How Language Changes Which Brand AI Recommends

13 October 2026 · The Cresia team

Yes. The language of the question changes which brands an AI assistant recommends, so an English-only visibility check tells you how you appear to English speakers and nobody else. Search Engine Journal reported research covering 1,650 AI responses in which brand recommendations shifted from one language to another. If you sell in more than one market, your AI visibility is several numbers, and most teams have only measured one of them.

Key takeaways

The same question in a different language can return a different set of recommended brands, so an English-only check covers one market, not all of them.
Treat each language as its own baseline, with its own prompts, competitors and sources, rather than a translation of the English set.
Answer engines assemble answers from retrieved passages and recognised entities, so local-language coverage and third-party mentions count for more than a translated homepage.
Build prompt sets natively with in-market speakers, and record brand, position, citation and sentiment per language and per engine.
Don't chase every language or every daily wobble. Start with the markets that carry revenue and close the largest gaps first.

What the research reported

Search Engine Journal covered a study built on 1,650 responses from AI systems, asked for brand recommendations across several languages. The finding is plain: the brands that came back changed with the language of the question. The outlet's framing was that the visibility check most teams ran was probably in English, and that is the part that stings.

It isn't evidence that AI is random. It's evidence that the assistant behaves like a different advisor in each language. It draws on different text, different review sites, different local press and different habits of naming things.

Two cautions before anyone rewrites a strategy. A sample of 1,650 responses spread across languages and categories is a signal, not a map of your market. And models and retrieval layers change often, so any specific brand ordering will age quickly. The durable lesson is the mechanism, not the leaderboard. Read the original for the languages and categories tested, then run your own check on the ones you actually sell in.

Why the same question gets different brands in different languages

The study reports the outcome. The reasons below are the likely drivers, and you can test each one against your own results.

Start with the text itself. The web is not evenly distributed across languages. A category might have a dense layer of English comparison articles, a thinner one in Portuguese and almost none in Thai. An assistant answering in Thai leans on whatever exists, including local publishers, forums and marketplaces that an English search would never surface.

Naming is the second driver. A brand with a strong English name may be written in a different script, transliterated three ways, or known by a local nickname. If the assistant sees your brand spelled one way in English and another way in local coverage, it may not connect them into one entity.

Third, the competitive set changes. The brands a German buyer considers are not always the brands an American buyer considers. A regional competitor with modest global reach can dominate a local answer simply because the local sources talk about it constantly.

Fourth, the framing of the question differs. A direct translation of an English prompt is often not how people in that market ask. Different phrasing pulls different passages, and different passages mean different brands.

How answer engines choose and cite brands

Most AI answer engines combine a model's built-in knowledge with a retrieval step. The model often breaks a question into several narrower searches, pulls passages that match, and writes an answer from what it found. Ahrefs published a study of 15 million queries on how ChatGPT does this fan-out, and it is worth reading for the mechanics. Each vendor documents its own behaviour, so check the current documentation for the engines you care about rather than assuming one pattern fits all.

Two consequences follow for multilingual work.

First, the sub-queries may or may not run in the user's language. Some may be issued in the language of the question, others in a different one. That is something to observe in each engine's visible search activity or citations, and it can change without notice. If an engine retrieves English sources to answer a Spanish question, your Spanish pages matter less and your English authority matters more. If it retrieves Spanish sources, the reverse holds.

Second, citations come from passages, not whole sites. A passage that states clearly what your product is, who it is for and how it differs gets lifted. A paragraph of marketing language doesn't. This is why a machine-translated page that technically exists in Italian often earns nothing: the passages are vague in two languages instead of one.

Search Engine Land has also written on what it calls the global consistency paradox, building local E-E-A-T for AI search. The tension is real. You want one brand, one set of facts and one positioning everywhere, yet the evidence an engine trusts is local: local authors, local references, local proof.

What to change in your AI visibility tracking

The fix starts with the measurement, because you can't improve a number you haven't separated by language.

List the languages that carry revenue or pipeline. Not every language you have a page in. Rank them by commercial weight and start with the top three or four.
Write prompts natively. Have an in-market speaker, or a trusted agency, write the questions as a buyer there would ask them. Don't translate the English list.
Define the competitor set per market. Include the regional players your English tracking has never named.
Capture the same fields every time. Whether your brand appears, its position in the answer, whether it is cited, which domains are cited, and the tone of the mention.
Run each prompt on each engine you care about. Behaviour differs across assistants and across AI search modes inside search engines.
Store raw answers. Summaries hide the useful detail, such as a competitor spelled a particular way or a source you didn't know mattered.

Teams in growth functions often own the prompt sets, while regional marketers hold the language judgement. Give both a seat, or the prompts drift back to English phrasing.

