We asked ChatGPT where to eat in London 32 times. It never gave the same answer twice.
Updated: Jul 26
AI restaurant recommendations are now shaping where Londoners eat, and almost nobody has checked what the machines actually say. So in July 2026 we ran the test ourselves: 32 queries through ChatGPT with live web search enabled, covering twelve London neighbourhoods and eight occasions, and we logged every restaurant named and every source cited.
Two findings surprised us. The first is who AI listens to, and it is not who the industry assumes.
The second is that asking the same question twice gets you a different answer — which has uncomfortable implications for anyone selling guaranteed AI rankings, including us.
Here is everything, including the parts that are inconvenient.
How did we run the study?
We sent 32 prompts to ChatGPT with web search enabled between 24 and 25 July 2026, using the GPT-5.4 model via the DataForSEO API. Twenty were unique questions covering London areas and occasions. Three questions were then repeated five times each to test whether answers held steady.
The queries covered Shoreditch, Soho, Covent Garden, Mayfair, Clapham, Hackney, Islington, Peckham, Brixton, Camden, Notting Hill and Borough Market, plus occasion-led questions: date night, birthday dinner, Sunday roast, cheap eats, vegan, Italian, Indian and taking a client to dinner.
We recorded every venue named and every URL cited. The full response set and citation log are published below — no cherry-picking, and you can check our working.
What this study is not. It is one model, on two days, from one location. It is not a census of AI behaviour, and we have not tested Gemini, Perplexity or Claude. Treat it as the first UK data point in a space that had none, not the final word.
What does AI cite when you ask where to eat in London?
Time Out and the Michelin Guide account for 56% of every citation. Across 32 responses ChatGPT made 280 citations to 37 different domains. Time Out was cited 86 times and appeared in 23 of 32 answers. The Michelin Guide was cited 71 times across 22 answers. Nothing else came close.
Source | Citations | Share | Appeared in |
Time Out | 86 | 30.7% | 23 of 32 answers |
Michelin Guide | 71 | 25.4% | 22 of 32 |
OpenTable UK | 27 | 9.6% | 8 of 32 |
The Infatuation | 15 | 5.4% | 10 of 32 |
Visit London | 14 | 5.0% | 5 of 32 |
Condé Nast Traveller | 10 | 3.6% | 5 of 32 |
That's Up | 9 | 3.2% | 6 of 32 |
Eater London | 5 | 1.8% | 4 of 32 |
SquareMeal | 2 | 0.7% | 2 of 32 |
Tripadvisor | 1 | 0.4% | 1 of 32 |
Harden's | 1 | 0.4% | 1 of 32 |
Now the numbers that should genuinely change what you do on a Monday morning:
Tripadvisor was cited once in 32 answers. The platform restaurants lose sleep over barely registered.
Instagram was cited zero times. So were Facebook and TikTok. Social media accounted for 0% of 280 citations.
TheFork and ResDiary were cited zero times — though OpenTable UK was cited 27 times, which is worth understanding if you are choosing a booking platform.
SquareMeal and Harden's, two of the best-known names in British restaurant coverage, managed three citations between them.
The overall split: 90% of citations went to third-party guides and directories. Just 10% went to the restaurant's own website.
Does ChatGPT give the same restaurant recommendations twice?
No. In three questions asked five times each, not one restaurant appeared in all five answers. Not one, across 45 distinct venues.
We asked "Where should I eat in Shoreditch tonight?" five separate times. We got 14 different restaurants. The most consistent — Brat, Plates and The Clove Club — each appeared in four runs out of five. Nothing appeared in all five.
Question, asked 5 times | Distinct restaurants | In all 5 answers | Named only once | Average overlap between two runs |
Where should I eat in Shoreditch tonight? | 14 | 0 | 7 | 32% |
Where to eat in Peckham | 19 | 0 | 14 | 14% |
Best Sunday roast in London | 12 | 0 | 8 | 28% |
Across all three questions, 64% of the restaurants named appeared in only one of the five runs. Average overlap between any two runs of the same question was 25%.
Read that again if you are being sold guaranteed AI rankings. Three quarters of the answer changes between one asking and the next.
So should I just try to get into Time Out?
By all means try — but "get into Time Out" is an outcome, not a plan. It is roughly as useful as being told to fill more covers. Nobody sells a place in the Michelin Guide, Time Out does not take applications, and every restaurant in London would already be in both if wanting it were enough.
The more useful question is why those guides wrote about the restaurants they wrote about. And when you look at the data rather than the fantasy, three things become obvious.
First, the filter comes before the citation. Google's own example of an AI restaurant query is "find a table for two at a dog-friendly Italian restaurant in Shoreditch for Saturday at 7 p.m." Look at how much filtering happens there — location, cuisine, party size, dog policy, day, time. If your opening hours are wrong, your cuisine is tagged vaguely, or nobody has ever recorded whether you allow dogs, you are removed from consideration before any guide citation gets a chance to help you. Being in Time Out does not save you from being filtered out at the first step.
That layer is entirely yours. It is also, in our experience, wrong or incomplete at most independent restaurants.
