Article contents
A generative engine answers in the language of the prompt, but it does not choose its sources from a single language
The answer follows the language of the question: a prompt in Spanish produces an answer in Spanish. The sources, however, are chosen from an index that mixes languages. In our observations, run in French, Spanish and English from Alicante, Valencia, Madrid and Paris, most of the sources cited are in the language of the prompt, but English sources appear regularly on technical and B2B topics, and a French or Spanish source can be picked up, translated, for a question asked in the other language when no local equivalent exists.
Two consequences follow. First: the language versions of the same site are not independent; they share an entity, figures and a reputation, and the engine reads them together. Second: multilingual dilution has a precise shape. It occurs when the versions say different things, when the entity's name varies from one language to another, or when a version is merely a word-for-word translation that does not answer the questions actually asked on its market.
This article deals with those risks in order: the language of the sources and what it implies for writing, adaptation by market, signalling versions with hreflang, consistency of the entity across languages, then measurement by language. It builds on the complete method, whose four conditions, accessible, citable, identified, corroborated, apply to each version separately.
Language of the prompt, language of the sources: what happens in between
The mechanism is that of any generative search: the prompt is broken down into sub-queries, each sub-query brings back passages, and the model writes. Language comes into play at two moments. At the sub-query stage, the engine generally formulates in the language of the prompt, and sometimes also in English for topics whose documentation is mostly in English. At the selection stage, it prefers a passage in the language of the answer, because it can quote it without translating; a passage in another language is retained if it is clearly better, or if it is the only one.
- 01Prompt in Spanish“¿Qué mantenimiento necesita una bomba de calor aerotérmica?”
- 02Sub-queriesFormulated in Spanish; one or two in English if the topic is technical.
- 03RetrievalSpanish, English and sometimes French passages, from the same index.
- 04SelectionAt equal quality, the Spanish version is preferred: it can be quoted without translation. Without a Spanish version, your French page may be translated… or ignored.
- 05Answer in SpanishMostly Spanish sources, a few English ones; your brand is among them if a readable version existed.
What this implies for writing: each language version must be citable on its own, with its own passages of forty to eighty words, its own evidence and its own date. A Spanish version that refers to a table in French, or an English version that cites a French decree without explaining it, is not a citable passage for a Spanish or English prompt. The rules of the citable passage apply to every language, not only to the original one.
A word on English. On B2B and technical topics, English-language sources are over-represented in answers, including to French or Spanish prompts. An English version of your key pages is therefore not just a door to the English-speaking market; it is sometimes the version that will be cited, translated, for a question asked in French.
Translate or adapt? Each version must answer the questions of its market
Word-for-word translation produces a grammatically correct version that answers the questions of the source market. Yet questions change from one market to another, even with the same intent. A French buyer asks “quel délai”; a Spanish buyer asks “cuánto tarda”, but also “qué documentación necesito”, because the procedures differ; an English-speaking buyer often starts with the price. Question headings, evidence and examples must come from the target market, not from the source version.
- 1The questionsHeadings and FAQ built on the market's real phrasings, collected in its language.Otherwise: your headings match no local sub-query.
- 2Local evidenceRegulations, currency, units, dates and bodies of the country.Otherwise: the passage is true elsewhere, therefore false here.
- 3The examplesCases and situations that speak to the market: cities, types of business, uses.Otherwise: the text sounds foreign, and the engine reads it as such.
- 4The local entityAddress, registration, team and author of the country, attached to the same organisation.Otherwise: three weak entities instead of one.