A measurement plan you can run per market

A table keeps the work honest. Fill one in per language and review it on a fixed rhythm.

Step	What to record	Owner	Why it matters
Prompt set	15-30 native-language questions across awareness, comparison and purchase intent	Regional marketer with SEO lead	Translated prompts don't reflect how buyers phrase things
Brand presence	Mentioned or not, and position in the answer	SEO or AEO analyst	The core visibility number, per language
Citations	Which domains are cited and in what language	SEO or AEO analyst	Shows which sources shape the answer in that market
Competitors	Who appears instead of you, including regional names	Regional marketer	Reveals the real local competitive set
Entity naming	How your brand and products are written in answers	Brand or content lead	Catches transliteration and naming splits
Accuracy	Wrong facts, outdated pricing, invented features	Product marketing	Errors in one language can persist unnoticed

If you track this by hand, a spreadsheet works for a handful of prompts. Beyond that you need a tool and a consistent tagging scheme, which is where MediaPilot comes in for teams tracking AI-search visibility. Whatever you use, keep the language and engine as dimensions from day one. Adding them later means re-running everything.

Pay attention to how you tag the data too. If your analytics are loose about language and market labels, you'll struggle to join AI visibility with traffic and conversions. A written tracking specification for those labels saves a lot of reconciliation later.

Closing the gaps once you can see them

Gaps will look different by market, so diagnose before you produce anything. A few patterns are common enough to check first.

You're absent and competitors are present. Look at the cited sources. If local review sites, directories or publications do the work, your gap is third-party coverage, not your own pages.
You appear, but with wrong or old facts. Fix the source the engine is reading. That may be your own local page, a partner listing or an old press piece.
You appear under a different name. Standardise how the brand and product names are written in each language and use them consistently on your pages, profiles and listings.
Your local pages exist but are never cited. They're probably thin or translated mechanically. Rewrite them with local examples, local pricing, local regulations and a named local author.

A worked example makes this concrete. A B2B software company sells project tooling in Germany and Brazil. Its English tracking looks healthy. The German prompts show a regional competitor named in most answers, cited from two trade publications the company has never pitched. The Brazilian prompts show the company mentioned, but with a pricing tier it retired two years ago, drawn from a reseller page. Different problems, different owners, different fixes. A single English score would have shown neither.

Note that this is mostly work outside your own site. That's uncomfortable for teams used to fixing things by editing pages, but it follows from how these answers are assembled.

What not to do

There is a lot of noise around AI search, and the multilingual angle adds more. A few things aren't worth your time.

Don't machine-translate your whole site and call it coverage. Pages that say little in English say even less after translation, and they dilute the signal that you have real local expertise.
Don't expect stable rankings. AI answers vary run to run. Look at patterns across many prompts and repeated runs, not a single response.
Don't spread across twenty languages at once. A thorough baseline in three markets beats a shallow one in twenty.
Don't treat AI visibility as separate from classic search. The same strong local content and credible mentions help in both. Keep the two programmes joined up.
Don't over-read one study. Use the research as a reason to test your own markets, then act on your own data.

The case for acting is straightforward. If you're recommended in English and missing in the language your next market speaks, that's a fixable gap, and finding it now is cheaper than finding it after a competitor has settled into the answers. If you want to see how a platform approach fits your stack, the platform overview is the place to start.

Frequently asked questions

Does AI really recommend different brands in different languages?

Research reported by Search Engine Journal, covering 1,650 responses, found that it does. The brands returned changed with the language of the question. Your own categories and markets may differ in degree, so test the languages you sell in rather than relying on the study alone.

Do I need a separate AEO strategy for every language?

You need a separate baseline and separate prompts for each language that matters commercially, but not a separate philosophy. The principles are the same: clear passages, consistent entities and credible third-party mentions. What changes is the local evidence, the local competitors and the phrasing buyers use.

Is a translated page enough to appear in AI answers in that language?

Usually it isn't enough on its own. Engines pull passages that answer a question clearly, and a mechanical translation tends to carry over the same vagueness. Pages written for the market, with local examples and named local authors, give an engine something specific to cite.

How often should I re-check each language?

Often enough to catch real shifts, not so often that you react to noise. Pick a fixed rhythm per market, run the same prompts on the same engines each time, and compare across runs. Re-check sooner after a major content change or a notable change in an engine's behaviour, which you can confirm in the vendor's own documentation or release notes.

Which languages should I start with?

Start with the languages tied to your largest revenue or pipeline, then add the ones where you're investing to grow. A narrow, well-run baseline in a few markets gives you more to act on than a thin check across many.

Sources

https://www.searchenginejournal.com/lost-in-translation-how-language-changes-which-brand-ai-recommends/591585/
https://ahrefs.com/blog/query-fan-out-chatgpt/
https://searchengineland.com/topic/generative-engine-optimization