Going multilingual in the AI era isn’t translation

You want your site in English. You want to explain your company accurately to customers and partners overseas. So you run the Japanese pages through machine translation, and the English that comes back does make sense. Yet when an English native reads it, something feels off. The grammar is correct, and still you can tell at a glance it was written by a non-native. In B2B, that faint dissonance quietly erodes trust.

Why even accurate translation reads as unnatural

Japanese and English differ in word order, in rhetoric, and in what they assume of the reader. Japanese drops the subject, saves the conclusion for last, and entrusts context to the space between the lines. English names the subject, leads with the conclusion, and spells things out. Ignore that gap and swap sentence for sentence, and you get prose that is grammatically fine but sits awkwardly as English.

Translate a subjectless passive as-is, and you get an English sentence where who does what is unclear. Carry a chain of Japanese abstract nouns straight across, and you get a stiff, verbless construction. A metaphor that feels natural in Japanese becomes a baffling literalism in English. So-called Japanglish is not a matter of translation accuracy — it is a matter of approach.

Don’t translate — write from scratch

The industry’s answer to this is transcreation. Rather than translating the source, you write fresh copy in the target language from a brief that captures the intent, the reader, and the tone. The person who does it is a copywriter, not a translator — that is the standard view. Put another way: the Japanese page is not a source text to translate, but reference material for writing the English.

Search engines favor this too. Google treats only mechanically duplicated, untranslated pages as duplicate content. A page written from intent in English is not seen as a duplicate; it is evaluated as a page in its own right.

AI is what makes it practical

Writing from scratch used to be expensive. It required a bilingual copywriter, which put it out of reach for a small company.

With AI in the loop, you can run this process without the human back-and-forth of copying and pasting. Extract the intent from the Japanese into a brief. Write the English from that brief. Have several AIs, each with a different native perspective, review it in parallel to surface literal-translation smells and awkward collocations. Then verify factual consistency on a separate line of AI. Because AI can run those quality checks internally, writing from scratch has become a realistic option for a small or mid-sized company.

What we learned doing it on our own site

Nihonbashi AI Lab took this approach when we put our own site into English. We reviewed the English in parallel through three native perspectives — a US executive, an English-language editor, and a Japanese-to-English localization specialist — and treated the spots that several of them independently flagged as the ones that genuinely needed fixing. Dissonance a single reviewer would miss surfaces when several judgments converge.

What turned out to be hard, though, was not the first pass into English — it was keeping it running. Once something is in English, if you update the Japanese and leave the English behind, the mapping between the languages (hreflang) breaks, and the search evaluation degrades with it. So we made the discipline structural: when we fix a core Japanese page, we fix the English through the same flow. If the English version is missing, the build itself stops — which prevents forgotten updates by design.

Going multilingual is not something that ends with running text through a translator. You design what to say (business design), decide how Japanese and English map to each other (data structure), run the writing and the quality checks through AI (AI use), and hold a mechanism that keeps the English in sync on every update (operations). Only when these four are in place does a multilingual site that reads naturally to natives begin to run as an asset, rather than a side task.

If you want to run it yourself (the skeleton)

This method needs no special tooling. You have an AI agent follow these steps:

• Turn intent into a brief — write out the purpose, the reader, the claims to keep, and the tone (you write from this, not by translating the source)
• Write in the target language — compose fresh, at native quality, from the brief
• Review through several perspectives — have multiple native viewpoints review in parallel, and fix what more than one of them flags
• Check consistency on a separate track — apart from the prose, confirm with a different eye that the links and rendering are not broken
• Make the follow-through structural — “fix the source, fix the translation,” enforced by the build rather than by willpower

We have written this up as a skill you can copy and use directly (for Claude Code), with the implementation steps and pseudocode, in a Zenn article.