Local AI search

How does ChatGPT find local businesses?

Three piles: what both parties confirmed, what a larger study refuted, and what nobody has published.

How does ChatGPT find local businesses?

OpenAI discloses almost nothing. Its help documentation says only that ChatGPT may share a user location with trusted third-party providers to supply nearby business listings, and it names no local data vendor. One partnership is confirmed by both sides: Yelp reviews, ratings, photos and business details, plus reservations and waitlist, announced by Yelp in August 2026 and described as non-exclusive. The widely repeated claim that Foursquare supplies most of ChatGPT's local results was refuted by a 2,879-query study that measured Foursquare-controlled domains at 0.00 percent of citations. How businesses are ordered once retrieved is not documented by anyone.

What OpenAI has actually said

The entire first-party disclosure on local data is one phrase. OpenAI's ChatGPT search help page states that ChatGPT may share a user's location with trusted third-party providers that help supply more local information, such as nearby business listings or weather reports, and that the IP address itself is not shared with those providers.

It does not name a single local data vendor. OpenAI's crawler documentation covers four bots and contains nothing about ranking, local business data, maps or providers. The ChatGPT search launch announcement names publishers and data partners across categories including maps, but the named list is media, not a points-of-interest vendor.

That is the whole primary record. Anything more specific that you read about ChatGPT local sourcing is inference, measurement by a third party, or invention, and it is worth knowing which.

What is confirmed by both parties

Yelp. On 10 August 2026 Yelp announced an integration bringing Yelp Reservations and Waitlist inside ChatGPT along with Yelp reviews, ratings, photos and business details across thousands of top restaurants in the United States and Canada. The underlying licensing arrangement was reported in July 2026 and described as non-exclusive.

Yelp says nothing about how businesses are selected or ordered within that surface, and neither does OpenAI. A confirmed data supply relationship is not a confirmed ranking mechanism, and the two get merged constantly.

Independent measurement supports the supply side. Ben Fisher's study found Yelp occupying the structured business-card slot in 95.83 percent of runs.

The Foursquare claim, traced from origin to refutation

This is the most repeated statement about ChatGPT and local business data, and it is worth walking through because the failure pattern repeats everywhere else.

Origin. In May 2025 Natzir Turrado ran 50 prompts across three verticals in five Spanish cities and inspected ChatGPT's internal business data structure rather than the rendered answer. That work produced a figure in the range of 60 to 70 percent for Foursquare, from a small sample, in first position only, in dense European cities.

Transmission. The qualifiers fell off. Fisher traced eight downstream repetitions and found seven had dropped every caveat. One version added an invented qualifier about smaller towns, which is the reverse of what the original density data showed.

Refutation. In August 2026 Fisher ran 2,879 successful ChatGPT queries across 12 verticals and 12 US metros on the primary surface, plus 1,728 prompts on a second surface as a cross-check, and deliberately stacked the sample toward food and nightlife in dense metros, which are the conditions most favourable to the hypothesis he was testing. Foursquare-controlled domains appeared in 0.00 percent of citations on the primary surface and 0.06 percent on the cross-check. Apple Maps and Bing measured zero. Google showed slightly higher resemblance than Foursquare, at 7.5 percent against 5.7.

The claim usually arrives attached to a recommendation to claim your Foursquare listing. That is a purchasable action resting on a refuted premise.

Local data claims about ChatGPT, sorted by evidence
ClaimStatusWhat the evidence is
Yelp supplies business details, reviews and bookingConfirmed by both partiesYelp announcement, August 2026, plus trade reporting of a non-exclusive licence
Foursquare supplies most local resultsRefuted2,879 queries measured Foursquare-controlled domains at 0.00 percent of citations
Google Business Profile feeds ChatGPT directlyNo evidenceNo documentation from either company; an indirect path exists only because profile-derived content spreads across indexed pages
Mapbox supplies business listingsCategory errorThe only source is a Mapbox post about map rendering. Rendering a map and sourcing listings are different jobs
Bing Places is a major sourceContradictedMeasured at zero in the largest available test
Something ranks the businesses once retrievedUnknownThe one study that inspected internals found ordering by provider response order, and it predates the Yelp era by more than a year

The ordering question, which is genuinely open

Turrado's most interesting finding is the one nobody repeated. Inspecting the internal structure, he characterised the ordering of businesses as concatenation by provider response order, essentially a race of latencies, with hard cuts when a provider returned more than twenty venues. He found no evidence of sorting by rating, alphabetically, or by opening hours.

That observation is from May 2025 on a small Spanish sample. It predates the Yelp arrangement by more than a year. Whether any ranking step exists inside the current business card surface is an open question with a literature of zero.

So when a guide tells a local business how to rank in ChatGPT, ask which measurement of ChatGPT ordering it is describing. As of today there is one, it is small, it is old, and it found no ranking at all.

