5 min read · updated 5 August 2026
How AI engines choose which business to name
Short answer
Each engine resolves local recommendations through a different index. ChatGPT and Copilot lean on Bing, Foursquare and Yelp. Gemini and AI Overviews weight Google Business Profile. Perplexity draws heavily on community discussion. Optimising for one covers roughly a quarter of the surface.
The mistake almost everyone makes
Treating “AI search” as one thing.
It is at least four things, each reading a different set of sources. A practice can be named consistently by Gemini and appear in nothing from ChatGPT — not because one engine likes it more, but because the two are reading different data.
What each engine reads
| Engine | Primary index | What moves it |
|---|---|---|
| ChatGPT | Bing | Bing Places, Foursquare, Yelp, review volume and recency |
| Gemini / AI Overviews | Business Profile completeness, schema, Google reviews | |
| Perplexity | Own index + community | Community discussion, forums, recent coverage |
| Copilot | Bing | Bing Places plus editorial coverage |
The practical consequence: an agency working only on Google is working on roughly one surface out of four, and the reporting it hands you will look fine.
Why third-party sources dominate
Across studies of AI citation behaviour, brands are consistently found to be around six and a half times more likely to be cited through third-party sources than through their own domain.
For local healthcare the gap widens further, because the engine is answering a question your website structurally cannot answer well: which of these businesses is actually good? Your site is not a credible source on that. Reviews, directories and independent coverage are.
This is the single most useful thing to internalise. The instinct is to improve the website. The leverage is almost entirely elsewhere.
Review quality is doing more work than it appears
Every engine, through one path or another, ends up weighting reviews — because reviews are the closest available proxy for the question being asked.
Two properties matter:
- Volume, with a practical floor near 30 at 4.3 stars, higher in competitive metros
- Recency, weighted comparably to volume
A steady trickle beats a historical pile. This is why review velocity is a system to build rather than a campaign to run.
Structured data resolves identity, not ranking
Schema markup is frequently oversold. It does not persuade an engine to recommend you.
What it does is let an engine resolve which entity a page is about — this practice, this practitioner, this service, this address, these hours. For multi-location groups that disambiguation is genuinely hard, and getting it wrong means an engine may not connect your locations to your brand at all.
Necessary. Not sufficient.
The signals move
Any document like this has a shelf life.
In recent measurement, community content’s share of AI citations shifted by double digits in a single quarter, and the most-cited social platform changed inside a year. Source maps built twelve months ago are describing a different system.
That instability is not a reason to ignore the channel. It is the reason measurement has to be continuous rather than one-off — you cannot manage against a snapshot of something changing this fast.
What is knowable, and what is not
No platform publishes its source-selection logic. Everything above is inferred from observed output: running fixed prompt sets repeatedly across engines and recording what comes back and which sources are cited.
That is empirical measurement of behaviour. It is not a claim about internal ranking factors, and anyone presenting it as one is overstating what can currently be known.