You can check a chatbot's answers in a language you don't speak with two layers of review. First, translate each answer back into your own language and compare the facts with your source page; that catches wrong numbers, dropped conditions and invented details. Second, ask a native speaker to read a sample for tone and natural wording, which back-translation cannot judge. Neither layer needs much time, and together they cover the mistakes that matter.
Our post Can a Chatbot Answer in Spanish From English Pages? explains how a chatbot answers in another language and where it gets weaker. This one is the checking routine itself, for any language.
Two kinds of error, two kinds of check
A wrong answer in another language is usually one of two things.
A fact error. The answer says 14 days when your page says 30, leaves out "unless faulty", or names a service you don't offer. These matter most, and you can catch them without speaking the language.
A wording error. The facts are right, but the reply sounds stiff, uses an overly formal or overly casual form of address, picks a term that means something slightly different locally, or reads like a translation. These matter for trust, and you need a native speaker to see them.
Back-translation handles the first. A native reviewer handles the second. Translation guidance in health research, such as the ISPOR task force principles for translating patient questionnaires, uses back-translation as one step and pairs it with testing by native speakers, because a translation can come back word-perfect and still read oddly to the person it was written for. The same logic applies to a chatbot, on a much smaller budget.
Layer one: back-translation spot checks
Back-translation means taking the bot's answer and translating it into your language, then reading what it says. You can do this yourself in a few minutes per answer.
- Write questions in the target language. Use real messages from customers who wrote in that language if you have them. If not, ask a speaker to write your ten most common questions the way a customer would type them. Avoid running your English questions through a translator, because machine-translated questions are cleaner than real ones and make the bot look better than it is.
- Ask the bot. The preview link in your dashboard opens the chat on its own page, which is the quickest place to test.
- Translate the answer back with a machine translation tool.
- Compare against the source page, not against what you hoped it would say.
- Mark each answer as correct, incomplete, wrong or invented.
When comparing, check these specifically, because they are where meaning shifts:
- Numbers and units. Prices, time limits, sizes, weights, distances.
- Currency. The bot should keep your currency, not convert it.
- Dates and times. 3/4 is the 3rd of April in some countries and March 4th in others. 24-hour and 12-hour times can be confused.
- Conditions and exceptions. "Within 30 days, if unused" must not lose "if unused".
- Negations. A dropped "not" reverses an answer. Read every sentence about what you do not do.
- Names. Product, plan and place names should stay as they are.
Back-translation has a blind spot of its own: the tool translating back may smooth over an awkward reply, so the English reads better than the original. It tells you the facts are right, not that the wording is good.
Layer two: a native speaker review
For each language you get real traffic in, find one fluent reader for an hour: a staff member, a trusted customer or a paid freelancer. Give them:
- the questions and the bot's answers, side by side;
- the source pages in your language, so they can check facts too;
- three questions to answer for each reply: Is it accurate? Does it sound natural? Would you trust a business that wrote this?
Ask specifically about form of address. Many languages have a formal and an informal "you", and the right choice depends on your business. If the reviewer says the bot is too formal or too casual, add an instruction to the bot's persona field, for example "In German, address visitors with Sie."
Tell the reviewer what is fixed. In SpideyChat the replies follow the visitor's language, but built-in labels such as Talk to a person and Report an issue, and the support ticket confirmations, are English. Your welcome message and suggested questions appear exactly as you wrote them. The dashboard is English too.
The post on the role of human review in AI chatbot quality covers review habits more widely.
Keep checking after launch
One round of review is a snapshot. Content changes, and new questions arrive.
Read transcripts. Conversation transcripts in the dashboard show exactly what was said in the visitor's language. Once a month, back-translate a handful of non-English conversations.
Watch the knowledge gaps list, with one caveat. SpideyChat logs a question when nothing in your content matched it closely, and when a visitor marks a reply unhelpful, in any language. It also flags replies that admit not knowing, but that check looks for English phrasing, so a reply in Polish saying it has no information may not appear. Transcripts fill that gap.
Save language-proof test questions. The bot's test suite saves a question with text the answer must contain and re-runs it after any knowledge change. For non-English questions, make the expected text something that survives translation: a number, a price, a product name. "30" or "£45" passes in any language; an English phrase never will.
Fixing what you find
Most problems trace back to the source content, not the language. Add Q&A pairs written in the target language for questions that keep going wrong, state currency and units plainly on your pages, and name local terms your customers use. The training docs show each way to add knowledge, and our guide to how multilingual chatbots answer in your customer's language covers the background.
Run your first check today
Paste your address into the free trial builder. It reads up to ten of your public pages and gives you an assistant you can question in any language. Ask it five real questions in the language your customers use, back-translate the answers, and you will know within half an hour where your content needs work.