--- title: "Multi-language AEO: hreflang and translation rules | Answerly" description: "AP Education shipped EN, UA, PL on every page. Multi-language architecture that compounds, hreflang implementation, what we never machine-translate." url: https://answerly.agency/blog/multi-language-aeo/ lang: en updated: 2026-04-08T00:00:00.000Z --- Frameworks · 6 min read # Multi-language AEO: the hreflang and translation rules that compound AP Education shipped EN, UA and PL versions on every priority page. Here is the multi-language architecture that compounds across regions, the hreflang implementation, and what we never machine-translate. Dmytro Popryadukhin · 2026-04-08 ## Key takeaways - Hreflang is mandatory; without it, AI extractors confuse language variants and dilute authority. - Schema and structured content translate cleanly; marketing copy never machine-translates. - Native voice on each language version — bilingual editor, not Google Translate. - Each language gets its own llms.txt path and prompt cluster. ## Quick Facts | Parameter | Value | | --- | --- | | Hreflang implementation | in every page head | | Translation policy | Schema yes (machine), copy no (native editor) | | Languages where AI tracks separately | Per-language LLM model behaviour varies | | Routing pattern | /\[lang\]/path or path?lang=\[lang\] | | llms.txt per language | Recommended on multi-language sites | ## Why multi-language matters for AI LLMs respond differently to the same prompt in different languages. A Polish-speaking buyer asking about “najlepsze krypto-licencje dla startupów” gets a different answer set than the English-speaking buyer asking the same question — different sources weighted, different brand positions in the answer. If your brand only ships in English, you are absent from every non-English prompt. For brands with EU, MENA or LATAM market exposure that is meaningful pipeline. The [AP Education engagement](/cases/ap-education) shipped UA, EN and PL on every priority page. The PL version pulls separate citations from Perplexity in PL and ChatGPT querying in PL. None of that surface existed in the English-only version of the site. ## Hreflang as the foundation `` in every page head, covering every language variant including `x-default`. Example: ``` ``` Without hreflang, AI extractors treat `/services/foo` and `/ua/services/foo` as duplicate content with different language. Authority dilutes; both versions perform worse than they should. This site (answerly.agency) deploys hreflang on every page out of the BaseLayout — see the head source of any page, the alternates are emitted per-locale automatically. ## What machine-translates and what does not **Machine-translate (acceptable):** - Schema markup (Article, FAQPage, Person property values for non-name fields) - Quick Facts table contents - Technical labels (form fields, navigation items) - Footer boilerplate **Never machine-translate:** - Hero copy - X-is-Y intro - H2 questions and direct answers - FAQ questions and direct answers - Author bios - Anything where the brand voice should come through The reason: machine translation passes the meaning but kills the voice. AI extractors weight content quality alongside structural compliance, and machine-translated copy reads as low-effort. We have measured this — pages with machine-translated marketing copy underperform native-edited versions by 40-60% on citation rate in the secondary language. The discipline: native editor or bilingual writer per language. The [AP Education engagement](/cases/ap-education) had separate UA, EN and PL editors on retainer. The translation cost is real but the citation lift in the secondary language pays for it within months. ## Per-language prompt cluster Each language gets its own prompt cluster. The PL prompt cluster for crypto licensing is not a translation of the English prompt cluster — Polish buyers ask different questions because the Polish regulatory landscape is different. Mining prompts in each language uses the same tools (Searchable, Profound) with language-specific filters. Five seed prompts per language, plus the conversational variants. For Scale or Enterprise engagements with three languages, that is 150–200 tracked prompts total — but spread across three teams (one editor / writer per language). ## The llms.txt per language For a multi-language site, the cleanest pattern: one canonical llms.txt at `/llms.txt` covering the primary language, plus reference to language variants: ``` # Brand > Primary description. ... ## Language versions - English: https://yourdomain.com/llms.txt (this file) - Ukrainian: https://yourdomain.com/ua/llms.txt - Polish: https://yourdomain.com/pl/llms.txt ``` Each language variant is a complete llms.txt in that language. AI systems can discover and cite either depending on the language of the prompt. This site does it differently — single llms.txt at `/llms.txt` with explicit pointers to UA paths inside it. Either pattern is valid. ## What you should do for a multi-language brand Start with two languages, not five. Pick the secondary language with the largest commercial pipeline (often Polish for Ukrainian and Romanian businesses, German for many EU SaaS, Arabic for MENA-targeting brands). Ship the secondary version of: - All five service pages - Pricing - Methodology - Top three blog posts That is roughly two months of native-editor work and gives you the citation surface in language 2. Expand only after that surface is producing citations. If you want this implemented end-to-end with native editors per language: that is the [Scale tier](/services/aeo-geo-scale-integrated) with explicit multi-language scope. Two languages on Scale, three on Enterprise, four+ on custom. ## See also - [**Free AI Visibility Audit** — 60-second score across 8 categories](/ai-visibility-audit/) - [**What is AEO?** — definitional pillar](/blog/what-is-answer-engine-optimization/) - [**Best AEO tools 2026** — comparison](/blog/best-aeo-tools-2026/) ## Related reading - [Cross-engine prompt research — finding the prompts that actually drive LLM traffic](/blog/cross-engine-prompt-research) - [Refresh cadence — a 4-week operational rhythm for AI citations](/blog/refresh-cadence-rhythm) - [The AEO checklist — 32-point answer engine optimization checklist for 2026](/blog/aeo-checklist) ## Run a free AI visibility audit 60 seconds to submit, full 47-check report in your inbox within 24 hours. No signup wall, no call required. [Get the free audit](/ai-visibility-audit/) [See pricing](/pricing/) Last updated 2026-04-08.