News & analysis
Same AI, Different Language, Different Hotel Pick
New testing found Claude recommended different airlines and hotels for the same London-to-Tenerife trip depending on whether it was asked in English, Arabic, Russian, Persian or Chinese.
By Roomer Editorial · · 3 min read
What did the testing find?
A Search Engine Journal analysis ran the same six shopping and booking questions through Claude Sonnet 5 in five languages, 40 times each, with web search turned off. That's 1,200 responses total.
Two of the six scenarios were travel questions: a return flight from London to Tenerife, and a hotel in Tenerife. Both showed the widest swings in the entire dataset.
In Persian, Claude picked British Airways 52.5% of the time and easyJet 12.5% of the time. In Russian, the same question flipped entirely: easyJet won 72.5% of responses, and British Airways was recommended zero times.
For hotels, Bahía del Duque topped 43.6% of Chinese-language responses but only 7.7% of Arabic ones. Hotel Botánico took nearly a third of English responses and didn't appear at all in Russian.
Why do airlines and hotels swing more than other categories?
The article tested four other categories: headphones, CRM software, robot vacuums, and project management software. None of them moved anywhere near as much.
Roborock won the robot vacuum question regardless of language. Monday.com beat Asana in every language tested, consistently, in a category the author calls a two-horse race. Sony dominated headphones everywhere except English, where it faced real competition.
Travel categories didn't behave that way. The author points to the sheer number of hotel options in a market like Tenerife as one likely factor — more candidates means more room for the model's language-specific reasoning to land on a different answer. Beyond that, the author is candid that the reason some categories stay consistent while others swing this much remains unclear.
Does this connect to a broader pattern in how AI models behave by language?
Yes. The testing was prompted by an Anthropic report published in July, which analyzed close to 310,000 real Claude conversations across 20 languages and three model versions. That report found the values a model expresses — warmth versus rigor, caution versus deference — shift depending on the language used, even when the underlying question is identical. Claude leaned toward warmth in Hindi and Arabic, and toward rigor in English and Russian.
The SEJ testing is separate from that Anthropic work and looks at something Anthropic didn't measure: whether a change in language also changes which specific brand gets named.
Does this matter outside English-speaking markets?
It matters especially there. Statcounter's August 2026 figures show ChatGPT accounted for roughly three-quarters of AI chatbot referral traffic in Saudi Arabia, Kazakhstan, and Singapore. Add Claude and Gemini, and the three U.S.-built models combined make up more than 90% of AI referrals in each of those markets.
This goes beyond geography. More than half of Los Angeles County residents speak a language other than English at home, including an estimated 74,000 Persian speakers, according to Census data cited in the piece.
Roomer Editorial's take
If you're a travel creator building out destination or hotel content, the airline and hotel findings are the part to sit with. These were the two categories in the entire test that moved the most by language — more than headphones, more than software. Travel recommendations appear to be unusually sensitive to which language an AI model is reasoning in.
That's a problem if your audience isn't monolingual. A creator covering Tenerife hotels for an English-speaking audience and a creator covering the same destination for Persian or Russian speakers may be competing in what are effectively two different recommendation pools inside the same chatbot. Checking your brand's AI visibility once, in English, tells you nothing about how a model behaves when asked the same question in Arabic or Chinese.
This is Roomer Editorial's view, not something the source claims directly: the fix isn't trying to game every language variant of every model. It's making sure your own content — the thing you actually control — doesn't depend on being picked by a chatbot in the first place. A bookable page on your own site works the same whether a reader found you through a Spanish-language search, an English one, or a direct link, because the booking rail and the commission sit with you, not with whichever language the AI happened to reason in that day.
If you publish hotel or airline content for an audience that isn't exclusively English-speaking, pull up your own past recommendations and check what a chatbot says about the same property or airline in your audience's other main language — the gap may be bigger than you'd guess.
Source: Search Engine Journal — Lost In Translation: How Language Changes Which Brand AI Recommends