On our own pages, produced in Spanish, French and English, we proceed as follows: the source version is written by an expert in their language, then each other version is entrusted to a native speaker who knows the trade and who rewrites the questions, replaces the evidence and the examples, and keeps the figures and the structure. An automatic first draft is acceptable to save time; it is never published without that rewrite.
| Element | French version (source) | Adapted Spanish version | What changes |
|---|---|---|---|
| Question heading | “Quel délai pour être cité par ChatGPT ?” | “¿Cuánto tarda una empresa en aparecer en ChatGPT?” | The market's real phrasing, not a translation |
| Evidence | Timeframe observed on French engagements | Timeframe observed on Spanish engagements, or an explicit note that the observation is common to both markets | The origin of the evidence |
| Example | An SME in Lyon | An SME in Valencia | The local anchoring |
| Entity | SEO360, Paris office, French-speaking author | SEO360, Alicante, Valencia and Madrid offices, Spanish-speaking author | The local attachment of one and the same organisation |
| Figures and dates | Identical | Identical | Nothing: a figure that differs between versions destroys corroboration |
Table: scroll horizontally.
hreflang and URLs: telling the engine which version serves whom
The hreflang signal indicates, for each page, its equivalents in the other languages. It is documented by Google. For generative engines, no vendor has, to our knowledge, published a specific rule; those that rely on Google's or Bing's index benefit from it indirectly, and the others read at the very least the lang attribute and the URL structure. The cost is low, the absence is costly: without hreflang, the engine may cite the English version for a French prompt, or treat three equivalent pages as duplicates.
<!-- In the <head> of each version: the same four links, reciprocal -->
<link rel="alternate" hreflang="fr" href="https://www.exemple.com/fr/entretien-pompe-a-chaleur/">
<link rel="alternate" hreflang="es" href="https://www.exemple.com/es/mantenimiento-bomba-de-calor/">
<link rel="alternate" hreflang="en" href="https://www.exemple.com/en/heat-pump-maintenance/">
<link rel="alternate" hreflang="x-default" href="https://www.exemple.com/en/heat-pump-maintenance/">
<!-- And a single language declared per document -->
<html lang="es">| URL structure | Example | How crawlers read it | Recommendation |
|---|---|---|---|
| Subdirectory | exemple.com/es/ | A single domain entity, clearly separated versions | Recommended in most cases |
| Subdomain | es.exemple.com | Separate versions, domain entity sometimes diluted | Acceptable if the infrastructure requires it |
| Country domain | exemple.es | Strong local anchoring, entity to rebuild per domain | Justified for a distinct subsidiary |
| URL parameter | exemple.com/page?lang=es | Often ignored, or mistaken for a duplicate | To be avoided |
| Redirect based on IP address | exemple.com to the version chosen by the server | Crawlers see only one version, chosen from their address | Rule it out: offer a selector, do not redirect |
Table: scroll horizontally.
Three practical rules. The hreflang links are reciprocal on each version. Each URL returns a single language, with no mixing, including in the title, the meta description and the structured data. The x-default version designates the page that welcomes visitors whose language is not served, often English. And one rule about crawlers: AI search crawlers arrive with no declared language preference; an automatic redirect based on browser language or IP address hides the other versions from them, and sometimes the right one.
A single entity in three languages: the consistency that makes the source
The most discreet risk of multilingual sites is the fragmentation of the entity. If the French site says “SEO360”, the Spanish site “SEO360 Alicante” and an English profile “SEO 360 Agency”, the engine reconstructs three organisations, each with a share of the mentions and the evidence. The “identified” condition fails three times. The rule is simple: one name, one organisation, equivalent descriptions, and offices attached to the same entity.
Person per author, with their publication languages, attached to the same organisation.In the markup, this translates into a single Organization, referenced by an identical @id from all three versions, with the same sameAs and the same legalName; only the textual properties are translated, and each page declares its language through inLanguage. Building the entity, registers, profiles, Wikidata, is covered in the article on the brand entity; here, the essential point is that it is one, whatever the language.
The three multilingual dilutions, and how to avoid them
Three defects recur in our audits of multilingual sites, and each is corrected by a production rule rather than by a technical fix.
- Contradiction between versions. A six-week timeframe in French, two months in Spanish; a price updated on one version and not on the other. The engine that reads both sees an inconsistent source and prefers a third party. Remedy: figures are centralised, the three versions are updated on the same day, and the update note exists in each language.
- Unreviewed machine translation. It produces a correct, generic text whose headings match no local question and whose examples sound foreign. Remedy: a native speaker who knows the trade rewrites each version, as described above.
