The language barrier in e-commerce is mostly gone now, and its removal exposed a second barrier standing behind it.
For decades, answering a Dutch shopper’s question at 22:00 required a Dutch speaker awake and on shift. Modern assistants have made the linguistic half of that nearly free: detect the language, reply in it, switch when the shopper switches. When I demo this, it is the moment merchants lean in.
It is also where I slow the demo down, because fluent language in the wrong context is a new way to lose a customer, delivered more convincingly than ever.
Why language matters more than teams assume
Most international shoppers read English adequately. They hesitate anyway, and the hesitation is not about comprehension. A purchase is a small bet, and nobody likes placing bets in a language where they might be missing a nuance in the returns policy. It is confidence, not vocabulary.
The scale of that effect is measured: CSA Research’s survey of 8,709 consumers across 29 countries found 76 percent prefer buying with product information in their own language, and 40 percent will not buy from a website in another language at all.
So stores translate: storefront pages, checkout, emails. And then the translated storefront’s answering layer, the part that handles questions, stays monolingual and office-hours. The shopper can browse in Dutch but must hesitate in English. That asymmetry is exactly where an assistant earns its place.
The mechanism merchants don’t see
A shopper types “iets warms voor herfstkamperen” (something warm for autumn camping). The catalog is written in French. The right sleeping bag comes back anyway.
That small miracle is modern embedding retrieval: text from different languages lands in a shared space, so a Dutch question can sit next to a French product description because the meanings align even when no words do. It is quietly remarkable, and quietly fragile. Cross-lingual retrieval quality is an active research problem, not a solved one; the MIRACL benchmark spans 18 languages because performance varies meaningfully by language rather than arriving for free everywhere. High-resource European languages behave well. The further you get from them, the less you should assume.
The practical translation of that research for a merchant: treat cross-lingual matching as the safety net, not the plan. It will rescue a Dutch question against a French catalog. A properly translated catalog still retrieves better, because the product text is the retrieval surface, in every language you sell in.
And a machine-translated catalog is the worst of both: the assistant will answer from whatever the text says, so wrong units and untranslated size charts become wrong answers delivered fluently. Bad data does not stay quiet in an assistant the way it does on a static page. It gets spoken.
Why “just translate the assistant” only partially works
Here is the failure mode I warn every multi-market merchant about, because it is the one fluency invites:
Shopper in Milan: Quanto costa la spedizione, e in quanto arriva?
Assistant, in flawless Italian: Shipping is 4.90 and arrives in two days!
The price is the German market’s. The two days are from the German warehouse. The product does not ship to Italy at all.
The checkout, ten minutes later, breaks all three promises at once.
Linguistically perfect. Commercially, the assistant has misled a customer with total confidence, and the fluency makes it worse: the shopper has no signal that anything is off until the checkout contradicts the conversation.
Markets differ in what you may sell, what you charge, what you promise for delivery, which warehouse has stock, and which storefront the links should open. None of that is language. All of it is context. An assistant that switches language without switching context is a well-dressed liar.
The requirement worth writing down: for each market, the assistant answers from that market’s catalog, prices, and policies. The language of the reply is almost a cosmetic layer on top of that.
The structure that holds: one brain, market overlays
The tempting architecture is one assistant per country. It fails slowly: every improvement to how the assistant sells must now be copied ten times, which within a quarter means drift and within a year means ten mediocre salespeople with different personalities.
The structure that has held up in practice is one assistant, per-market overlays. One shared brief defines how it sells, when it hands off, what it never promises; improve it Monday and every market improves Monday. Each market overlays its own catalog or search backend, its greeting and starter prompts in the local language, and its own policy lines, with a default overlay catching markets you have not localized yet. The storefront declares its market, and everything follows from that declaration. Probedots is built around exactly this shape (we call them locale configurations), not because it is elegant but because we watched the per-country alternative decay.
The limit: overlays solve consistency, not culture. Whether the assistant’s tone lands as helpful or as oddly formal in Japanese is not something any architecture fixes.
Launching a market whose language you don’t speak
This stops most merchants, and it mostly shouldn’t, because trust-destroying errors are fact errors, and facts are checkable without fluency.
Ask the assistant checkable questions in the target language through a translation tool: this product’s price, delivery to Rotterdam, the returns window. Verify every number, currency symbol, and link against that market’s storefront. Then have one native speaker read ten transcripts before launch, not for grammar but for whether it sounds like a helpful person in that culture or a translated press release.
After launch, watch metrics per market, never blended: a market with a broken catalog connection hides comfortably inside a healthy average, and ratings and clicks are language-blind, so an oddly low market almost always has a findable data problem rather than a language problem.
The shift
Merchants file multilingual support under translation, a content problem, something you buy by the word. The assistant era inverts that. The words are the cheap layer now. What a shopper in another country actually needs from you is the same thing every shopper needs, delivered in the only language they fully trust: correct facts about the products you will actually sell them, under the promises you will actually keep. Language was never really the product. Confidence was.