Blog / AI & search /
AI translation for multilingual sites: good enough?
A store owner wants a French version of the site by Friday. Two hundred product pages, a dozen collections, a blog. A human translator quotes weeks and a figure that makes everyone go quiet. ChatGPT, DeepL or the translation app in the store’s admin quotes an afternoon.
So the real question isn’t whether AI can translate. It obviously can, and much better than the machine translation of a few years ago. The question is whether AI translation is good enough to publish and expect to rank in another language.
My answer: it’s good enough for a first draft, rarely good enough to publish untouched, and the gap between those two is where most of the SEO value lives.
What Google actually says
Google doesn’t ban machine translation. What it objects to is translation published at scale with nobody checking it.
The spam policies list “text translated by an automated tool without human review or curation before publishing” as an example of the kind of thing that crosses the line. Since the March 2024 update, that sits under the scaled content abuse policy, which is about producing lots of pages mainly to manipulate rankings, whatever the method.
Read that carefully and the line isn’t “AI or human”. It’s “reviewed or not”. A translated page that a fluent person has read, corrected and adapted is just a page in another language. Two hundred pages pushed live straight from an API, with nobody who speaks the language looking at them, is the pattern the policy describes.
I covered where that line sits for AI content in general in scaled content abuse and AI: where the line is.
Where AI translation holds up
It’s genuinely useful for:
- Structured, factual text. Specs, dimensions, materials, care instructions. There’s not much room for tone to go wrong.
- Getting the bulk done. A reviewer correcting a decent draft works far faster than a translator starting from nothing.
- Consistency. Give it a glossary (“always translate ‘collection’ as X, never translate the brand name”) and it sticks to it better than a tired human on page 150.
- Checking a human translation. Asking a model to back-translate a page into English is a quick way to spot sentences that drifted.
Where it quietly fails
The failures aren’t usually wrong words. They’re right words that nobody in that market would use.
Keywords don’t translate
This is the big one for SEO. Translate “trainers” into American English and you get “sneakers”, fine. But translate your English keyword into another language word for word and you often get a phrase people understand but don’t search for. The term shoppers actually type may be a borrowed English word, a regional variant, or something completely different.
My linguistics degree comes in handy here: language is what people use, not what the dictionary says. Keyword research has to be done again in each language, with real search data, before the translation starts. Otherwise you get beautifully translated pages targeting phrases with no demand.
Regional variants
French for France isn’t quite French for Canada or Belgium. Portuguese for Portugal isn’t Portuguese for Brazil. Spanish varies a lot across countries. AI tools default to whatever variant dominates their training data unless you tell them otherwise, and even then they slip.
Units, prices and conventions
Sizes, measurements, date formats, currency placement, how addresses are written. A good model handles some of this if asked. It won’t know your shipping rules or that a size chart needs converting rather than translating.
Tone and trust
Product copy that sounds slightly off makes people hesitate at checkout. It’s the same feeling you get from a scam email: nothing’s exactly wrong, but something doesn’t feel local.
The technical side still matters more than the words
Even perfect translations won’t perform if the setup is wrong. The basics:
- One URL per language version, in a subfolder or subdomain, never the same URL switching language by cookie or browser setting.
- hreflang connecting the versions, with return links. I wrote hreflang basics for multilingual sites for the details.
- Translated metadata. Titles, meta descriptions, image alt text and URL handles where the platform allows it. These get forgotten more often than body text.
- No automatic redirects by IP. Let people (and Googlebot, which mostly crawls from the US) reach every version.
On Shopify, Markets and the translation apps handle much of this; the setup choices are in Shopify international SEO with Markets.
A process I’d trust
If I were translating a site with AI today, this is the order:
- Keyword research in the target language first. Build a short glossary: your key product terms, the phrases people search, words to leave in English.
- Write the AI instructions properly. Target language and country, the glossary, the tone, what never to translate (brand names, model numbers).
- Translate the important pages first. Homepage, top collections, best-selling products, shipping and returns. Not the whole blog archive.
- Have a fluent person review every page before it goes live. Not a skim. They should change wording to what locals say, fix the keyword phrases, and flag anything that reads oddly.
- Launch in batches and watch Search Console for the new folder: are pages indexed, which queries appear?
- Translate more only when the first batch shows demand.
The review step is the expensive part, and it’s the part people want to skip. It’s also what makes the pages useful and keeps them on the right side of the spam policies.
Is it worth doing at all?
Sometimes the honest answer is no. A second language doubles the pages you have to maintain. Every price change, product update and new collection now happens twice. If you don’t have customers, shipping or support in that market, a translated site is a lot of upkeep for traffic that can’t convert.
If you do sell there, AI makes the first draft cheap. Just budget for the human review, because that’s where the pages go from “understandable” to “local”.
FAQ
Will Google penalise AI-translated pages?
Not for being AI-translated. Unreviewed, mass-published translations fit the scaled content abuse examples in Google’s spam policies. Reviewed translations are treated like any other content.
Should I use browser auto-translate instead of translated pages?
For SEO, no. Google can only rank pages that exist at their own URLs in that language. Browser translation helps visitors, but it creates nothing for search.
Which pages should I translate first?
The ones that make money: top products, top collections, and the pages that answer buying questions like shipping and returns.
Related reading
- Hreflang basics for multilingual sites
- Scaled content abuse and AI: where the line is
- International SEO for African businesses selling abroad
- Comparing AI writing assistants for SEO drafts
Before you translate anything, look up your top five product keywords in the target language with real search data. Do they match what a straight translation gives you?