Multilingual AI Search Visibility: Diagnose Language and Locale Mismatches

Diagnose multilingual visibility by separating three things: language, country and commercial locale. A translated page can be perfectly readable and still answer the wrong market’s pricing, availability or policy question. Check the content and the URL relationships first, before assuming that an assistant simply prefers English.

In short

• Decide, per URL, which language and which market it is for, and write that down.
• Make the visible answer complete in each language. Translating the navigation is not localising the page.
• Give every language version a reciprocal hreflang cluster and a canonical in its own language.
• Test each language and each market as its own panel of questions, with the exact context recorded.
• When a wrong-language source is cited, record it and check your own pages before theorising about the engine.

Define the intended audience per URL

Create a locale record for every URL that is meant to serve a market: its language, the country where one applies, the currency, availability, support conditions and the canonical URL. A language-only page and a country-specific offer serve different purposes, and a country flag is not a definition of a language.

Field	Why it matters
Language	Decides which readers and which questions the page answers
Country or region	Decides which prices, delivery terms and legal text apply
Currency and price	The most often quoted commercial fact, and the most often stale in a translation
Availability and support	Whether the product can be bought or supported in that market
Canonical URL	The version you want treated as the source for this locale

A fictional English and Hungarian product guide might explain the same capability while linking to different local policies. The Hungarian version should not silently inherit a commercial claim that is unavailable in Hungary.

Check the answer in the visible content

Read each translation with someone who understands the product and the market. Verify terminology, units, qualifications and exceptions. A translated heading does not mean the underlying answer was localised.

Google determines a page’s language from its visible content, not from lang attributes or the URL, and recommends one language for the content and navigation of a page, with no side-by-side translations (language signals, grade A). The same guidance treats localised versions as duplicates only when the main content stays untranslated, and lists many automatically translated pages with little value to users as scaled content abuse (translation parity, grade A). The practical reading: translate the substance, and publish fewer, better pages rather than many thin ones.

Mixed-language navigation or a boilerplate introduction can obscure a page’s intended language. Inspect the initial HTML, and the page without its layout, to confirm that the substantive answer is present. Account for language selectors that send a visitor to a generic page instead of the equivalent translated section.

Align canonical and alternate relationships

Each language version must list itself and all the other versions in its hreflang annotations. If two pages do not point to each other, Google ignores those annotations, though it still processes the pairs that do. Codes must be ISO 639-1 language codes with an optional ISO 3166-1 Alpha 2 region, and x-default names the fallback page for unmatched languages (hreflang reciprocity, grade A). With hreflang, specify a canonical in the same language rather than pointing every translation at the English page (canonical and hreflang, grade A). A declared canonical is a hint, not a rule, and links to the canonical URL from your own navigation strengthen it (canonical consistency, grade A).

Do not redirect visitors automatically by guessed language or IP location. Google advises against it, because most crawls come from the United States with no Accept-Language header, so other versions may never be crawled (geo-targeting and currency, grade A). Offer a visible selector and distinct URLs instead.

These are Google’s instructions. Do not describe hreflang as a universal rule that every assistant uses when choosing citations.

What the retrieval research suggests, and what it does not

Multilingual embedding models place equivalent text in different languages close together, so a question in one language can retrieve a passage in another. In multilingual retrieval experiments, retrievers preferred documents in the language of the query and in high-resource languages, and the generators preferred the query language (multilingual retrieval, grade B). That supports writing a complete page in the language your buyers ask in. It does not tell you what any particular production engine does, so measure it.

Build a language-and-market observation grid

Test field	What to record
Question	Natural wording written for the intended reader
Language	Actual prompt and answer language
Market	Country or location setting, where supported
Engine and surface	Exact observation context
Expected source	Localised URL that owns the answer
Observed source	URL and locale actually cited
Failure	Unavailable setting, request failure or no answer

Keep the wording stable within each language cohort. A literal translation of an English buying question may not match how local customers ask it. Label discovery prompts separately from the panel used for month-to-month comparison, so a change in the panel never looks like a change in visibility.

Diagnose a wrong-locale citation

1. Check the intended page. Does the local page contain the complete answer, in visible text, in the initial HTML?
2. Check the plumbing. Internal links, indexability, canonicals, reciprocal alternates and contradictory market facts.
3. Compare sources. Is an English source more current or more specific for this task than the local one?
4. Record the result. An assistant may select a third-party or foreign-language source for reasons you cannot observe. Write down what happened rather than inventing a ranking explanation.

A local simulation can highlight answer gaps in your content. It cannot reconstruct the assistant’s language selection.

Maintain translations as a set

When a product fact changes, identify every locale that contains it and assign a review owner. Track whether each translation is updated, intentionally different or pending. Keep publication and update dates accurate: do not make every version look current after one edit.

Common mistakes

• Canonicalising every translated page to the English version.
• Translating the template and leaving the body in one language.
• Redirecting by IP address or browser language.
• Carrying English prices or policies into a market they do not apply to.
• Comparing months of results after silently changing the question panel.

Questions people ask

Do AI assistants use hreflang?

Hreflang is documented by Google for its own search. Whether another assistant uses it is not documented, so treat it as good hygiene, not as a lever you can count on.

Should I machine-translate every page?

Not without review. Automatically translated pages that add little for users are listed by Google as scaled content abuse, and a wrong price in translation is worse than no page.

Is a single English page enough?

Sometimes. If your buyers ask in English in every market, a well-built English page may do the job. Test the questions your buyers actually ask before you decide.

Related

Vercatus’s markets guide explains how markets and languages are configured. The entity-consistency guide helps reconcile facts shared across locales, and the freshness workflow records real revisions. Measure local citation outcomes separately from the correctness of your international setup.