Guide

Multilingual customer support on your website, done right

Multilingual customer support is more than translation. Learn how one source of truth, language detection and clean handoffs keep answers consistent.

Key takeaways
  • Keep one source of truth for prices, policies and hours, and generate every language's answer from it, so no translated page can drift out of date.
  • Route handoffs by language skill, and give the team a short summary in their own language rather than the visitor's original wording.
  • Have a native speaker test 20 real questions in each language before launch, paying special attention to prices, fees and dates.

Multilingual customer support often starts as a translation project, and that is exactly where the trouble begins. A visitor lands on your site in Spanish, asks what your cancellation fee is, and receives a fluent, confident answer. Whether that answer is correct depends on which page it came from. If the Spanish version was translated two years ago, before the fee changed, the visitor is now reading a plausible sentence that is wrong, written in their language and published under your name.

Translation gives you pages in more languages. It does not give you consistency across them, and consistency is what visitors actually need. This guide covers how to keep facts aligned across languages, when to translate by hand and when a machine is enough, and how to handle the details that quietly differ from one market to the next.

Why translation alone creates contradictions

A translated page is a snapshot. It is accurate on the day it is finished and starts drifting the moment someone edits the original. Prices change, opening hours shift for the summer, a service is retired. In most small businesses, the English page gets updated because someone notices. The French, German and Spanish pages get updated when someone remembers, which is rarely.

The result is a predictable pattern. Take a clinic that publishes its cancellation policy in three languages. The English page, updated last month, says 24 hours’ notice and a £20 fee. The Spanish page, translated the previous year, says 48 hours and no fee. The Polish page, copied from an old PDF, says 24 hours but gives an earlier fee of £15. A visitor reading any one page gets a coherent answer. A visitor who compares pages, or who asks the chat and then checks the site, sees that the business contradicts itself. That erodes trust faster than a slow reply ever does.

The root cause is that facts were stored as sentences in each language, rather than as facts that get expressed in each language. Once you see that distinction, the solution becomes clearer.

Use one source of truth for facts

The most reliable setup keeps the facts in one place, in one language, and generates answers in other languages from that place. Your fees, hours, service list and policies live in a single canonical source. Every language reads from it. When the fee changes, you change it once, and every language reflects it the next time someone asks.

Businesses tend to take one of three approaches, and they differ in effort and in risk.

Approach Upkeep effort Consistency risk Best suited to
Separate translated pages for each language High: every edit happens in every language High: pages drift apart A few stable pages, such as an about page
Translated FAQ alongside the main site Medium: FAQ kept in sync by hand Medium: the FAQ is often right while the main pages are not Small catalogues that change rarely
One source of truth, answers produced per language Low: edit the source, not each language Low, as long as the source is clean Sites where prices, hours or availability change often

The third option sounds like a purely technical feature, and in part it is one. But the real work is editorial. You still have to write the source clearly, remove contradictions inside it and keep it current. A chatbot that answers in nine languages from one bad source will simply give nine wrong answers instead of one. For more on why the assistant should answer only from your own content, see how to keep an AI assistant from making things up.

Detect the language, then handle the messy cases

Language detection is usually the easy part. A visitor’s first message is often enough, and browser settings give a useful hint. The harder cases are the ones real visitors produce.

Short messages are a challenge. “Cuánto?” is clearly Spanish, but “Hours?” could come from anyone. Mixed-language messages are common in markets where people move between languages in ordinary speech, and a visitor may write in English and quote a product name in their own language. Some visitors start in one language and switch when they realise the answer is clearer in another.

The sensible rule is to answer in the language of the most recent clear message, and to offer a switch when the signal is weak. “I can answer in English or Spanish. Which would you prefer?” costs one exchange and avoids a long conversation in the wrong language. Right-to-left languages such as Arabic and Hebrew also need the chat layout and the site to flip correctly, which is a design question as much as a language question.

Names and addresses need their own care. A visitor’s name may contain accents or characters your system doesn’t expect. Keep the original spelling in the record and make sure your team sees it exactly as typed. A handoff that mangles someone’s name is an unforced error that a visitor will remember.

If you are only just starting, don’t try to launch every language at once. Look at which languages your visitors already use, using your analytics and the language of the first message in your chat or enquiry logs, and start with the two or three that account for most of your real traffic. Each additional language brings its own glossary, its own review and its own handoff routing. Two languages done well are worth more than nine done carelessly, and you can add the rest once the process is working.

Translate some things by hand, and say why

Not every piece of text carries the same risk. Opening hours are low-risk, because a wrong translation is noticed quickly and is easy to correct. A cancellation fee, a refund policy or a safety instruction is high-risk, because a plausible but wrong sentence can cost money or cause harm.

