Does ChatGPT use Yelp or Google ratings? We checked 170 businesses on its map

In the US and Canada, the star rating on a ChatGPT map card usually matches Yelp; in the UK and Australia it usually matches Google. We asked ChatGPT 30 local questions twice each on 27 September 2026, in Houston, Miami and London, and took every business shown on the map cards in its answers: 170 in all. Then we looked each one up on Google Maps, Yelp, Apple Maps, Bing Maps and, for restaurants, Tripadvisor, and compared the rating on the card with each platform.

Many businesses have the same rating everywhere, so a match on its own proves little. The telling cases are those where Yelp and Google disagree. In Houston and Miami there were 68 such businesses: the card matched Yelp for 55 of them, Google for 8 and neither for 5. The next day a second run gave the same result in Chicago (18 of 20) and Toronto (11 of 11). In London the pattern reversed: 9 of 12 cards matched Google.

This is consistent with the one public deal. Yelp and OpenAI announced on 23 July 2026 that ChatGPT would get Yelp's reviews, ratings, photos and business details; the financial terms were not disclosed and no countries were named. Our check shows the result on the card, not the contract behind it, and it compared star ratings only: it does not show how ChatGPT chooses which businesses to list, or whether it reads the text of Yelp reviews.

Because in the US and Canada the card usually shows your Yelp rating, and that is often lower. Of the 116 businesses we could compare in Houston and Miami on 27 September, 59 had a lower rating on ChatGPT's card than on Google, 17 of them by half a star or more; only 6 were rated higher. The widest gap was 2.3 stars. On 28 September, 17 of 24 Chicago cards were below Google, 3 of them by half a star or more; in Toronto 14 of 28 were below, and a larger share, 10, by half a star or more. In London, 9 of 53 were lower. A business that has built up its Google reviews for years can appear on ChatGPT's map with a score drawn from a different, often much smaller, set of reviews.

The examples below come from our screenshots. In the first, the card took Yelp's 3.0 from four reviews rather than Google's 5.0 from more than 150. The London row shows the reverse: the card followed Google. A reader of the card cannot tell these cases apart, because only 12 of the 170 cards in the first run showed a review count at all, all of them car repair shops.

In Toronto, yes; in the UK and Australia the cards matched Google instead. In Toronto, all 11 businesses whose Yelp and Google ratings differed had the Yelp rating on the card, and none had Google's. Yelp's Canadian site blocked our browser with a CAPTCHA, so we took Yelp's rating from Apple Maps and Bing Maps, which show it under Yelp's name; the method section explains how we handled Bing's rounding.

In Manchester the card showed the Google rating for all 22 service businesses, and in Sydney for all 23 we found on Google, including figures from 4.5 to 4.8 that coincide less often by chance. Yelp blocked us with a CAPTCHA in both cities, so this is a match with Google, not a test against Yelp. In Chicago, for contrast, the card equalled Google for only 5 of 19 service businesses. London, from the first run on 27 September, points the same way: where Yelp and Google differed, the card matched Google in 9 of 12 cases and Yelp in 1. Only 18 of the 54 London businesses had a Yelp rating at all, against 82 of 116 in Houston and Miami.

Restaurants outside North America leaned towards Tripadvisor. The card matched Tripadvisor for 5 of the 7 London Italian restaurants whose Tripadvisor and Google ratings disagreed, and for 2 of 4 in Sydney; in Berlin all six restaurant cards showed Tripadvisor's rating. We have left out Berlin's other businesses and a second set of London niches: that part of our check did not meet our evidence standard and is being repeated.

We cannot say whether this reflects agreements or simply where Yelp has data. Yelp's own announcement of table reservations inside ChatGPT, on 10 August 2026, covers restaurants in the US and Canada; the July ratings deal names no countries. Beyond Houston and Miami our figures rest on a few dozen businesses per city, so treat them as an early observation. For the wider picture of the American market, see our guide to GEO in the USA.

We matched the rating on each ChatGPT card, within 0.05, against five platforms on the same day, and drew conclusions only from businesses whose Yelp and Google ratings differed. We did not ask ChatGPT where its data came from. In our pilot runs on 22 September 2026, the model said it had not searched the web in answers that showed web citations, so its account of its own sources is not evidence.

Counting any match would have misled us. 93 of the 170 cards showed 5.0 or 4.9, and such ratings often coincide on several platforms at once; counted that way, Google looks like the leading source. Apple Maps and Bing Maps are not independent sources either. 69 of the 82 Apple ratings we recorded were labelled as coming from Yelp, and of 108 Bing ratings, 28 were labelled Yelp and 7 Tripadvisor.

