Ask ChatGPT for Paris hotels in any language — it cites the same English sources
We ran 4,875 ChatGPT hotel searches for Paris across 5 languages and 5 countries. The hotels it names shuffle. The sources it reads almost never do.
A traveller in Tokyo asks ChatGPT, in Japanese, for the best hotels in Paris. A traveller in Madrid asks in Spanish. Do they get a different Paris — drawn from their own language’s web — or the same one? We held the city fixed (Paris), and varied two things: the language of the question and the country it’s asked from. 39 hotel prompts, 5 languages (EN, FR, ES, DE, JA), 5 proxy countries (US, FR, ES, DE, JP), 5 repeats each — 4,875 live ChatGPT searches with web search on.
The sources don’t move
Share of citations that come from the local-language top-level domain, by the language of the question. If language pulled in local sources, these bars would be tall. They’re not.
French scrapes a token 4.9% from .fr; German, Spanish and Japanese are under 1%. Across every language, ~44–51% of citations are plain.com. And the same outlets top the list regardless of the asking language:
Top sources across all 4,875 searches
Citation counts pooled across all languages and countries. Every one of these is an English-language outlet.
Even when both signals say “Japan”
The proxy country doesn’t rescue local sources either — asking from a French IP nudges.fr to 5.8%, but German, Spanish and Japanese IPs all stay near zero. The sharpest case is the one where both the language and the location point the same way:
.jp) sources when asking in Japanese, from a Japanese IPBut the hotels do shuffle
This isn’t to say language does nothing. The specific hotels named do change — only ~40–46% of the top picks overlap between English and each other language (averaged over the five repeats, so it’s signal, not run-to-run noise). French queries surface iconic French luxury (Ritz Paris, Plaza Athénée) that English misses. So roughly half the recommended hotels differ by language — they’re just selected from the same English-language source pool.
Top-5 hotel overlap (Jaccard), English vs each language, US IP, averaged across 39 prompts.
It answers in your language, reads in English
The disconnect is most vivid in Japanese: 100% of Japanese queries got an answer written in Japanese — fluent, native, properly localised. And underneath that Japanese answer, the citations were Time Out, oyster.com and Reddit. ChatGPT speaks your language; it just doesn’t read it. The map panel of hotels surfaced ~79% of the time, identically across all five languages — the experience is localised on the surface and English underneath.
Why it matters. If you run a hotel in a non-English market, this is the English-corpus tilt one layer deeper than we’ve seen before: being visible to AI in your own language’s web isn’t enough, because AI isn’t reading it. The sources that decide your Paris ranking are English-language ones — Reddit threads, Time Out lists, oyster.com reviews — no matter who’s asking or from where. That’s where the visibility work has to go.
Method & limits
4,875 live ChatGPT searches (web search on), captured via Bright Data, June 2026. One city (Paris) for clean comparison. 39 distinct hotel prompts (control, star tiers, audiences, neighbourhoods, landmarks, amenities), each natively written in English, French, Spanish, German and Japanese, run from US, French, Spanish, German and Japanese IPs, repeated 5×. We log every citation and the structured map/hotel panel. “Local source” here = the citation’s domain on the language’s ccTLD.
.com) — but the dominant outlets are unambiguously English (Reddit, Time Out, Expedia), so the direction is firm. One city, five languages, one model snapshot; ChatGPT is non-deterministic, which is exactly why every cell is repeated and the hotel-overlap is averaged. Results describe ChatGPT’s web-search behaviour, not other engines.FAQ
The specific hotels change — only ~40–46% of the top picks overlap between English and French/Spanish/German/Japanese, so roughly half differ. But the sources ChatGPT cites barely change: it reads the same English-language web (Reddit, Time Out, oyster.com, Expedia) regardless of the asking language. The hotels shuffle; the sources don’t.
Summarize with AI
The English tilt, one layer up
We also measured which hotels are even in the open crawl that trains AI — and found 39% absent.
Read: Are hotels in Common Crawl?