A practical rule is to match the level of human review to the consequence of an error. The table is a starting point, and your own list should reflect your industry and your regulators.

Content type Machine translation acceptable? Review rule
Opening hours and location Yes Spot-check once after each change
General service descriptions Yes, with a glossary Native speaker reads a sample
Prices, fees and deposits Only with review Native speaker signs off every figure
Refund, cancellation and warranty terms No, not without legal review Written or checked by a native speaker against the source
Privacy and consent wording No, not without legal review Follow your legal advice for each market
Brand and product names Keep the original Never translate
Health or safety instructions No Qualified native speaker review

Small words do a lot of work in policy text. “May” and “will”, or “up to” and “at least”, can flip the meaning of a sentence, and machine translation handles those distinctions unreliably. This is why the review rule for prices and policies is a human sign-off rather than a spot check.

Fluent is not the same as correct: machine-translated fees and policies read smoothly, which is exactly why their errors go unnoticed. Have a native speaker sign off every price and every policy sentence before it reaches a visitor, in every language you serve.

Multilingual customer support needs clean handoffs

A handoff in a second language has three extra requirements. The visitor should reach someone who can speak their language, or a clear fallback. The team member should receive a summary they can read in seconds. And the visitor should never have to explain their situation twice.

If you have staff who speak several languages, route by language first and by topic second. If you don’t, be honest about it. “Our team replies in English within one working day. If you’d prefer a reply in Spanish, I can take a callback number and note that for whoever calls you” is far better than a handoff that promises a Spanish reply and delivers an English one.

The summary matters too. A visitor’s message written in French should reach your English-speaking team as a one-sentence English summary, with the original text attached. The team can then reply in the visitor’s language if they can, or use a translation tool carefully if they can’t, and in either case they know what was actually asked.

Dates, money and local expectations

Some of the most common multilingual errors are not about language at all. They are about formats that differ quietly between markets.

A date written as 03/04/2026 is the fourth of March in some countries and the third of April in others. Write dates in full, with the month name, in every language. “Thursday 15 October 2026” is unambiguous everywhere. Money needs its currency and the decimal convention for the market, because “1,250” can mean one thousand two hundred and fifty in one place and one and a quarter in another. Phone numbers and addresses follow local conventions too, and a visitor will notice when yours don’t look like theirs.

Register is subtler. Many languages distinguish a formal and an informal “you”. Decide which your brand uses in each market and keep it consistent. A bank and a neighbourhood bakery often make different choices, and both can be right for their customers. What matters is that the chat doesn’t switch between them within one conversation.

Test each language before you trust it

Testing multilingual support is slower than testing English, and it is where most businesses cut corners. A fair test has three parts.

First, collect twenty real questions per language. Use transcripts, emails and phone notes rather than inventing them, because real visitors phrase things in ways you won’t predict. Second, have a native speaker check each answer for accuracy against your source, for natural phrasing and for register. Third, test the handoff. Send a question in each language that requires a person, and check that the routing, the summary and the promised reply time all make sense in that language.

Repeat the test after any change to the source content. A fee change that is correct in English and stale in Spanish is exactly the kind of error testing is meant to catch.

Keep a glossary of fixed terms: list the 20 to 30 words that must never vary between languages, such as your service names, policy titles and product names. Give each one a single approved translation, and check new content against the list.

Why it’s worth the effort

The business case is straightforward. In its 2014 study “Can’t Read, Won’t Buy”, the research firm Common Sense Advisory reported that roughly seven in ten consumers said they were more likely to buy a product when the information was in their own language. That study is a decade old, and the exact figure will have shifted with the market, but the underlying point has held up. A visitor who can read your answers in their own language is more likely to stay, act and trust what they read.

That trust is fragile in both directions. A correct answer in the wrong language is a barrier, and a fluent answer in the right language that happens to be wrong is worse. Consistency is what makes the investment pay off. If you are also looking at how quick, clear answers affect the rest of the funnel, our piece on website conversion with instant answers looks at that side of the picture.

One sourceof truthENESFRDEPLEach answer is generated from the same facts
One source of truth feeds every language. Change a fee once, and every answer reflects it.

What to take away

Multilingual customer support is a consistency problem before it is a translation problem. Store the facts once and write them clearly. Match review to risk, with native speakers signing off prices and policies. Route handoffs by language, and test every language with real questions. Do those four things and additional languages stop being a liability and start being a service.

Voxy answers in nine languages from your own content, so the fee on your page is the fee the visitor hears in Spanish, French or Polish. It also brings in your team when a person is needed. You can see how the assistant handles questions on the features page, or start a free 14-day trial with no credit card and test it on the languages your customers actually use.