London group dining: the names AI gives and the rooms behind them

TLDR: Across 1,572 usable answers to 200 London group-dining questions, Hawksmoor was named in 12.6% and Dishoom in 12.0%. For a restaurant team, a name was only the start. In ten room and booking cases we checked, the exact room, branch and terms decided whether a suggestion worked, and three answers disagreed with the restaurant's own page.

“Dinner in Soho, London, for 14 friends. Where won't feel cramped?” Ask an AI assistant that and a shortlist arrives in seconds. The host still has to find out whether any of those places can seat 14 together, and on what terms.

On 20 October 2026 we put that question and 199 others about group meals in London to ChatGPT, Google Gemini and Google AI Overviews, three times each. We counted which restaurants were named, watched how the names moved with the question, and checked ten room and booking details against the restaurants' own pages.

A table for 14 in Soho

The Soho question has a headcount, a neighbourhood and a worry about space. A useful answer deals with all three.

One ChatGPT answer did. It put Quo Vadis first and named two rooms: the Blue Room, seating 14 at one long table, and the larger Marx Room, for up to 24 at one oval table. The Quo Vadis private dining page bears out both and adds a second Marx Room layout, 32 across four tables. Non-members can book the event spaces too.

That answer gave the host a room name, a number and a table shape to check. A bare restaurant name would have left all three to find.

Across the study, a handful of names came up again and again. The harder test was whether the detail behind each name matched what the restaurant offers.

Which names came up most often

The study made 1,800 attempts: 200 questions, three services, three runs each. That produced 1,572 usable answers: 598 from ChatGPT, 586 from Google Gemini and 388 from Google AI Overviews. Every overall rate below uses those 1,572.

Captured answers count every name in the answer, including venue cards, the short boxed listings some services show. Written answers count names the assistant wrote about in its sentences or headings, or described in substance under a card heading; isolated card labels drop out. On this stricter count, closer to what the assistant said about a place, Hawksmoor (191) and Dishoom (184) still lead, and Big Mamma edges Brasserie Zédel, 95 to 94. The rest of the order holds, with HIDE next on 63.

Each row is a declared dining unit, and the units differ in size. Hawksmoor combines its named branches; its Guildhall branch alone has a 24-seat private dining room. Big Mamma counts the family name plus Circolo Popolare, Ave Mario, Gloria, Jacuzzi and Carlotta, once per answer. Gordon Ramsay Restaurants counts the family name and its declared portfolio, including the currently listed Pizza East. The Ivy counts its declared family, including Ivy Collection, Ivy Asia and Granary Square Brasserie. The Wolseley is narrower: Piccadilly and City count, while the Wolseley Hospitality Group alone doesn't. Brasserie Zédel and The Delaunay keep separate rows.

A group with many restaurants has more ways into an answer than one room at one address, so the table compares declared units, not estate size or individual rooms. We checked the ten against current official London dining pages, but not the ownership or size of every name.

Below the ten, HIDE appeared in 64 captured answers (4.1%), The Coal Shed in 58 (3.7%), Caravan in 57 (3.6%) and Brother Marcus in 56 (3.6%). Two Coal Shed answers name its Brighton restaurant, so that count doesn't establish a London branch.

The study started from 19 restaurant brands, whose median was named in 33 answers (2.1%), a figure about those brands rather than a typical London restaurant. Names found in the answers widened the list to 343 dining brands or families named at least once, including hotel restaurants and food halls. Beside it sit 16 adjacent, unresolved or separately excluded names, such as caterers, event venues and sub-brands counted elsewhere, two of them with no named answer. All are in the measurement record. A missing name doesn't show that a restaurant is absent from AI search generally.

The question changes the shortlist

The 200 questions fall into five groups of 40, each with more than 300 usable answers. Shares are captured answers within each group or service; the measurement record has every name broken down both ways.

The leaders swap with the question. Hawksmoor was named in 52 of 323 answers about party size, privacy and room layout (16.1%); Dishoom in 20 (6.2%). On choosing a restaurant and neighbourhood, Dishoom had 50 of 304 (16.4%) and Hawksmoor 40 (13.2%). The Ivy reached 11.8% on neighbourhood questions and 0.6% on food, diet and drinks. Quaglino's peaked on room layout (9.3%); Gordon Ramsay Restaurants (11.7%) and The Wolseley (10.2%) on occasions.

