How to run a multilingual chatbot without a translation budget
If you sell to people outside the country your site is written in, your visitors arrive in five or six languages by week two. The "right" thing — translated landing pages per locale, professional translators on retainer — is a budget most SMB and creator-economy teams don't have. There are three pragmatic patterns; only one of them actually works for AI chat in 2026.
Pattern 1 — Translated landing pages per locale
Hire translators, maintain locale folders (/en, /de, /es), translate each content change. This is the gold standard for SEO and for visitors who scroll past the chat widget.
Reality: most teams maintain one locale (their primary), let the rest go stale, and watch the "translated" pages drift months out of sync with the primary. The translated chatbot then has to choose: rely on stale translated content or ground on the fresh primary-language content the visitor does not read.
Pattern 2 — A Google Translate badge
Put a "Translate this page" widget in the corner. Visitor clicks, page gets machine-translated client-side, chat operates on the translated text.
Reality: the translation quality is uneven, brand voice evaporates, and the chat widget itself either ignores the locale switch (so the agent answers in your primary language inside a translated page — disorienting) or runs on the machine-translated content (so the agent grounds on text that doesn't match your real source).
Pattern 3 — Single-source content, agent translates at conversation time
Keep your content in one language (whatever you write it in natively). When a visitor arrives speaking a different language, the agent:
- Detects the visitor's language from their first message
- Retrieves the most relevant passages from your single-source content (in your original language)
- Translates the answer into the visitor's language at the moment of replying
This is what modern AI chat agents actually do. Your content stays single-source — you write it once, in your language, and it serves visitors in their own language.
The catch (one)
Retrieval grounding works best when the question and the content share a language. If your visitor asks a question in Mandarin and the answer lives in English content, the retrieval step (vector similarity) is slightly less accurate than if both were in the same language.
The two mitigations: (a) for high-volume target languages, write some content natively in that language; or (b) look for source-side translation — the agent translates the question into your content's language before retrieving, then answers in the visitor's language. NovixAgent AI does this today whenever your site's language is known and differs from the visitor's.
What the visitor sees
The widget detects the browser language and the message language. Greeting messages are generated automatically in each language you select when you set the site up, so a visitor in one of those languages lands on a localized opening screen rather than an English one inside an otherwise translated chat. The visitor does not have to pick a conversation language — the agent picks for them based on what they wrote.
What you see in your dashboard
Transcripts in your dashboard display in the original conversation language. For English-speaking sales teams reading a German transcript, the captured form fields (email, name, qualifier responses) are language-agnostic and surfaced separately — the German prose is supplementary context. Live chat is the exception: a thread your own operator is handling can be switched into their language from the thread itself, translated in their browser rather than sent to a translation service.
The one thing not to do
Don't ship a chatbot that pretends to speak the visitor's language by maintaining hand-translated canned responses for the top 20 questions, then falls back to "I don't understand" for the rest. The visitor types in Polish, gets a perfectly-translated "Hi, how can I help?", asks an actual question, and gets a wall of English. That's worse than English-only — it's a bait-and-switch.
Auto-detect-and-respond is the only modern pattern that holds up in production. Pick the agent architecture that does it natively, then write your content in the one language you can keep fresh.
See the multilingual use case for our implementation.