Translating a website used to be slow and expensive, so most businesses didn’t do it, or did it once and never updated it. AI has changed the cost. A model will translate every page you have into five languages before lunch.
That is the problem as well as the opportunity. A job that was slow enough to be checked is now fast enough to be published without anyone reading it.
What goes wrong
The translations are mostly fluent. Fluent is not the same as right. The errors I see are the ones that matter most to the business:
- Product and trade names get translated. A brand name becomes an ordinary word. A technical term gets a different translation on every page.
- Your own vocabulary disappears. Every industry has words it uses in a particular way. A general model picks the common meaning, which is often not yours.
- Nobody knows what has been checked. Fifty pages in five languages is 250 versions. Without a record, “the German site” is a mix of reviewed pages, raw machine output and pages that were translated before the English changed.
- It all goes live at once. Somebody installs a plugin, turns it on, and every mistake is published in every language the same afternoon.
None of that is the model’s fault. It was given no rules.
Rules first
The fix is to treat translation as a system, not a button.
A glossary of approved terms. A list of the words that must be translated a particular way, and the ones that must not be translated at all. Product names, technical terms, the phrases your sales team actually uses in that market. The model is given the glossary every time, so the same term comes out the same way on every page.
A status for every page. Editors should be able to see, for each page and each language, whether it is untranslated, machine translated, reviewed or out of date because the original changed.
A test step. The translation runs somewhere it can be read before it is public.
Two separate decisions. Deploying the workflow and switching a language on are different things. The first is technical. The second belongs to whoever answers for what the site says in that market.
I’ve built this inside a client’s CMS for a multilingual industrial site. It works from a glossary of the client’s own approved terms.
The parts outside the translation
Getting the words right is half the job. Search engines also need to know which version is for whom. That means an address for each language and the markup that tells search engines which page is the equivalent of which. Leave that out and the versions compete with each other, or the wrong one turns up in the wrong country.
It is worth checking how AI assistants describe you in each market too. They read your translated pages, and they repeat what they find.
What I’d do
Before translating anything, write the glossary with the people who sell in that market. It is a day’s work and it is the most valuable day in the project.
Then build the workflow where your editors already work, with a status for every page and a test before anything is public. Start with one language and the pages that earn money. Read them properly. Switch it on when someone who speaks the language says it is right.
AI has made translation cheap. It hasn’t made it somebody else’s responsibility.