What a business website contributes

BrightLocal ran 800 manual local searches across 20 verticals and 20 US cities and recorded the first ten sources for each. Business websites accounted for 58 percent of sources, business mentions 27 percent, and directories 15 percent. Within the directory slice, smaller review and roundup sites dominated. Yelp, Facebook and Google Maps did not appear as directory sources in that sample at all.

That was November 2024, before the Yelp arrangement, and the sample is dense US cities. Treat it as a dated snapshot of a moving surface rather than a current composition.

The durable point survives the date. The largest single category of cited source was ordinary business websites, which is the one asset a local owner fully controls.

One correction while we are here. Claude has no local business data layer at all. Its web search tool documentation exposes an approximate location object with city, region, country and timezone, and no points-of-interest interface. Guides that bundle ChatGPT, Claude and Gemini as though all three had a local surface are describing something that does not exist on one of them.

What is actually established, and how

Sorted by how strong the evidence is, not by how convenient it is.

ClaimBasis
OpenAI discloses only that ChatGPT may share location with trusted third-party providers for nearby business listings, naming none of them.Documented by the platform
Yelp reviews, ratings, photos, business details, reservations and waitlist are integrated into ChatGPT, announced by Yelp on 10 August 2026.Documented by the platform
The Yelp licensing arrangement is non-exclusive according to trade reporting from July 2026.Documented by the platform
Foursquare-controlled domains appeared in 0.00 percent of citations across 2,879 ChatGPT local queries in 12 US metros.We measured this
Yelp occupied the structured business-card slot in 95.83 percent of runs in that same study.We measured this
Business websites were 58 percent of first-ten sources across 800 manual local searches in November 2024.We measured this
Claude exposes an approximate location parameter and no points-of-interest or business listing interface.Documented by the platform
How businesses are ordered inside the current ChatGPT local surface.Not publicly documented

What nobody can currently tell you

Stated because the alternative is implying a certainty that does not exist.

  • Nothing is publicly known about whether an ordering step exists in the Yelp-powered business card era. The only internals-level observation predates it by more than a year and found ordering by provider response latency rather than by any quality signal.

  • OpenAI has never named a local data provider in its own documentation, so the full provider list is unknown and any published list is inference.

  • Whether a business appearing in a ChatGPT answer produces any contact, quote request or sale has never been measured by anyone.

  • Coverage outside dense metropolitan areas is largely unmeasured. The refutation study deliberately concentrated on dense metros, and the one study that looked at density found coverage thinning outside it.

  • Whether the composition of cited sources measured in late 2024 still holds after the Yelp arrangement changed what sits in the business card slot.

What people get wrong about this

  • Foursquare supplies 60 to 70 percent of ChatGPT local business data, so claim your Foursquare listing.

    What is actually the case

    The origin was 50 prompts in five Spanish cities in first position only. A study at roughly fifty times the sample size, deliberately biased toward conditions favouring that hypothesis, measured Foursquare-controlled domains at zero.

  • Your Google Business Profile is a data source for ChatGPT.

    What is actually the case

    Neither company documents any integration, and where it has been examined no direct link was found. Profile-derived facts do propagate onto indexed pages that a model can read, so an indirect path exists. That is a different sentence from the one being sold.

  • Mapbox listings matter because OpenAI uses Mapbox.

    What is actually the case

    The only evidence offered is a Mapbox announcement about maps rendering inside ChatGPT. Rendering a map is not sourcing a business listing, and the conflation is usually followed by an offer to manage listings across dozens of platforms.

  • The same local playbook works across ChatGPT, Claude and Gemini.

    What is actually the case

    Claude has no documented local business data layer whatsoever. Any advice about ranking locally in Claude is describing generic web search results.

How to check this on your own site

You should not have to take our word for any of it.

  1. Ask ChatGPT for businesses in your category and town ten times, logged out, and record every name and every citation. Then repeat with a nearby town of very different density and compare how many businesses each returns.
  2. For any answer that shows a business card, note which platform supplied it. That single observation tells you more about current sourcing than most published guides.
  3. Search your own business name and check which pages get cited. If your own site is not among them, that is a more actionable finding than any directory theory.

Questions we get asked constantly

Which directory should a local business prioritise for ChatGPT?

The evidence supports Yelp for restaurants specifically, because that partnership is confirmed by both parties. Beyond that, directory advice in this space is largely untested, and the most-cited source category in the one composition study was ordinary business websites.

Does ChatGPT always search the web for local questions?

No. OpenAI documents that ChatGPT decides when to search. What it does for a local query where no search is triggered is not documented, and almost the entire published literature assumes retrieval happened.

Why do published numbers about this disagree so much?

Different surfaces, different countries, different densities, different dates, and samples ranging from 50 prompts to several thousand. A figure without its sample and date attached cannot be compared to another one.

Where this came from

Every factual claim above traces to one of these. Each entry says what it supports and the date it was read, because platform documentation changes without notice.

Want this measured on your own site rather than explained?

AI Search Optimization