Ten languages, one question: of five global categories, only one returns the same brands
180 answers, 10 languages, 3 AI assistants. Overlap with English falls to 43% for online payments and 52% for web hosting - and stays intact for running shoes.
In this article
- What we measured, precisely
- Asking the same English question twice does not give the same answer
- The result, category by category
- Why shoes are the exception
- What a customer sees in their language, and English never shows
- English is a market too, not a neutral vantage point
- What this changes for an international brand
- Limits, and what we did not find
Short answer: the language you ask in changes which companies get recommended - in four of the five categories we tested, and not at all in the fifth. We asked the same question in ten languages, naming no country, to ChatGPT, Gemini and Perplexity. 193 companies were cited. 34 in all ten languages. 76 in exactly one.
This is not a translation nuance. For a hosting provider, a payment processor or a booking platform, a visibility audit run in English says almost nothing about what a customer in Tokyo, Seoul or Riyadh sees. For a shoe manufacturer, it says exactly the same thing. The difference between those two worlds is measurable, and it has a simple explanation.
What we measured, precisely
One variable moves: the language. The question names no country. We do not ask for "the best web hosts in Germany" - we ask "what are the best web hosts?" in German. If the list changes, the change can only come from the language, because nothing else moved.
Five global categories, chosen so the same multinationals are candidates everywhere: CRM software (customer relationship management), running shoes, online payments, hotel booking, web hosting. Ten languages - English, French, Portuguese, Spanish, German, Polish, Japanese, Korean, Chinese, Arabic. Three assistants. That is 150 answers, plus 30 control draws.
Asking the same English question twice does not give the same answer
This is the control, and it is the part of the experiment without which nothing is publishable. An assistant is not deterministic: ask it exactly the same English prompt again and it does not return the same list. So we drew English three times per category and per engine, to measure how much an engine contradicts itself.
Two draws of the same English prompt overlap by only 62-76% depending on the category. Any gap between languages has to be read against that floor - otherwise you publish noise and call it a finding.
That floor is what makes the table below readable. The "English vs English" column is what chance alone produces; the next column is what changing the language produces. The gap between the two is the only number that matters.
The result, category by category
| Category | English vs English | Nine other languages | Gap |
|---|---|---|---|
| Running shoes | 74% | 74% | 0 pts |
| Hotel booking | 64% | 52% | 12 pts |
| CRM software | 72% | 59% | 13 pts |
| Online payments | 62% | 43% | 19 pts |
| Web hosting | 76% | 52% | 24 pts |
Running shoes do not move by a single point. Asking in Arabic or Korean returns the same brands as English, within the variation you would get by asking in English twice. Nine brands - Nike, Adidas, ASICS, Brooks, New Balance, HOKA, Saucony, Altra, Salomon - are cited in all ten languages.
Why shoes are the exception
The axis is not language, it is how locally the service has to be delivered. A shoe is the same object in Lyon, Osaka and Jeddah: the maker is global, the product is identical, there is nothing to localise. A payment provider needs a local banking licence; a host needs data centres and support in the language; a booking platform needs the country's hotel inventory.
The more a service must be delivered on the ground, the more the language reassigns the recommendation. That is why web hosting is the most sensitive category in our panel (24 points below the noise floor) and running shoes the least (zero).
What a customer sees in their language, and English never shows
The counting vocabulary was built by reading all 150 answers, then searched identically across all ten languages. So when a name appears in only one language, it is not because we failed to look for it elsewhere.
Japanese, web hosting
: エックスサーバー, ConoHa, ロリポップ. Japanese is the only language where the three assistants cite just ten hosts - 29% overlap with English, the lowest figure in the whole measurement.
Korean, payments
: Toss, 네이버페이 (Naver Pay), 카카오페이 (Kakao Pay), 페이앱, 이노페이. None appears in any of the three English draws.
Arabic, payments
: Tap Payments, HyperPay, PayTabs, MyFatoorah, Paymob, Fawry - six Gulf and Egyptian providers that no other language cites.
Polish, web hosting
: LH.pl, Seohost, home.pl, nazwa.pl, cyber_Folks - five Polish hosts, plus OVH.
Japanese, hotel booking
: 楽天トラベル, じゃらん, 一休, トラベルコ, alongside Booking.com and Agoda, which are cited everywhere.
English is a market too, not a neutral vantage point
This is the counter-test, and it matters: of the 76 companies cited in a single language, 12 belong to English - Helcim, Venmo, GoCardless, Paytm, ScalaHosting, Flywheel, Rocket.net, Nimble, On Running, Accor, Wyndham, Tablet Hotels. English does not give you "the global view" plus a few local quirks elsewhere. It gives a market view, largely North American, exactly as Polish gives its own.
That Accor, a French hotel group, is cited only in English in our measurement is enough to rule out the idea that each language simply pulls its national champions. It is not that mechanical.
What this changes for an international brand
Do not infer your Japanese visibility from your English visibility.
In hosting and payments, the two lists share fewer than a third of their names. A dashboard that only queries English measures one market and displays ten.
Work out which side your category falls on.
If your service needs a licence, a data centre, local stock or local-language support, expect a per-language ranking. If you sell an identical product everywhere, an English measurement is a reasonable approximation.
Your real rivals in a language are not the ones from your home market.
A European payment provider targeting the Gulf is not compared with Stripe and Adyen in Arabic answers: it is compared with PayTabs and MyFatoorah.
Measure more than once before concluding.
The 62-76% noise floor applies to you too: a single measurement cannot tell a real loss of visibility from an unlucky draw.
Local sources weigh more than global ones inside a local language.
Names that appear in only one language come from a corpus that exists only in that language - the country's comparison sites, directories and forums.
Limits, and what we did not find
Five categories, one night, one model per engine: this is a sample, not a census. The nine other languages get a single draw where English gets three - but that is precisely what the floor measures, since it also compares one draw against one draw. The counting vocabulary was built from the answers we obtained: a company nobody named could not be counted.
We did not use our own product's competitor extractor for this study. It is tuned to fill a table of eight rivals from one answer, not to compare ten corpora with each other. A curated vocabulary, applied uniformly, was the only honest instrument here.
Finally, one hypothesis the measurement refuted: we expected Chinese answers to cite mostly Chinese providers. They do cite 阿里云, 腾讯云 and 华为云 - but also SiteGround, Hostinger, Kinsta, WP Engine and DigitalOcean, and they name more hosts than average. The language adds local players; it does not replace the global ones.