So a team with a private room for 20 should start with the party-size figures, and a team known for its menu with food, diet and drinks.

The service matters too. Each produced a different number of usable answers, so compare shares within one service.

Dishoom appeared in 58 ChatGPT answers (9.7%) and 98 Gemini answers (16.7%). Hawksmoor had 48 in ChatGPT (8.0%), 94 in Gemini (16.0%) and 56 in AI Overviews (14.4%), where Brasserie Zédel had 43 (11.1%). Gaucho ran the other way, strongest in ChatGPT (6.7%). These shares describe captured answers, not how many people use each service. A strong showing in one service told us little about the other two, so check all three.

Room, branch and terms decide the booking

Another ChatGPT answer to the Soho question named Brother Marcus Soho, with a private dining room for 14 and a feasting table for up to 18. It suggested asking about the table first if space was the main worry. The Brother Marcus Soho groups page lists both.

Those are two different bookings, a room of the group's own or one big shared table, and the host has to choose before the capacity means anything.

The branch can decide it too. AI Overviews answered a question about “dinner with family the night before our London wedding” with The Wolseley's private room for up to 15 at one oval table, reached by its own staircase. The Piccadilly private dining page confirms all three details. The City restaurant has different rooms and layouts, so the name alone points to two restaurants with different answers.

Then come the terms. An AI Overviews answer for a 50th birthday said parties of 8 to 24 at Circolo Popolare must book a group sharing menu. The Circolo Popolare page confirms the set menu for groups of 8 to 24 and opens ordinary bookings 30 days ahead at 9am. Gemini suggested Dishoom, Carnaby included, for a relaxed team dinner; Dishoom Carnaby takes online bookings for up to 20 guests up to four months ahead, with enquiries for larger parties. Neither page shows whether a table is free on a given night.

For “Private dinner in London for about £60 each before drinks. Is that realistic anywhere?”, an AI Overviews answer named Khari Baoli at Cinnamon Bazaar in Covent Garden. The Cinnamon Collection events page describes it as a room for up to 30 with its own bar, and lists a separate 18-seat mezzanine at Cinnamon Kitchen Battersea. We didn't verify the prices in the answer. The room checked out.

Where answers and official pages disagreed

Many details matched. A ChatGPT answer to “London private dining rooms: where should I start looking?” said Hawksmoor Guildhall's private room seats around 24, which the Guildhall page confirms. Three answers did not match.

For “Where in central London can 30 of us eat together?”, a ChatGPT answer described Brasserie Zédel's “large dining spaces and private rooms”. The reservations page takes ordinary bookings of up to 16 for lunch and 10 for dinner. The FAQ limits large group bookings to weekday lunchtimes and says there is no private dining room; Crazy Coqs and Bar Américain can be hired. Neither page offers a seated dinner for 30. A host would have to ask the events team whether either hired space can take one.

The £60 answer said Bancone Covent Garden's private room seats up to 12. The current Bancone page says 8 to 14, and adds that Kensington has no private room and takes groups of 8 to 28 in the main dining room. We didn't verify the answer's menu price, so the capacity, branch, prices and service charge all need checking with Bancone.

For “I'm inviting journalists to lunch in London. Where can we have a quiet room and good food?”, an AI Overviews answer described a downstairs “Butcher's Locker” at The Quality Chop House. Its private room page describes a first-floor room for 8 to 14. For a guest who cares about stairs or privacy, the floor and the room's identity are booking facts.

We chose these ten cases to show how room and booking details behave in real answers; they give no accuracy rate for any service or restaurant. In all three mismatches the error was in the room description, the detail a host books on.

What the sources show

The services often list source pages alongside their answers. These eight hosts appeared most.

One answer can list several hosts, so the shares overlap. SquareMeal appeared across all three services. OpenTable's UK site appeared in 220 ChatGPT answers, 36.8% of ChatGPT's usable set. Source pages were recorded for 1,262 usable answers; 310 had none.

A page in a source list differs from a citation beside a sentence. We haven't treated a listed page as proof that it was read, trusted, quoted or responsible for any restaurant being named.

