Translation produces a second version of your facts
Localisation is supposed to adapt how something is said. What frequently happens is that adaptation extends to what is claimed: a feature boundary softens, a price model changes, a limitation disappears, a support commitment is stated more strongly in one market than another. Each version is locally reasonable, and collectively they are a set of conflicting facts about one product.
This matters more than it did for search ranking. A search engine can be pointed at the right language version. A model answering in English has no such control, and the versions it finds do not come with an indication of which market they describe.
- Audit the same attribute across language versions and record where the wording differs, not only where the facts differ. The site scan reads the pages of one site and reports the descriptions that conflict, which is the first place a language-version conflict becomes visible.
- Treat differences in pricing, feature limits, support terms and compliance claims as conflicts, because those are the attributes buyers check
- Decide which version is canonical for each market, and make that decision explicit rather than emergent. The brand check reports the attributes a model currently returns for your brand, which is how you see whether it picked the version you expected.
How a model resolves a conflict
When sources disagree, a model does not resolve the disagreement; it picks. The material it selects depends on which version is easiest to retrieve, which matches the language of the question most closely, and which reads as more authoritative. All three of those are properties of your site rather than properties of the truth, which means the version that gets quoted is often not the one you would choose.
- The more versions of a fact that exist, the more ways there are to be quoted incorrectly
- Consistency in the canonical attributes reduces the number of quotable variants far more than adding new pages does
- Machine-readable facts that state scope and market explicitly help a model attribute a claim to the right version
Converge the facts, keep the language local
The mistake to avoid is the opposite one: standardising every market onto a single language, which produces accurate pages that nobody reads. The distinction that matters is between the facts and the expression. A price model, a feature limit, a support commitment and a compliance claim should be identical across markets. The way you explain them, the examples you use and the objections you address should be local.
- Publish a canonical attribute set covering product scope, pricing model, limits, support terms and legal entity, and require every language version to match it
- Keep the canonical set in a machine-readable file so the same facts are available to crawlers regardless of page language
- Let local pages differ in examples, terminology and objection handling, and specify those as the permitted variance
- Review conflicts on a schedule tied to product changes, since a pricing or packaging change invalidates the agreement automatically
Checking the market and language dimensions separately
A multilingual site needs a measurement design that can tell a language problem from a market problem, because they have different causes. A question asked in English that returns a Chinese-market answer indicates a retrieval or targeting problem. A question asked in English that returns a French price model indicates an entity conflict. Both look like 'the answer is wrong' if the test set does not separate them.
- Run the same question set in each market language, with the market named in the query where buyers would name it
- Record the language of the answer separately from the market it describes
- Watch for answers that mix a language with another market's facts, which is the signature of an unresolved entity conflict. The crawler access check confirms that each language version is actually reachable, so a missing version is not mistaken for an entity conflict.
- Include a small set of questions with no market qualifier, because those are where cross-market bleed shows up first
Multilingual consistency is not about writing the same sentence everywhere. It is about making sure the facts a model can quote do not contradict each other, while leaving everything else genuinely local. Sources reviewed on 2026-10-20. This article records a method with verifiable boundaries; it does not claim that any search or AI platform will index, cite or recommend anything.