Paris Restaurants in ChatGPTWho gets named, in English and in French

The most named restaurants

The questions were written the way people ask: by cuisine (“best ramen in Paris”), by area (“where to eat in the Marais”), by occasion (“date night restaurant Paris”), by landmark (“restaurants near the Louvre”), by budget, and a few generic ones (“best restaurants in Paris”). The French set is a translation of the English one, so every question has a twin. Details are in the method.

Le Tout-Paris, the rooftop restaurant of the Cheval Blanc hotel, leads in English with 20 of 472 answers. In French the leader is La Taverne de Zhao, a Chinese restaurant serving food from Xi’an, in 21 answers. 8 restaurants are in both top 20 lists.

Answers out of 472 per language. A restaurant counts once per answer. Stars are the 2026 Michelin guide.

English and French get different restaurants

Of all the restaurants named in either language, 49% are named in both. Of each language’s top 50, 25 are shared. When I compare the two answers to the same question (“best Korean restaurants in Paris” against “meilleurs restaurants coréens à Paris”), the median overlap is 27%.

The French answers lean more on local bistros and neighbourhood places (Bistrot Instinct, Bombance, Le Comptoir du Relais, Au Bourguignon du Marais). The English answers include more fine dining: 4 three-star restaurants are in the English top 15, against 3 in the French top 15.

Michelin

ChatGPT names 76 of the 120 matched starred restaurants in English and 81 in French. Starred restaurants account for 10% of English mentions and 8.6% of French ones. The share depends on the question.

Starred in the 2026 guide. “Budget” covers cheap and splurge questions; the splurge questions carry the starred restaurants.

On the generic “best restaurants in Paris” questions, starred restaurants take 35% of English mentions and 23% of French ones. On cuisine, area and landmark questions they stay under 9% in both languages.

Restaurant websites and booking engines

I fetched the website of every named restaurant that has one (1,626 reachable) and of a random sample of 1,000 Paris restaurants that ChatGPT never named (869 reachable). Each site was checked for an embedded booking engine, the tool used to build it, and structured data. Sites behind a bot wall (50 named, 43 never named) were recorded as such and not bypassed.

A booking engine counts when a widget or script is embedded, or a “book” link goes to the provider. A plain link to a listing page does not count.

53% of named restaurants’ websites have a booking engine, against 25% of the never-named sample. Zenchef, a French booking company, is on 34% of named sites and 14% of the others. TheFork’s widget, OpenTable and SevenRooms follow.

Bigger restaurants are more likely to be named and more likely to pay for software, so I compared the two groups within bands of review count. The gap stays between about 1.6 and 3 times in every band.

Sample sizes, named / never named: 0–100 reviews 26 / 108; 100–300 152 / 203; 300–700 400 / 240; 700–1,500 499 / 174; Over 1,500 549 / 144.

How the websites are built

WordPress is the most common builder in both groups. Zenchef’s own website builder is three times as common on named restaurants’ sites (11.9% against 4%). The restaurant site builders Uniiti and DISH are more common among restaurants ChatGPT never names, and those restaurants also link more to delivery and ordering services.

Method and limits

• Questions. 472 per language: cuisine, area (arrondissements and neighbourhoods), cuisine in an area, occasion, occasion in an area, landmark, budget, and a small set of generic and “top 10” questions. The French questions are translations of the English ones.
• Capture. English questions sent to ChatGPT from the US, French questions from France, with web search, on 4 and 5 November 2026. When ChatGPT replied with a questionnaire, the capture answered it and kept the second turn. This is the first of three runs; the second and third start on 8 November 2026, and this article will be updated with how stable the lists are across them.
• Matching. Restaurant names were extracted from each answer and matched to a registry of 14,130 Paris restaurants built from Google Maps (exact name, then website domain, then a fuzzy match inside the same arrondissement). 98.9% of mentions matched. Restaurants outside Paris were kept out of the rankings.
• Michelin. Stars from Wikipedia’s list of Michelin-starred restaurants in Paris, 2026 column. 120 of the 127 starred restaurants were matched to the registry; the rest were left unmatched rather than guessed.
• Websites. One fetch of each homepage, checked for booking widgets and scripts, “book” links to a booking provider, site-builder fingerprints and schema.org Restaurant data. Pages behind a bot wall were recorded, not bypassed.
• Prices. Google Maps returned a price level for 3 of 14,130 restaurants, so price fit uses the price ChatGPT itself states.
• Disclosure. I work at Lighthouse, a hospitality data company. This article describes no Lighthouse product.