Multilingual GEO: how brands win in Spanish AND English AI search at the same time
Most agencies translate. That is the wrong unit of work. Multilingual GEO is about building two parallel entity graphs — one per language — so AI assistants cite you natively in each market. Here is the implementation sequence.
Founder 6 min read
Summarize with AI Open this article in your preferred assistant
In this article
Most agencies treat multilingual websites as a translation problem. You write the English version, you push it through a translation pipeline (machine, human, or hybrid), and you publish. The pages exist. The hreflang tags exist. The job is done.
Inside an AI assistant, the job is not done. It has not started.
Multilingual GEO is the discipline of making your brand citable in every language your buyers ask AI assistants in — not just present, but cited by name, in context, against the right competitors, for the right prompts. That requires parallel entity graphs, parallel content authority, parallel measurement, and a shared entity backbone that ties them together. Translation gets you one of those five things and skips the rest.
This article is the implementation sequence Citable Agency runs for bilingual EN–ES brands. The same principles apply to any pair of major languages. The mechanics scale.
Why translation alone fails in AI search
There are three structural reasons.
1. AI assistants run language-aware retrieval
When a user types a prompt in Spanish, the retrieval layer behind ChatGPT, Perplexity, or Gemini preferentially pulls Spanish-language content from its index. English content with a Spanish translation slapped on top is retrieved less often than content that was written in Spanish in the first place and has consistent Spanish-language entity signals to back it up.
2. Translation artefacts are detectable and downweighted
Modern LLMs can identify machine-translated text with high reliability — repetitive sentence structure, calque phrasing, register mismatches between paragraphs, untranslated proper nouns left awkwardly inline. Translation artefacts are not a citation positive. In our engagement data, machine-translated landing pages produce a fraction of the citation frequency that native-written equivalents produce, even when the literal information content is identical.
3. Entity signals do not auto-translate
Your Wikidata item is a single entity. But the labels (label), descriptions (description), and aliases (alsoKnownAs) on that item are per-language. If you only have an English label, Spanish-language prompts have nothing to match. Translation pipelines do not fix Wikidata. Schema descriptions, FAQ schema answers, and disambiguating descriptions all need native-language versions too. Translating the page content without translating the entity layer leaves you visible in one language and invisible in the other.
The minimum multilingual GEO stack
Here is what you need in place before you start measuring.
1. Correct hreflang on every page
Every page that exists in multiple languages must declare its language siblings via hreflang. Self-referential hreflang on the canonical page. Reciprocal hreflang on the translated version. x-default set to the most globally-relevant version (usually English for most international brands). Hreflang errors are the most common cause of AI assistants treating two language versions as duplicates and demoting both.
2. Language-tagged schema
Your JSON-LD must carry a top-level inLanguage field on Article, WebPage, FAQPage, and any other content schema. Without it, schema parsers default to the page’s HTML lang attribute — which is correct in most CMSes and broken in many. Make the language explicit. It costs you four characters per page and prevents an entire category of misindexation.
3. Wikidata labels in every target language
Open your Wikidata item. For each language you operate in, set:
label— your brand name, with the language code.description— a single short sentence describing what you are, in that language.alsoKnownAs— any alternate forms or common misspellings in that language.
Free, public, and load-bearing. Most brands have only an English label and wonder why they do not show up in Spanish ChatGPT.
4. Native content, not translated content
Write each language separately. If you cannot afford a native writer for both, you cannot afford bilingual GEO yet — pick the higher-priority language and ship it cleanly, then build the second one when you have the budget. Translation pipelines produce content the model will not cite. There is no shortcut here.
The exception: structural elements (navigation, schema metadata, technical labels) can be translated mechanically with human review. The body of your service pages, journal articles, and FAQ content cannot.
5. Per-language Share of Answer measurement
Run a separate prompt set in each target language. The prompts should reflect how native speakers in that market actually phrase questions to AI assistants — not direct translations of your English prompts. A Spanish ICP does not type what is the best GEO agency for B2B SaaS; they type qué agencia de GEO recomiendas para una startup B2B en España. Different surface form, different competitor field, different winning content.
A brand with 30% SOA in English and 0% in Spanish is, by definition, invisible to half its market. If your measurement is single-language, you do not know which half.
Implementation order
You cannot do everything in week one. The order matters because some fixes compound and some block others.
