Same question, different language, different brands: a four-market measurement of query-language effects on brand recommendations from AI answer engines

Brands increasingly monitor how AI answer engines recommend them, and most monitoring is done in English. We test whether the language of a buyer's question changes which brands an engine names when the location is held fixed. In four markets (Germany, France, Japan, Brazil) we asked 10 coffee-buying questions in the local language and in English, from the same country, on up to five engines (Perplexity, Google AI Mode, Gemini, Microsoft Copilot, and ChatGPT in Germany), twice each, on 4 October 2026 (about 650 answers analysed). We compare brand sets with the Jaccard index and use the overlap between two runs of the same question in the same language as a noise baseline. In every one of the 15 market-engine cells, the median overlap between local-language and English answers (0-21%) was lower than the run-to-run baseline (20-78%), and the language gap exceeded the noise on at least 8 of the comparable questions in each cell. The median share of citations on the market's country-code domain was higher for local-language answers in 14 of 15 cells, with Google AI Mode in Japan as the clear exception. Perplexity, which received no location in our setup, was highly stable within a language (55-78% overlap) and almost entirely different across languages (4-11%). The results suggest that English-only monitoring of AI visibility can describe a different set of recommendations from the one local-language buyers see. The study is small (one category, one day, two runs) and we report all figures as P25-P75 ranges.