For an operator, the table is a reading list. Open your listings on the hosts that appear for your kind of question and check the room, group menu and booking route, starting with capacity and floor, the details the Bancone and Quality Chop House answers got wrong.

Three steps for restaurant teams

1. Name each bookable space and describe it the way a guest books it. Give the room or table a name. State seated capacity for each layout, whether the space is closed off, where it sits in the building and which branch it belongs to. Quo Vadis gives capacity by table layout and Brother Marcus separates its private room from its feasting table, so a host can match either to a headcount.
2. Put the terms next to the space. State the minimum spend, service, deposit, cancellation rule, menu choice and how far ahead bookings open. Circolo Popolare and Dishoom Carnaby publish different booking windows. A page that only says it suits groups leaves the host to ring round and gives an answer nothing specific to repeat.
3. Test one real question against your own page. Pick a question that matches a meal you serve, such as a dinner for 14 in your neighbourhood. Read the answers from all three services and open the sources they list. Compare branch, room, capacity and terms with your official page and directory listings, correct whatever disagrees, then ask the same question in the same words later.

If you'd like a second view on how your rooms read in these answers, you can contact Schmitdy.

How the study worked

We wrote the 200 questions to cover recognisable guest decisions in Greater London. They don't come from search logs and don't measure query volume or diner demand. Some are broad: a first-time visitor question also drew sightseeing plans, which we kept as answers to the literal question even without a restaurant name. The original English questions appear in full in the appendix.

Each of the 1,800 attempts on 20 October 2026 was joined to its exact question. We excluded 210 empty bodies of unknown cause, six refusals, four fragments, seven clarification-only replies and one header-only reply, leaving 1,572 usable answers. An empty AI Overview was never counted as a zero for any restaurant.

A name counts once per usable answer, however often it appears or however many branches are named. A link destination alone doesn't count. We spot-checked two card formats in the captured view, where names and source lists appeared separately, but not every card, the original Google interface, or opening-status and price details.

A separate recall check reviewed 95 candidate matches and accepted six missed names. Fortnum & Mason, Duck & Waffle and Smith's Bar & Grill were hidden by how an ampersand or space was encoded. Two were Coal Shed cards naming Brighton. Harper Privé stays with the catering alternatives and isn't counted in written answers.

Comparisons by question type or service need at least 30 usable answers and ten distinct questions. Three runs of one question are not three independent guests. This is a dated snapshot with no revenue, conversion or causal evidence. We coded no sentiment or recommendation strength and assigned no overall accuracy score. The name list is incomplete, and current operating checks cover only the top ten and the ten cases.

Sources

The measurement record holds all 200 questions, the shared denominator, every name by service and question type, and source definitions. Every restaurant page linked above was checked on 20 October 2026, as were these pages for the top ten:

• Hawksmoor: private dining
• Big Mamma: London restaurants
• Gordon Ramsay Restaurants: portfolio, private dining, kitchen tables
• The Ivy: London locations, Ivy Asia, Ivy Asia private dining
• Bob Bob Ricard: Soho, City (no City capacity certified, as the page conflicts)
• Quaglino's: private dining
• Gaucho: private dining

Questions restaurant teams ask

Does being named mean the restaurant was recommended?+

Not necessarily. It means the name appeared under the counting rules above. A name can appear as an example, a comparison or a qualified suggestion, and we didn't code sentiment or recommendation strength.

Can a single-site restaurant compare itself with a large restaurant group?+

Yes, question by question rather than on the overall table. The overall figures combine declared families: Hawksmoor's count covers its named branches, and Big Mamma's covers its family name and five named London restaurants. Ownership and small-business status weren't used to build a league table.

Can a guest book straight from an AI answer?+

Only after confirming the branch, room, date, access, menu and terms with the restaurant. Brasserie Zédel, for example, takes ordinary bookings of up to 10 at dinner and has no private dining room, whatever an answer says. The study records answers and checks selected facts; it doesn't confirm availability or a booking.

What should a restaurant change before the next measurement?+

Fix verified room, menu and booking details first. Keep this question set so a rerun stays comparable, and record any deliberate change to the questions separately, because a new question set starts a new baseline.