147 names appeared, 28 stayed: most AI citations are passages, not places
Three identical runs, five sectors, two engines: two cited names out of three appear only once. And a citation's value depends on the engine - one in two holds on Perplexity, one in twenty on Gemini.
19%
of cited names hold all three runs - 28 out of 147
In this article
Short answer: being cited by an AI and holding a place with it are two different things, and the gap can be measured. We asked the same question three times over, across five sectors, in two languages, to two engines. 147 business names appeared at least once. Only 28 show up in all three runs. 98 - two out of three - appear exactly once: passages, not places.
5 names
the median length of a recommendation list, after cleaning
1 in 2 vs 1 in 20
the chance a citation is a stable place, on Perplexity vs on Gemini
Of 147 names that appeared across three runs, 28 stayed. The 98 that appeared once were not places: passages.
What we measured
The same reference question as every measurement since 23 August - « what are the best X? » - asked three times in a row, identically, across five global sectors, in English and French, to Perplexity and Gemini. 60 runs requested, 9 API failures, 51 usable, collected on the evening of 29 August.
For each sector-language-engine combination we count three things: the union (every name that appeared at least once), the core (names present in all three runs) and the tail (names that appeared exactly once). The core is the only list you can build on; the tail is randomness no optimisation can target.
Both the raw and the cleaned counts are published. Our extractor still lets through fragments that are not business names, and every fragment invents a fictitious place. After cleaning, a median answer contains 5 names (8 raw) - 7 on Perplexity, 4 on Gemini, 5 in English, 4 in French.
One engine builds a core, the other barely does
| Combination | Names appeared | Core | Core share |
|---|---|---|---|
| Hotel booking (en) - Perplexity | 8 | 8 | 100% |
| Payments (en) - Perplexity | 8 | 7 | 88% |
| Payments (fr) - Perplexity | 9 | 4 | 44% |
| CRM (en) - Perplexity | 9 | 3 | 33% |
| Hotel booking (en) - Gemini | 13 | 2 | 15% |
| CRM (en) - Gemini | 13 | 1 | 8% |
| Payments (en) - Gemini | 14 | 0 | 0% |
| Web hosting (en) - Gemini | 12 | 0 | 0% |
Same sector, same language, same day: Perplexity on hotel booking returns the same 8-name list three times - a 100% core. Gemini on payments surfaces 14 names across three runs and keeps none of them from one run to the next.
On the seven combinations where both engines answered, the count is stark: on Perplexity, 23 of the 47 names that appeared hold all three runs - one in two. On Gemini, 3 out of 65 - one in twenty. A « Gemini citation » and a « Perplexity citation » are not the same currency, and a score that adds them up without saying so is mixing gold with tokens.
The median combination has no core at all
The hardest number in this measurement is not an average: the median core per combination is zero. Pick a sector-language-engine combination at random and the most likely case is that no name holds three runs there, while five names pass through once and vanish. Being « recommended by the AI » on a Tuesday says almost nothing about Wednesday.
Three collection failures all sit on Perplexity's side.
Hotel booking in French and web hosting in both languages have no Perplexity runs at all (API rate limits). The engine duel is therefore read only on the seven combinations where both answered - and those are where the 1-in-2 and the 1-in-20 come from.
The running-shoes sector is a failure of ours.
Our noise filter drops labels containing a digit - but shoe models are called « Pegasus 41 » or « Clifton 10 ». That sector's lists come out nearly empty after cleaning: the filter empties them, not the engines. We publish the sector as is rather than quietly removing it.
Three runs, one evening, two engines: a sample, not a census.
Three runs are enough to separate a core from a tail; they are not enough to date how fast names rotate. And nothing says the next night or week would give the same proportions.
What to do with this
Aim for the core, not the list.
A name that appears once across three runs does not « have a place in the AI »: it passed through one run. Before celebrating a citation, ask whether it comes back.
Always ask which engine.
The same sentence - « you are cited » - describes a 50%-stable position on Perplexity and a 5% lottery on Gemini. A report that does not name the engine tells you nothing.
Re-measure, identically, over time.
It is the only way to tell a place from a passage - a single run cannot, by construction. That is exactly why our monitoring re-asks the same question every 3 days.