Support every language your customers write in

You do not have a Spanish help centre, a German one, and a Japanese one. You have one, and customers who do not read it in English.

Start free Read the setup guide

No credit card. No second help centre.

ES¿Hacen envíos a España?

FRPuis-je modifier ma commande ?

DEauto-detected

Wie lange dauert der Versand nach Österreich?

Standardversand nach Österreich dauert 3–5 Werktage. Ab 50 € ist der Versand kostenlos.

Shipping & delivery

Answers in their language: Detected from the question, not from a picker nobody finds.
Content maintained once: No second help centre to keep in sync.
Citations still work: The link points at your real source page.
Refuses in every language: The honesty does not degrade outside English.

The counter of a small family-run shop after closing, warm bulb light meeting cold blue daylight from the shopfront.

Going international multiplies the support surface

Every new market is another language of inbound and, in theory, another translated help centre to maintain. In practice the translation lags, so international customers get a worse answer than domestic ones — or they get English and are expected to cope.

Support quality drops noticeably outside your primary language
Translated help articles fall behind the originals after the first quarter
You cannot hire a native speaker for every market you sell into
Non-English questions sit in the queue longest

One knowledge base, every market

The usual approach to going international is a translated help centre per market, which falls behind the original within two quarters. Answering in the language of the question from a single maintained source removes the sync problem entirely.

Detected, not selected

The language comes from what the visitor wrote. Language pickers are consistently ignored.

Written once

Your source pages stay in one language; the answers do not have to.

Verifiable everywhere

The citation points at the page the answer came from, so it stays checkable in any language.

The same refusal behaviour

Which matters most here — fabrication is hardest to catch in a language nobody on your team reads.

Timezones you do not staff

International markets ask at their evening, which is the middle of your night.

Demand by market

Conversations are logged as asked, so you can see which languages generate real volume before you localise.

Written in one language, answered in theirs

The bot understands the question in the language it arrives in and answers in that language, drawing on the content you maintain in one place. There is no second help centre to keep in sync, because there is no second help centre.

Reply in the asker's language: Whatever they type in, the answer comes back in.
Maintain content once: Nothing forks. One page to edit when the policy changes, whatever the reader speaks.
Citations still work: A reader can open the source and check it, even when it is not in their language.

Widget voice settings with a header title, a greeting, and a language dropdown set to follow the visitor's browser, noted as governing the interface only. — Auto is almost always the right choice here — it only decides what the chrome says.

The same honesty in every language

Refusal behaviour does not change with the language. A question your content does not cover gets a plain 'I do not know' and a route to a human in that language, rather than a fluent invention — which is a particular risk when nobody on your team can read the reply to check it.

See demand by market

Because questions and refusals are logged as asked, you can see which markets are generating volume and which questions they are asking. That is usually a better signal of where to localise properly than a forecast.

The honest part

Fluent is not the same as correct

A model will happily invent a policy in flawless German, and the failure is far harder to catch than it would be in English, because nobody on your team is reading it. So the safeguard has to be the same rule everywhere: a passage in your content, or a plain refusal.

The reply follows the language of the question, not a picker
The refusal follows it too — no confident invention outside English
Citations point at the source page you actually maintain
Conversations are stored in full, so a speaker can sample them

Precision is still better in the languages you publish in. Translating your source pages improves the answers in that language; it is just not a prerequisite for answering at all.

A desk at night lit only by an off-screen monitor: keyboard, notepad, folded reading glasses and a last inch of tea.

No matching passage

ESNo tengo esa información en la documentación.
DEDazu habe ich keine Information in der Dokumentation.
FRJe n’ai pas cette information dans la documentation.

Same rule, every language

Setting it up, and checking it honestly

There is almost no configuration. The work is in the verification, and specifically in the step most teams skip.

1
Index what you already maintain
In whatever language you maintain it. There is no per-language setup, because answers are generated from that single source.
2
Leave the language on auto
Widget language sets the menus and buttons. Answers follow the question regardless, so auto is usually correct.
3
Test the boundary, not the greeting
Ask something outside your content in each language you expect and confirm you get a refusal, not a fluent invention.
4
Have a speaker sample the logs
Conversations are stored in full. A periodic review by a native speaker is the practical quality control.

The chat widget open on a web page with a greeting and suggested questions, ready to answer a visitor. — The same widget, whatever language the visitor opens it in.

Before you sell into another market

Short answers. If yours is not here, the assistant on this page will try it — and tell you honestly if it cannot.

Do I need to translate my documentation first?

No. The bot works from the content you have and answers in the language of the question. Translating your source material improves precision in that language, but it is not a prerequisite for answering at all.

How do I check the quality of answers in a language nobody on my team reads?

Conversations are logged in full, so they can be reviewed or sampled by a native speaker. The refusal behaviour matters most here: because the bot declines rather than improvises, the failure mode in an unreviewed language is silence rather than confident fabrication.

Does the citation link to a translated page?

It links to the source page the answer came from, which is the page you actually maintain. If you publish a translated version of that page and it is indexed, that becomes citable too.

All solutions