Translating your website with AI: automatic or reviewed?

Translating your website with AI quickly solves the problem of turnaround time; what remains is the question of review. Having a person review everything is expensive on a catalogue of several hundred pages, and reviewing nothing lets errors through in languages that nobody in the company reads. The right split is decided page by page, depending on what each page commits the company to. This comparison goes through it in detail; it extends our articles on website translation and opening new markets.

What does Google ask of a website translated with AI?

For a text produced by an AI model, the reference is Google’s guidance on using generative AI content on your website. In its English version, updated on 1 October 2026, it points out that a generative model does not retrieve facts: it predicts a likely sequence of words, and can therefore be wrong. On review, its two language versions say the same thing. The English version puts it this way:

The page extends this check to metadata: the title tag, the meta description, structured data and the alt text of images, which can appear in the results.

In our reading, a translation produced by an AI model falls into this case: the text published in the target language is written by the model, even if the original is not. Two consequences for a translated website.

Metadata is reviewed like the text. The title and description of a translated page are what the reader sees first, in the results in their language. An error there costs the click before the page has even been read.

Volume does not protect you. The same page points out that generating many pages without adding anything for readers can break Google’s rules on scaled content abuse. Translating your pages for readers who speak the language gives them something; opening languages that nobody targets, just to multiply addresses, comes close to it. We go into what separates a useful catalogue from scaled content in a separate article.

Which pages should you have reviewed, and which can you validate on a sample?

A review takes the time of someone who reads the language. That time goes where an error costs the most:

The catalogue entries row is the only one where not every page is reviewed. The template, though, is reviewed in full and in each language, because it is the template that repeats an error across all the entries. For the pages themselves, our method provides, on the first pass, for a sample reviewed in each language by an AI judge (a second model, more expensive than the one that translates, which reads only the samples) and by a person, then spot checks. The compromise is real.

Two safeguards frame the sample. Note, in each sample, the share of segments corrected, and review more if it rises. And if a language is not worth the cost of a review, ask yourself whether it is worth opening.

How do you check an AI translation, step by step?

The check comes down to six steps:

1. 01Write the glossary before the first pass.

   Note the brand names, industry terms and tones to avoid. The translation memory then keeps every approved sentence: see what a translation memory brings to a company.

2. 02Fix the addresses of each language before the catalogue grows.

   On Palmora Property, a property listings site in six languages that we operate, articles in English carry French addresses: a debt that is no longer worth correcting.

3. 03Translate by segments, with a translation memory.

   At each update, only new or modified segments go back for translation, and a correction made in the memory applies to every page that reuses the sentence.

4. 04Set up the checks.

   Every content key must exist in every language (the parity check), no page must announce a version that does not exist, and the facts, sources and names cited must be checked on all generated content. These checks avoid the costliest hreflang mistakes on a multilingual site.

5. 05On the first pass, review each template in full, then a sample of pages per language.

   Titles and descriptions are part of the review. After that, monitor with spot checks.

6. 06Have the pages that commit the company reviewed in full.

   An offer, terms or a page that sells are reviewed before publication, then at every change.

Where to start

Start with one language, that of your first market, and check that there is demand before translating everything: a few pages that answer the searches you have already spotted, with a threshold written in advance, are enough for a first test. We go through this test in International SEO: choose your markets before translating. A small campaign in the language gives an answer even faster: see how to choose between Google Ads and Meta for a B2B SME. And translation never makes up for a structural disadvantage: if your offer already loses against the players in the country, a better translation will change nothing.

To decide which pages to have reviewed, and which language to start with, tell us about your website and the target markets: we start from the pages that sell.

Frequently asked questions

Sources

1. Google’s guidance on using generative AI content on your website (new tab)developers.google.com
2. Google’s Keyword Planner (new tab)support.google.com