi-engine happens to ~1 in 17 tail domains per week, and once you are multi-engine you mostly stay.

The 34 new members, in full

Which engines are the door

Admission is not evenly distributed across engines. Gemini participated in 25 of the 34 admissions and Perplexity in 23; ChatGPT in 11 and Claude in 14. The single most common combination is Gemini + Perplexity (11 of the 29 two-engine admissions). That is consistent with what we measured on retention: Gemini and ChatGPT keep less than half of their own citation sets week to week, so they are also the engines most often handing out new citations. The engine most likely to admit you and the engine most likely to keep you are not the same engine.

Limits

• Sampling luck is real at the entry boundary. Multi-engine membership takes as little as one citation from each of two engines; with 3 samples per question, a domain can cross the line by draw. The direction of the bias is known: some of the 34 “new” members are variance, so 34.3% is an upper bound on true six-day turnover — the incumbents' 65.7% retention is the harder number.
• Six days, two runs, one panel. 12 buyer questions about AI visibility for small businesses; the instrument runs weekly, so the turnover series accumulates from here.
• We have a commercial interest. We sell AI-visibility audits. Every number is reproducible from the two dated citation cuts, shared with anyone who asks.

The practical reading: yesterday's cut said a multi-engine citation is the only durable kind; today's says the club that confers that durability admits new members every week — about 1 in 17 of the single-engine tail climbed in, in a single six-day window, and 8 domains went from invisible to multi-engine in one step. Which side of the door your business is on is a measurable fact, not a guess — run the four-engine measurement on your business in 60 seconds.