Second, it is breadth that holds, not one placement. Brat appeared in four of five Shoreditch runs. Not because of one write-up — because Michelin, Time Out and The Infatuation all cover it, and ChatGPT read all three. One guide is a lottery ticket. Being visible across several is the actual mechanism.
Third, one mention does not hold the position. With 25% overlap between one asking and the next, a single write-up in 2024 is not doing the work you think it is. Presence has to be maintained, which is why this is a monthly job rather than a project with an end date.
Where does a restaurant actually start?
Not at Michelin. You start by becoming the kind of restaurant that is easy to find, easy to verify and easy to write about — because that is what every one of those guides is looking for, and because the 10% you control is the only part of this you can act on directly this week.
That 10% is not nothing. It was 27 citations in our study, and it is the only share of the answer any restaurant owns outright.
Here is the honest order of work, and it is the order we work in.
Fix what the machines read about you. Hours, cuisine, location, menu in a format that can be parsed, and the attributes that filtered queries depend on. Unglamorous, and it decides whether you survive step one.
Make yourself consistent everywhere. If your opening hours disagree across three listings, you are ambiguous — and ambiguity is what gets a restaurant dropped from a shortlist of four.
Build a body of your own material worth citing. Our study found guides dominate partly because most restaurants publish nothing an AI can quote. A restaurant with a real menu page, a proper story and answers to the questions diners actually ask is a restaurant that has something to be cited for.
Get talked about, consistently. Guides and journalists do not discover restaurants by magic. They find the ones that are already visible, already being mentioned, already easy to check.
Watch what changes, and keep going. Because the answer moves every time it is asked.
None of that is glamorous and none of it is instant. But it is the work that turns a restaurant from invisible into the kind of place that ends up in the guides — rather than sitting around hoping to be discovered by them.
What about social media, if it gets cited zero times?
Our own data says social media is not what AI reads. Zero citations out of 280 — not Instagram, not Facebook, not TikTok. We are not going to tell you otherwise three paragraphs after publishing that number.
What social does is different, and worth being precise about, because plenty of people will happily sell you the vaguer version.
It brings diners directly and builds authority. It is where your reviews, your mentions and your word of mouth start. And it is what a Time Out writer looks at when deciding whether you are worth a visit — nobody covers a restaurant with a dead feed and no evidence anyone eats there.
So: social is demand, and it is the raw material that eventually becomes the coverage AI reads. It is not, on our evidence, a direct route into an AI answer but it helps. Anyone telling you your Instagram posts make ChatGPT recommend you directly is guessing, and our data suggests they are guessing wrong.
Why is this the first UK data on the subject?
Because the whole market has been quoting American numbers. The most-cited figure in UK hospitality content — that 45% of consumers use AI for local business recommendations — comes from BrightLocal's Local Consumer Review Survey, published 11 February 2026, and is a survey of 1,002 US adults about local businesses generally, not restaurants and not Britain.
Popmenu's restaurant-specific survey of 1,000 US consumers, fielded 4–5 February 2026, found 20% using AI tools to find restaurants. Same year, same country, restaurant-specific — less than half the figure being quoted at British operators.
We looked hard for a UK dataset before running our own. There isn't one.
Questions restaurants ask us
Which AI models did you test?
One: ChatGPT, on the GPT-5.4 model with live web search, on 24 and 25 July 2026. We have not tested Gemini, Perplexity or Claude, and we would not assume the results transfer. Yext's citation research found only around 5% overlap in citations between ChatGPT, Perplexity and Gemini, which suggests each engine is close to a separate problem.
If the guides get 90% of citations, why not just do PR?
Try, if you have the budget and the contacts. But guides cover restaurants that are already findable, verifiable and being talked about — and a write-up you do land still has to survive the filtering stage, which depends on data you control, not the guide. The two are not alternatives. One is the ground floor.
Does this mean AI recommendations are random?
Not random, but not fixed. Some restaurants appeared in four of five runs and others in one. The pattern that predicted repeat appearances was coverage across several trusted guides. Think of it as a weighted draw rather than a ranking.
I'm not in any guide at all. Is this hopeless?
No, but it is a starting position rather than a finishing one. Get the machine-readable layer right, build something worth citing, and become visible enough to be noticed. That is a matter of months, not weeks, and anyone promising faster is selling you something.
How often will you repeat this?
Quarterly, with the method and the raw data published each time. The value of a benchmark is in the trend.
Can I see the raw data?
Yes. All 32 responses, every citation and the consistency calculations are yours for the asking — email hello@gptable.ai and we will send the lot. If you spot an error in our working, tell us and we will correct it publicly.
What we would do first, if it were our restaurant
Find out where you actually stand. Not a guess, not a scare statistic borrowed from an American survey — the specific answer to what AI says when someone asks for a restaurant like yours, in your part of London, tonight.
We will run your restaurant through the same process we used for this study and show you exactly what comes back: which questions surface you, which competitors appear instead, which of the sources above already mention you, and which of the details AI reads about you are wrong.
Thirty minutes. You keep the findings whether you work with us or not.
Comments