- The mixed page. A body in Spanish, quotations in English, interface labels in French, a title left in the source language. The crawler reads a page whose language is not clear-cut, and selection suffers. Remedy: one language per URL, including for interface elements, title tags and structured data.
A partial version is often worse than none. Thirty pages in Spanish that link to French pages for the detail create thirty incomplete passages, and signal to the engine that the Spanish version is not the reference version. Ten complete Spanish pages, citable on their own, are better than thirty translations that stop halfway.
Measuring by language: one panel per market, not a translation of the panel
Citation share is measured by language, with a panel built in each language from the real phrasings of each market, not by translating the French panel. The questions differ, the dominant engines differ, and so do the third-party sources. We run our own panels in French, Spanish and English, and the results for one and the same brand vary greatly from one language to another: present on one market, absent on the other, often because the third-party sources that corroborate it exist in only one language. Building the panel is detailed in the article devoted to the prompt panel.
Illustrative data. What to take from it: the source version gains little; the translated versions, which answered no local question, are the ones that gain most once adapted.
Two readings complete the measurement. By language and by engine, first: ChatGPT, Perplexity, Gemini and Claude do not rely on the same indexes and do not have the same coverage across languages; a market can be won on one and lost on another. By language of the sources cited, next: if Spanish answers cite English sources, your English version is a lever for the Spanish market; if they cite Spanish directories and media, it is local corroboration that must be worked on before the pages.
The most frequent multilingual mistakes
The six mistakes below can be checked in an hour on any three-language site, and each costs citations on at least one market.
- A brand name varied by country, which fragments the entity into as many organisations.
- Word-for-word translated versions whose headings match no local question.
- Figures or timeframes that differ from one language to another.
- An automatic redirect based on IP address, which hides the other versions from crawlers.
- A non-reciprocal
hreflang, or one missing from the secondary versions. - A measurement panel translated from French, which measures questions nobody asks in Spanish.
What to remember
- The answer follows the language of the prompt; sources are chosen from several languages, with a preference for the language of the answer.
- Each version must be citable on its own: the market's questions, evidence and examples, identical figures across languages.
- Reciprocal
hreflang, one language per URL, no automatic redirect based on IP address. - A single entity: same name, same
sameAs, same organisation, equivalent descriptions. - One panel per language, built on local phrasings, read by engine and by language of the sources cited.
Frequently asked questions
Can a generative engine cite my French page for a question asked in Spanish?
Yes, it does happen, especially when no equivalent Spanish source exists: the engine translates the passage and cites the French page. But it is a fragile position. As soon as a Spanish source of comparable quality appears, it is preferred, because it can be quoted without translation. An adapted Spanish version is the only way to hold the position durably.
Should all pages be translated, or only some?
Only those that serve a valuable sub-query on the target market, and they must be adapted, not translated. Ten complete Spanish pages, with their local questions, evidence and examples, are cited more often than a hundred partial translations. The selection is made from the Spanish prompt panel: the questions it contains designate the pages to produce first.
Is machine translation acceptable for generative engines?
As a first draft, yes; as a published version, no. A machine-translated text is correct but generic: its headings do not use local phrasings, its evidence remains that of the source market, and its examples sound foreign. A native speaker who knows the trade must rewrite the questions, replace the evidence and the examples, and check that the figures have remained identical to the source version.
Do you need one domain per country or a single domain with subdirectories?
In most cases, a single domain with one subdirectory per language: the entity remains unique, the versions are clearly separated, and the hreflang signal is enough to guide the engines. A country domain is justified for a genuinely distinct subsidiary, with its own legal entity and its own third-party sources, accepting that corroboration has to be rebuilt for each domain.
How do you measure generative visibility across several markets?
With one prompt panel per language, built from the real phrasings of each market, run every week on the same engines. Citation share is read by language, by engine and by language of the sources cited. These three readings indicate, market by market, whether to adapt the pages, work on local corroboration or strengthen the English version that feeds the others.
An SEO practitioner since 2001, Kamel Malek runs SEO360 (Alicante, Valencia, Madrid, Paris). He has published three books on visibility in generative engines, including Generative Engine Optimization and Rétablir les faits.
