Ask an AI engine in the local language and it names different brands. We checked 4 markets.

If you sell in Germany, your buyers are likely to ask ChatGPT, Perplexity and Google in German. Would you see the same answers if you checked in English instead?

We wanted to know whether that matters. So we asked the same buying questions both ways, in the same country, and compared the answers. In our sample it mattered a lot: the local-language answers named largely different brands and leaned on local websites.

What we found

How we ran it

• Markets: Germany, France, Japan and Brazil.
• Questions: 10 buyer questions about coffee (best beans for a bean-to-cup machine, which subscription is worth it, where to buy freshly roasted coffee online, and so on). Each was written natively in the local language, with an English version of the same question. Both were asked from the same country, so the only thing that changed was the language.
• Engines: Perplexity, Google AI Mode, Gemini and Microsoft Copilot, and ChatGPT in Germany. We collected answers through the same data sources Stellarcast uses in production. In our setup Perplexity receives no location at all, which makes it the cleanest read of language on its own.
• Repeats: every question was asked twice in each language. AI answers vary from run to run, so the second run is our yardstick: a language difference only counts if it is bigger than the difference between two runs of the same question.
• What we measured: which brands each answer named (extracted with a language model, then normalized so that a product line counts as its brand and, for many well-known brands, a Japanese spelling matches the Latin one), and which websites each answer cited.
• Size: about 650 answers, collected on 4 October 2026.

Selling in more than one market? Get a free visibility audit and see which brands AI engines name when your buyers ask in their own language.

Finding 1: the local language changes which brands get named

For each question we compared the set of brands in the local-language answer with the set in the English answer, and with the set from a second run in the same language. The chart shows the median share of brands in common.

"Brands in common" is the number of brands named in both answers divided by the number named in either, so it reads lower than a simple match rate. "Questions" counts only the questions where both comparisons had brands to compare, which is why a few rows are out of 8 or 9.

Two things stand out:

• Perplexity is consistent within a language and different across languages. Asked the same question twice, it repeated most of its brands (55-78% in common). Asked in the local language instead of English, it kept almost none (4-11%). It received no location, so in our setup the only input that changed was the language.
• ChatGPT's gap was the smallest. In Germany its answers varied so much from run to run (28% in common) that the language difference, while still bigger on 8 of 10 questions, was the narrowest in the study.

What kind of brands moved? Across all ten questions taken together, many brands appeared in both languages somewhere, and large international brands tended to show up in both. The difference was in which brands a given question surfaced. The brands that appeared in only one language were mostly local names in the local-language answers (regional roasters, but also local retailers, supermarket own brands and domestic appliance makers) and, in all four markets, US coffee brands in the English answers.

Finding 2: local-language answers cite local websites

We also looked at the sources behind each answer: the share of cited pages on the market's own country domain (.de, .fr, .jp, .br). Median per answer, among answers that cited at least one page:

The clear exception was Google AI Mode in Japan, which cited almost no .jp pages, and slightly more in English than in Japanese. On a few engines the shift showed on only about half the questions, which the last column makes visible. Country domains also undercount local sources, since many local businesses use .com, so the real shift is probably larger than this table shows.

The likely mechanism is simple, though we cannot prove it from this data: an engine answering in German searches the web in German, finds German pages, and recommends what those pages recommend.

What this means if you sell in more than one market

These are cautious readings of a first study, not rules:

1. Measure in the language your buyers use. In our sample, an English-only check in Germany would have shown a mostly different shortlist from the one German buyers saw.
2. Treat each market as its own measurement. The brand gap appeared everywhere we looked, but its size varied, and the citation shift did not hold everywhere: Google AI Mode cited local sites in France and almost none in Japan.
3. Local-language sources may matter more than you think. If local-language answers lean on local sites, being covered by those sites (local reviews, local comparison pages, local press) is a plausible lever, and one an English-language content plan would miss.

Limits of this study

We would rather you cite this carefully than widely:

• One category. Coffee has strong local brands. A category dominated by global players (say, smartphones) may show a smaller gap.
• Ten questions, two runs, one day. Enough to see a consistent pattern, not enough for precise figures. Treat the percentages as rough.
• Not every engine everywhere. ChatGPT was checked in Germany only, and Google AI Mode and Gemini are not included for Brazil. We also tried Google AI Overviews in Germany, but an overview appeared for too few of these questions to analyze.
• Brand extraction is automated. A language model read each answer and listed the companies in it. It also counts retailers, review publications and certification labels as brands, and while we normalized the names, some mismatches will remain.

How to check this for your own brand

Pick five questions a buyer in one of your markets would ask. Ask each in that market's language and in English, twice each, and write down which brands are named and which sites are cited. If the lists differ more between languages than between runs, your English-language tracking is measuring a different conversation from the one your buyers are having.

On paid plans, Stellarcast runs this per market, in each market's language, every day, and reports a P25-P75 range instead of a single number so you can see how settled a reading is. The method is in how we measure AI visibility, and the market guides cover what changes from one market to the next.

This is a first study, and we plan to extend it. If there is a market or a category you would like us to test next, email [email protected].