Why Chatbots Reply in the Wrong Language (And How to Fix It)
OS
Oskar
·5 min read
Your website is in German, a visitor types a question in English, and the chatbot answers in German anyway, or it answers in English but gets the technical term wrong, or it starts a reply in German and finishes the sentence in English. This isn't a rare glitch. Someone posted the exact symptom, word for word, on the OpenAI developer forum: "my RAG chatbot does not reply in same language as query." It's a known failure mode with a specific, fixable cause.
For a DACH business, this isn't a cosmetic bug. A Swiss company with German and French customers, a Munich retailer with English-speaking tourists, or a B2B supplier fielding a Polish subcontractor's question all run into it the moment their audience isn't monolingual. Here's why it happens, and the specific setup that fixes it.
Frequently asked questions
Related Posts
Why AI chatbots answer in the wrong language
A RAG chatbot does two separate jobs, and only one of them is naturally multilingual. The first job is retrieval: turning the visitor's question into a vector and finding the closest matching chunks in your uploaded documents. Modern embedding models are trained on parallel text across dozens of languages, so "delivery time" and "Lieferzeit" usually land close together in vector space. Retrieval across languages generally works better than people expect.
The second job is generation: writing the actual reply. This is where things break. The model has the right German source chunk in front of it, but nothing forces it to answer in the visitor's language rather than the source document's language, or rather than whatever language dominates its own training weights (usually English). Without an explicit instruction, it guesses, and it guesses inconsistently between one message and the next.
The translation-layer trap
Some chatbot builders try to solve this by bolting on a translation step: detect the visitor's language, machine-translate the question into the document's language, search, then translate the answer back. It sounds reasonable and it's a common approach in off-the-shelf widgets. It's also where a second failure mode shows up.
A translation round-trip loses precision on exactly the terms that matter most: product names, technical vocabulary, and industry jargon. "Wärmepumpe" translates cleanly to "heat pump," but "Bandbreite" can mean network bandwidth or a range of values depending on context, and a naive translation layer picks one and moves on. Every extra translation step is another place for the original meaning to drift, and pricing or spec numbers are the last thing you want drifting.
The alternative is a multilingual embedding model that searches your original documents directly in whatever language they're written in, with no translation step at all. The visitor's question gets embedded in its own language, compared against your German source content in vector space, and the matching chunk is retrieved without ever being rewritten.
What this costs a DACH business in practice
Picture a Zurich-based online shop with German product pages and a mostly German-speaking customer base, but a meaningful share of French-speaking visitors from Geneva and Lausanne. If those visitors type a question in French and the bot replies in German, or replies in French but with a mistranslated spec, most of them won't correct it or switch languages themselves. They'll assume the shop doesn't serve them and leave. That's a lost sale on the very first message, not a minor annoyance.
For a Handwerk business with mostly German-speaking customers, the risk looks different: an occasional English-speaking expat or tourist asks about a service, gets an answer that's either in the wrong language or garbles a technical term like "Durchlauferhitzer," and the business never finds out, because nobody reports a bad chatbot conversation. It just quietly leaves.
A note on Swiss German and regional dialect
There's a related wrinkle specific to Switzerland. Website content is almost always written in standard Hochdeutsch, but customers typing into a chat box often write in Swiss German dialect, or in a casual mix of dialect spelling and Hochdeutsch. A chatbot's retrieval step still needs to make sense of that informal input against formally written source documents. This is less about picking the right output language and more about the model tolerating messy, dialect-influenced input without misreading the question entirely. It's a smaller problem than the cross-language one, but it compounds it for Swiss businesses specifically.
How to check if your chatbot has this problem
You can test this yourself in about five minutes, no technical background required.
Ask the same factual question twice, once in German and once in English. Compare the content, not just the language. If the two answers give different prices, specs, or availability, your retrieval step is broken, not just the language output.
Check whether the reply language matches the question language. Ask in French or Italian if you serve Swiss customers. If the bot answers in German regardless of what language you typed, the generation step has no language instruction.
Test your actual industry vocabulary, not generic phrases. "Wärmepumpe," "Frist," "Bandbreite," whatever terms your business lives on. Generic small talk translates fine in almost any chatbot; the failures show up in domain-specific words.
Type a question in casual, misspelled, or dialect-influenced language. Real customers don't type formal Hochdeutsch into a chat box. If the bot only understands textbook phrasing, it will misfire on real traffic.
Fixing it: what a properly built multilingual chatbot needs
Two things have to be true at once, and most off-the-shelf chatbot builders only get one of them right.
First, retrieval has to use a genuinely multilingual embedding model, so a German document and an English or French question land close together in vector space without a translation step in between. Second, the system prompt has to explicitly lock the reply language to the visitor's input language, not the source document's language and not a default. This is exactly what AmueAI does: German source content matches German, English, or French questions directly, and the widget's built-in test playground lets you check answers in multiple languages before the chatbot ever talks to a real visitor.
If you're running a chatbot today, don't take a vendor's "multilingual support" claim at face value. Open the widget, ask it your five most common customer questions in every language your visitors actually use, and read the answers for accuracy, not just for whether they came back in the right language. That test takes less time than writing this article did, and it catches the exact failure that quietly costs DACH businesses customers they never see leave.
AI Search VisibilityAI Agents
Why ChatGPT Doesn't Recommend Your Handwerk Business
Only about 1 in 100 local businesses gets actively recommended by ChatGPT, and a 4.8-star rating won't change that. Here's what actually earns a Handwerk business a mention in AI search, and why it overlaps with what a RAG chatbot needs too.
