How to Check an AI Translation in a Language You Do Not Speak

This is the objection that stops most multilingual projects, and it is a fair one. You can generate a Polish version of your site in an afternoon. You cannot read a word of it. Publishing it feels like signing a document in the dark.

The instinct is to hire a native speaker to read everything. That works, and it costs more than the translation did. The better answer is that you do not need to verify meaning to catch nearly every failure that matters, because machine translation fails in specific, findable ways.

What actually goes wrong, and what does not

We benchmarked eleven models on real product copy across several languages, twice each, scoring on verifiable criteria rather than impressions. Two results changed how we think about review.

Grammar is not the problem any more. Not one of the eleven produced a sentence a native speaker would call broken. The failures were terminology and register: a product family renamed halfway through, a trademark helpfully localised, a formal register where the brand is informal.

The defects are systematic, not random. Running the same segments twice produced the same mistakes. That is the single most useful fact for a reviewer: you are not sampling a lottery. If a model handles a construction badly, it handles it badly everywhere, so finding it once tells you where all the others are.

It also means a small, well-chosen sample is genuinely informative. Fifteen pages will not catch everything, but the categories of error you find in fifteen pages are the categories present in the other two thousand.

The checks that need no language skills at all

Start here, because these catch the failures that embarrass you publicly, and none of them require understanding a single sentence.

Numbers, units and codes. Prices, dimensions, weights, product references, SKUs, model numbers. Put the source and the translation side by side and compare the digits. A model that turns 250 g into 205 g, or rewrites a product reference, is a problem you can see without reading.

Layout and length. German and Polish run long. If your buttons, menu items or table headers are fixed width, they break. Open the target-language site at mobile width and look for text spilling out of its container, headings wrapping to three lines, or a navigation bar that has silently reflowed.

Untranslated leftovers. Scan for source-language words in the translated pages. A single English button in a Spanish checkout is more damaging to trust than a clumsy sentence, and it is trivial to spot.

Links and formatting. Every link in the source should exist in the translation. This one matters more than it sounds: a translation that flattens a sentence can silently drop the anchor and leave the text without its link. Compare link counts per page.

The functional path. Add to cart, checkout, payment step, confirmation page, confirmation email, invoice. Do it in the target language. You are not reading, you are checking that every step is in one language and that nothing falls back to the source halfway.

That list takes under an hour on a site of any size and it is where most of the real risk lives.

The checks that need a little help

For meaning, you need something more than your own eyes, but far less than a full human review.

Back-translation, used correctly. Translate a sample of the target text back into your own language with a different model, and read that. It will not tell you whether the style is good. It tells you loudly when the meaning has moved: a claim that got stronger, a guarantee that appeared, a negation that vanished. Use it on your legal pages, your guarantees and your top product claims, not on everything.

A second model as a reviewer. Ask a different model to list terminology inconsistencies and register shifts in the translated text, without rewriting it. It is good at exactly the failures we measured: the same term rendered three ways, formality that flips mid-page. Treat the output as a list of suspects, not a verdict.

Your own glossary as a test. If you have a list of terms that must never be translated or must always be translated a certain way, checking it is a search, not a reading exercise. Search the translated site for each term and count. This is the highest ratio of confidence to effort in the whole process.

Where a human is worth paying for

Some review still needs a native speaker. The trick is to buy an hour of it rather than a week, and to spend that hour on the right pages.

Give them your home page, your two best-selling product pages, your delivery and returns policy, and your checkout. Ask three questions, not "is this good":

1. Does this sound like a company you would buy from, or like a translation?
2. Is the level of formality right for this market and this kind of business?
3. Are there terms here that a customer in this market would not use?

That is a one-hour job for a freelancer, it costs less than a dinner, and it catches the register and vocabulary problems that no automated check will surface. Then apply what they tell you to the glossary, and the correction propagates to everything you translate afterwards.

The market itself is a reviewer

Once the site is live, you get feedback that no reviewer can give you.

Search Console tells you which queries your translated pages actually match. If a page ranks for terms that are not the ones you would expect, your translated titles are technically correct and commercially wrong. That is a signal, not a failure.

Bounce rate and add-to-cart rate per language tell you where the copy is not doing its job. A language whose product pages get traffic and no carts usually has a trust problem, and the trust pages, delivery, returns, contact, are almost always the ones nobody reread.

And customer emails are the most honest source you have. Two confused messages about the same sentence are worth more than an afternoon of review.

The pragmatic position

You are not choosing between a perfect translation and a risky one. You are choosing between a machine translation you review sensibly and no version of your site at all in that market.

Check the mechanical things, which need no language. Sample deliberately, because the defects repeat. Buy one hour of native review for the pages that sell. Put what you learn into a glossary so it applies to everything you translate later. Then publish, watch the numbers, and improve the pages the market tells you to improve.

That process is why we built the translation editor around editing published pages in place rather than around a promise of perfection: the first version gets you into the market, and the corrections you make afterwards are the ones that actually matter.