Multilingual GEO: the answer changes with the language

Ask an AI engine the same question in English and then in German and you are not sampling one opinion twice. You are running two different retrievals over two largely different pools of content, and the brands named in the two answers can differ completely. For any brand selling in more than one market, this is the single most underestimated fact in AI visibility.

The language outweighs the brand

Analysis of variance in LLM brand responses puts numbers on it. The language of the query accounts for 26.5 to 32.0 percent of the variance in how models respond about brands. Brand identity, which brand is actually being asked about, accounts for 1.5 percent. The language you are asked in explains roughly twenty times as much of the outcome as who you are.

That ordering sounds absurd until you consider how answers are grounded. Engines retrieve supporting content largely in the language of the question, compose from what that retrieval returns, and cite it. A brand with deep English coverage and nothing citable in French is, from the French retrieval pool's point of view, barely a brand at all. Your reputation does not travel across languages. Your content does, or it does not.

Publish in your markets' languages

The direct implication: if you sell in five languages, the citable substance has to exist in five languages. Quotations, statistics and cited sources move generated answers, and they can only move the answers in the language they are written in.

Written for the market is not the same as machine-translated from headquarters. A translated page carries the home market's examples, prices, regulations and assumptions into a market where they are wrong, and it answers questions the way the original audience asked them rather than the way this one does. Content that earns citations in a market cites that market's sources, uses its figures and currencies, and addresses the questions its buyers actually phrase. Translation tooling can start that work; it cannot finish it.

Map the versions with hreflang

Once language versions exist, they need to be connected. hreflang annotations tell Google which page is the equivalent of which for each language and region, so the right version is surfaced to the right audience instead of versions competing with each other. It is unglamorous plumbing, it is frequently misconfigured, and it belongs on the same checklist as the rest of technical SEO, because a language version that is not correctly mapped and indexed cannot be retrieved, and a page that cannot be retrieved cannot be cited.

Measure per language, or be misled

The variance numbers also condemn a common dashboard: AI visibility measured in English only, for a brand that sells in six markets. If query language drives 26.5 to 32.0 percent of response variance, an English-only measurement is blind to most of what actually varies. A multi-market brand can look dominant measured in English while being invisible in the languages where its revenue lives, and no amount of English sampling will surface that.

Defensible multilingual measurement runs the same prompt set in each market language, tracks share of voice per language, and reports each with its own confidence interval, for the reasons laid out in measuring AI search visibility honestly. The per-language numbers are the decision-grade ones: they tell you which markets have a content gap, and the fix for that gap is the publishing work above, not more measurement.

Where this fits

Multilingual GEO is not a separate discipline, it is GEO done once per market: substance worth citing, in the market's language, technically retrievable, measured in the language buyers use. The brands that treat non-English markets as a translation afterthought are leaving those answers to whoever bothered to write for them.