Byter research · 24 October 2026
We asked ChatGPT 60 questionsLondon diners ask, from “best Indian in Soho” to “Sunday roast in Clapham”, and studied 115 answers and the 195 restaurant cards it showed. The short version: the ratings came from Tripadvisor, not Google, the lists were shaped by restaurant guides, and you do not need thousands of reviews to be named.
86%
ChatGPT showed a star rating and review count next to 195 recommended restaurants. Its own place data labels 168 of those as Tripadvisor. 20 came from another data partner, 1 from OpenTable and 6 had no source. None were labelled as Google. Its "directions" links did go to Google Maps, in 63 of 70 answers.
430
Half of the restaurants ChatGPT recommended had fewer than 430 reviews, and 42 of 195 (more than one in five) had fewer than 100. The restaurant it named first had a median of 396 reviews, slightly fewer than the rest. Ratings averaged 4.33, and 26 restaurants rated below 4.0 still made the list.
86%
60 of 70 answers cited at least one editorial guide. Time Out was cited in 49% of answers, The Infatuation in 46% and the Michelin Guide in 44%. Harden's, SquareMeal, DesignMyNight and the Evening Standard followed. ChatGPT read an average of 12.2 search results per answer, from 155 different websites.
64%
OpenTable was the website ChatGPT searched most often: its pages came up in 45 of 70 answers. OpenTable was also the only booking partner ChatGPT attached to a restaurant, on 72 of 195 recommendations (37%). We saw no SevenRooms, Resy or other booking links.
84%
ChatGPT linked to the restaurant's own website for 163 of the 195 restaurants it recommended. A handful had only an Instagram or Facebook page in place of a website.
225
208 of them (92%) were named only once. The most-named restaurant appeared 3 times. ChatGPT does not keep recommending the same famous names: each area, cuisine and occasion has its own short list, and a restaurant can win a narrow question without being famous.
33 of 40
We asked the same questions again later the same day. ChatGPT picked the same restaurant first for 33 of 40 questions, and gave an identical list for 17. On average 77% of the restaurants overlapped. The answers are fairly stable, but a place on the list is not guaranteed every time.
Six steps, in the order we would take them, based on what ChatGPT actually read and cited.
Step 1
Time Out, The Infatuation, the Michelin Guide, Harden's and SquareMeal shape most answers. A review, a list entry or an award there is worth more to ChatGPT than another hundred reviews. That means a press list, a reason to write about you (a new chef, a menu, an opening) and invitations to the right critics.
Step 2
It is where ChatGPT's ratings came from. Claim the listing, keep the hours, menu, photos and price band right, reply to reviews, and ask every guest for an honest review. Never offer anything in return: incentivised reviews break Tripadvisor's rules and UK consumer law.
Step 3
OpenTable pages turned up in nearly two thirds of searches, and it was the only booking button ChatGPT showed. If you use another system, make sure an OpenTable profile still describes you accurately.
Step 4
ChatGPT linked to most recommended restaurants' own sites. Put your menu on the page as text (not a PDF or a photo), name your area and nearest station, and give occasions their own page: private dining, pre-theatre, Sunday roast, halal, vegan, groups.
Step 5
Almost every restaurant we saw was named for one specific question. "Best Lebanese in Marylebone" or "birthday dinner for 10 in Soho" is winnable. Pick the two or three questions that bring you real bookings and make sure every source says you are the answer.
Step 6
Ask it the questions your guests ask, in a temporary chat so your history does not skew the answer, and note which sources it cites. Repeat monthly. A place that was there last month can drop off.
On 24 October 2026 we asked ChatGPT 60 questionsa London diner might ask: 30 about a cuisine in an area (“What’s the best Turkish restaurant in Islington, London?”), 20 about an occasion (“Restaurant for a birthday dinner for 10 people in Soho?”) and 10 about London as a whole. 10 of them were asked twice in the first run, and the first 45 were asked again in a second run later the same day.
Every question was asked in a new temporary chat with personalisation turned off, using ChatGPT’s default fast model with web search on, from a London connection. We recorded the restaurants named, the websites ChatGPT searched and cited, and, in the second run, the rating, review count, rating source, website and booking partner attached to each restaurant card in ChatGPT’s own data.
Limits. This is one day, one city and one AI assistant. ChatGPT changes often, and answers can differ by location, account and model. The figures show what ChatGPT showed and cited, not why it chose one restaurant over another. Ratings and review counts are as ChatGPT displayed them on the day.
Download the questions and the restaurants named (TSV). Journalists can use these findings with a credit and link to this page. For the full dataset or a comment, email info@byter.com.
In our study, 86% of the ratings and review counts ChatGPT showed next to London restaurants were labelled in its own data as coming from Tripadvisor. A small number came from another data partner or OpenTable. None were labelled as Google.
Google reviews were not the source of the ratings ChatGPT showed in our study, although its directions links went to Google Maps. Google reviews still matter a great deal for Google Maps and Google's own AI answers, which we did not test here.
Fewer than most owners think. The typical restaurant ChatGPT recommended had 430 reviews, and more than one in five had fewer than 100. Being in the guides it cites mattered more than review volume.
Get covered by the guides it cites (Time Out, The Infatuation, the Michelin Guide, Harden's, SquareMeal), keep your Tripadvisor and OpenTable listings accurate and well reviewed, put your menu and area on your website as text, and aim to be the clear answer to a few specific questions.
We saw no paid or sponsored placements in the 115 answers we studied.
Mostly, but not always. Asked the same question again on the same day, it picked the same first restaurant 33 times out of 40 and gave an identical list 17 times.
Research by Lewis Banks, founder of Byter Digital, a London marketing agency for restaurants, bars and hotels. Published 24 October 2026.
