Brand AI Recommends can change when customers ask the same question in different languages, according to tests of Claude, ChatGPT and Gemini. A Search Engine Journal analysis found especially large differences in airline and hotel suggestions, while other product categories produced more stable results.
The tests compared six shopping or booking scenarios in English, Arabic, Russian, Simplified Chinese and Persian. The author collected 1,650 responses across the models and language combinations, then examined which brands each model named and recommended.
Brand AI Recommends Varied By Language
Airline recommendations showed the widest gap in Claude’s tests without web search. For a return flight from London to Tenerife, Claude chose British Airways in 52.5% of Persian responses and 12.5% of English responses. In Russian, easyJet received 72.5% of recommendations, while British Airways was not selected.
The hotel results also differed. Bahía del Duque led in 43.6% of Chinese responses but 7.7% of Arabic responses. Hotel Botánico appeared in nearly a third of English recommendations and did not appear in Russian results.
Other categories were steadier. Sony led the headphone recommendations in all five languages, including 50.0% in English and 100.0% in Russian and Persian. Roborock led robot vacuum results in each language, ranging from 87.5% in English and Arabic to 97.5% in Chinese.
Project management software was a closer contest. Monday.com received the most recommendations in every language in Claude’s core test, with shares from 45.0% in English to 55.0% in Russian and Chinese.
Live Search Changed Brand AI Recommends
When Claude could search the web, the leading brand changed in seven of the ten tested language and category combinations. In the hotel scenario, the winner changed in all five languages. JOIA by Iberostar took 31% of picks overall and 60% in English, while ten hotel names appeared that had not appeared in the earlier results.
The author connected JOIA’s appearance to a July 2024 rebrand of Iberostar’s Grand tier hotels. Without web search, Claude had recommended an Iberostar Grand hotel only twice in 198 responses. The test suggests that current web information can affect which brands a model recommends.
Web search also changed project management results in Arabic and Persian. Wrike won 46.7% of Arabic responses, compared with 5% when search was off. In Persian, Asana led with 40%, while Monday.com received 26.7%.
Those results have a caveat: 42.7% of search-enabled responses reached the output limit before completion. Truncation was especially frequent in Persian, at 83%, and Arabic, at 63%. The author said the model named its top choice early enough to capture nearly every recommendation, but the figures may still be uncertain.
Across the comparisons, Gemini matched Claude’s leading brand in five of ten combinations, while ChatGPT matched it in one. ChatGPT favored Wrike in four project management language tests.
Test Customer Questions In Their Languages
The author cautioned that the research used a small set of models, categories and languages, and that the tests took place on September 7, 2026, in one afternoon. The article says the percentages may fluctuate and should not be treated as a complete map of multilingual AI recommendations.
For businesses, the proposed next step is to test questions customers ask before buying across relevant models and target-market languages, then repeat the tests to observe patterns over time. This connects with broader work on AI search visibility and the ways public relations can shape search visibility. Brands can also review the requirements for Google Knowledge Panels as part of their wider search presence.
The tests do not identify why language-specific differences occur, and the author says the factors may be linguistic, cultural, regional or political. The immediate implication is narrower: a single English-language check may not describe what customers see in other languages.
Read the multilingual recommendation study for the full breakdown.