Weeks 1–2: schema and hreflang foundation
Fix the hreflang errors across the site. Audit every page that exists in both languages, confirm self-referential and reciprocal tags, set x-default correctly. Add inLanguage to every JSON-LD block. This is one-time technical work with compound long-term returns.
Weeks 2–3: Wikidata and entity layer
Add native-language labels, descriptions, and aliases to your Wikidata item. Update your Organization schema across the site to include sameAs to your Wikidata item and to your native-language LinkedIn, Crunchbase, and directory profiles where they exist separately per market.
Weeks 3–6: native content production
Identify your top 5 service or product pages. Commission a native-language writer (not a translator) to produce the equivalent pages in the second language, optimized for the prompts you have already mapped to your single-language SOA baseline. Publish, schema, hreflang.
Weeks 6–12: measurement and iteration
Run per-language Share of Answer monthly. Track separation between EN and ES citation frequency. Where you are winning in one language and losing in the other, the gap usually points at a missing entity signal in the losing language — fix that, re-measure, iterate.
What multilingual GEO is NOT
A short disambiguation, because the market sells some of these as multilingual GEO and they are not.
- Not translation. Translation produces a page in another language. Multilingual GEO produces citation in another market.
- Not auto-translation widgets. A live JavaScript translation widget on top of an English page is not indexable as a second language and AI assistants will not retrieve from it.
- Not subdomain duplication. A
es.yourdomain.commirror with translated pages and no Wikidata, no native authority, and no native prompt set is not multilingual GEO. It is mirrored translation. - Not international SEO with extra steps. International SEO optimizes for ranking in country-specific Google indexes. Multilingual GEO optimizes for citation in language-specific AI assistant responses. The technical foundation overlaps; the success metrics and the work product do not.
The Citable Agency take
Citable Agency operates in Spanish and English natively. Every page on citable.agency exists in both languages, written by Spanish-native and English-native writers separately, with reciprocal hreflang, language-tagged schema, and per-language SOA measurement. Our Wikidata item carries labels and descriptions in both languages. We track our Share of Answer in chat.openai.com Spanish-language responses and English-language responses as two separate metrics — and where the numbers diverge, we know exactly where to invest next.
If you operate in two or more major languages and your current SOA measurement is single-language, you are flying with one eye closed. The AI Visibility Audit at €1,200 includes per-language baseline measurement and a multilingual repair roadmap. If you only operate in one language today but expect to expand, the same audit will tell you what foundations to lay before the second language goes live, so you do not ship a translated mirror that needs rebuilding 12 months later.
Native vs translated content
Source: Citable Agency engagement data, Q1 2026 cohort
What happens when you machine-translate your way into AI search
| Signal | Native content | Machine-translated content |
|---|---|---|
| Share of Answer (Spanish prompts) | 18–32% within 90 days | 0–4% within 90 days |
| Perplexity citation frequency | High | Very low (translation artefacts detected) |
| Wikidata label completeness | All target languages | Usually only source language |
| Hreflang implementation | Correct + per-page | Often missing or self-referential |
| Cost over 12 months | Higher (€) | Lower (€) but produces no leads |
Frequently asked
Questions buyers ask before booking
Is multilingual GEO just doing GEO twice?
Operationally yes — you run the full stack in each language. Strategically no — the prompt sets differ, the competitor sets differ, the authoritative sources differ, and the schema, hreflang, and Wikidata work has to be coordinated across languages so the same brand entity is visible in each one. It is two parallel disciplines that share an entity backbone.
Can I just use machine translation for the second language?
You can, and you will underperform. Native AI assistants weight translation artefacts negatively in their citation logic — repetitive sentence structure, untranslated idioms, awkward register. The drop in citation frequency on machine-translated content versus native content is large enough that any agency claiming bilingual GEO with machine translation is selling you the wrong thing.
Which language should I prioritize if I cannot do both at once?
Prioritize the language with the higher commercial intent in your category, not the larger speaker base. For B2B SaaS in EU markets, English usually wins because most enterprise buyers default to it. For local services in Spain, LATAM, or France, the native language wins because the buyers research in their own language and rarely cross to English.
Does hreflang really matter for AI search?
Yes. AI assistants that crawl your site need to know which language version is authoritative for which market. Missing or incorrect hreflang causes the model to treat your translated pages as duplicates and demote both. Correct hreflang signals that each language version is an independent authoritative resource for its market.