As a control we asked Google AI Mode the same questions and checked 20 businesses from its cards against Google Maps. All 16 cards that showed a numeric rating matched, so the method does pick out a known source.

In the second run, Yelp's sites in Canada, the UK and Australia blocked our browser with a CAPTCHA. Where Apple Maps or Bing Maps showed a rating attributed to Yelp, we used that figure; because Bing rounds to half a star, we counted the two ratings taken from Bing as consistent rather than exact matches. We re-read 156 platform screenshots against the recorded numbers and dropped every value that had no screenshot of its own.

Yes: "Can you recommend…" questions got a ChatGPT map in 48 of 48 answers, "How do I choose…" questions in 7 of 12. We asked the second kind for dentists and plumbers in all three cities. Those answers named 2.8 businesses on average instead of 5.4, and two named none. Google AI Mode showed no map at all for the "How do I choose" questions, 0 of 12, though it still named businesses in the text.

When the map is there, it is the shortlist. In answers with a map, 92% of the businesses ChatGPT named in the text (280 of 305) also appeared on its cards, the pattern we first saw with law firms in our study of which lawyers AI assistants recommend. A business that is not on the map is rarely in the answer; our guide on why ChatGPT doesn't recommend you covers the usual reasons.

They do not show where the rating comes from. The Directions button on the first card opened Google Maps in 52 of the 55 answers with a map; that is a route, not the source of the stars. Search Engine Land reported in July that Yelp branding and links would appear where OpenAI uses Yelp content. In our screenshots a Yelp label or link was visible in the opened card in 9 of the 55 map answers, all nine in the US; Tripadvisor appeared in 5, and OpenTable, the Better Business Bureau and WeddingWire occasionally.

So the card rarely tells a reader, or a business owner, where the stars came from. Comparing the number on the card with each platform on the same day is the only reliable test we found.

In the US and Canada, check your Yelp profile first; in the UK and Australia, your Google Business Profile, plus Tripadvisor if you run a restaurant. On Yelp, claim the listing and check the name, category, address and hours. Do not ask customers for Yelp reviews: Yelp tells businesses not to request them, and its software may not recommend reviews that look prompted by the business. For Google and Tripadvisor the same rule applies: an accurate, complete listing and reviews earned without prompting.

Then look at your own card. Open ChatGPT in a Temporary Chat (a mode that does not use or save your chat history), ask two or three questions the way customers do, such as "Can you recommend a plumber in [your city]?", and screenshot the map. Compare the rating on your card with Google and Yelp on the same day, and repeat monthly; our guide on checking your AI visibility sets out a fuller protocol. If you would rather have it checked for you, our free audit looks at your card, which platform its rating matches, and how the map changes with the wording of the question.

On 27 September 2026, between 12:01 and 15:53 UTC, we asked ChatGPT and Google AI Mode the same 30 questions, twice each. ChatGPT ran on a free account in Temporary Chat with memory off, in a desktop browser; AI Mode ran in a private window without a login. Twenty-four questions asked for recommendations in eight niches (dentist, plumber, small-business accountant, car mechanic, vet, wedding photographer, engagement-ring jeweller, Italian restaurant) across Houston, Miami and London; six asked how to choose a dentist or a plumber. The city was named in each question, and the browser's location was not changed. That gave 120 answers, all recorded as screenshots. On the same day we checked 170 businesses from ChatGPT's cards and 20 from AI Mode's on the five platforms.

On 28 September 2026, between 17:12 and 18:04 UTC, we asked ChatGPT 30 more recommendation questions, once each and with the same setup: dentist, plumber, car mechanic, vet and Italian restaurant in Chicago, Toronto, Sydney, Manchester and Berlin, and electrician, physiotherapist, hair salon, locksmith and sushi restaurant in London. All 30 answers showed a map, with 166 businesses on the cards, which we checked on the same five platforms that evening and verified against the screenshots the next day.

We compared ratings only; this study does not show where ChatGPT gets its list of businesses or how it ranks them. The first run covers one day, three cities, eight niches and English. The second asked each question once, and in every city except Houston and Miami the figures rest on dozens of businesses, not hundreds. OpenAI's help page describes maps on iOS and Android; we recorded the desktop web version, which showed maps as well. Interfaces change, so treat these figures as a dated measurement. The full table of 190 businesses, with the rating on every platform, is published as a CSV file, and the 166 from the second run as a separate CSV file, where values we could not verify, including the Berlin and London checks we set aside, are left blank; our note on how we measure describes our general method.