Can website translation be fully entrusted to AI?

Short answer: which pages to hand over to AI

The logic is simple: the closer a text sits to the purchase decision or to a legal obligation, the more an error costs.

How good is automatic website translation

For ordinary descriptive text — quite good. Modern neural translators and language models produce grammatically correct, readable sentences on common language pairs such as English and Ukrainian. A blog article or a help page often just needs a read-through after that kind of translation.

Quality drops sharply in three places:

• Short fragments without context. A word on a button, a menu item, a form field label. The AI sees a single word and has no idea where it sits.
• Industry terminology. The model picks the most common rendering of a term, not the one your industry or your customers actually use.
• Brand voice. Translation carries the meaning but not the tone. A friendly voice turns neutral, a joke stops being a joke.

So the question “how well does AI translate” really has two answers: for running text — well; for interface copy and sales wording — unreliably.

How machine translation affects your rankings in Google

Directly — not at all, provided the translation is good and technically implemented correctly. Indirectly — through how people behave on the page, and through whether Google can see your language versions in the first place.

What changed in Google’s position

A few years ago Google’s documentation advised blocking automatically translated pages from indexing via robots.txt. In 2025 that advice was removed.

The trigger for the discussion was Reddit, which began translating pages at scale with AI. Asked whether this was a violation, Google replied that its policies do not strictly classify AI-translated content as spam. Translation is mentioned only in the rules against mass low-quality content: as an example of automated transformation, where someone else’s material or your own is used to generate many pages that give people no value.

What this means for your site:

• Translating your own site with AI is not a violation. What gets judged is the result, not the method.
• Translating thousands of pages and checking none of them is a risk. Especially if the translation is hard to follow and the pages duplicate one another.
• Translation quality shapes how people behave. If someone cannot understand the page and goes back to search, that is bad for the page regardless of who translated it.

Does an automatically translated page get indexed

It does, if every language version has its own address. It does not, or only partly, if the language switches on the fly.

Google’s documentation gives three direct recommendations on this:

• Separate URLs for each language. Not cookies, not browser settings, but separate pages: /en/, /pl/, /de/.
• Be careful with dynamic language switching. Google warns that it may then fail to find and crawl all the variants. The reason is technical: the crawler usually arrives from the US and does not announce which language it wants the page in.
• hreflang annotations between language versions, so that Google shows a person the right language.

The practical conclusion, which usually comes as an unwelcome surprise to site owners: a widget that translates the page right inside the visitor’s browser is almost useless for search. The person sees the translation, Google does not. That kind of “translation” brings in no search traffic at all.

And one more common mistake, this time at the settings level: a canonical from every language version pointing to the original. To Google that signals the translated pages are copies that should not appear in search. How canonical differs from hreflang and how to set it up correctly is covered in the article “What the rel=”canonical” attribute is”.

The downsides of AI website translation

• Context errors in the interface — the most frequent and the most visible type.
• Inconsistent terminology. The same term rendered differently on different pages.
• Date, currency and unit formats. The date 03/12 means March to an American and December to a Ukrainian.
• Text length. A Ukrainian translation is usually longer than the English original, and the text spills out of buttons and cards.
• Loss of brand voice. The meaning is there, the character is gone.
• Untranslated technical elements. Titles, descriptions, image alt text and error messages often stay in the source language.

Examples of machine translation errors on websites

Most errors appear where the AI gets a standalone word with no sentence around it. Here are typical examples from interfaces, translated from English into Ukrainian:

For the Ukrainian market there is one more class of error — calques carried over from Russian. They turn up both in machine translation and in human translation that no editor has seen:

None of these errors makes the text incomprehensible. But every one of them tells a native speaker the same thing: nobody proofread this.

Why automatic translation erodes user trust

The fact of machine translation does not itself cost you trust — the visitor has no idea how you translated. What costs you trust are the errors visible at first glance.

The mechanism is almost reflexive. Broken language, odd phrasing and untranslated fragments are exactly what marks out phishing sites and fly-by-night online shops. People do not analyse; they simply sense that something is off and move to a competitor whose text reads as if a human wrote it.

For a company entering a new market this hurts especially. CSA Research data shows that the wish to buy in one’s own language is not a whim: it determines whether someone will consider your site at all. But if that language reads awkwardly on the page, the native-language advantage disappears. A translation that a visitor spots as machine-made works about as well as no translation at all.

When machine translation is enough

There are situations where AI translation without deep editing is a sensible decision rather than a compromise:

• A large reference section. Hundreds of instruction pages where accuracy of content matters and style does not.
• Testing a new market. You need to know whether demand exists before investing in professional localisation.
• Blog and archive material. Articles that bring in informational traffic and do not sell directly.
• Product specifications. Figures, dimensions, materials — where there is little to get wrong.
• Internal material and working drafts for the team.

AI translation is not enough wherever the text sells, promises or commits: the home page, services, landing pages, payment and delivery terms, the public offer.

Translating into Ukrainian: what the law says

For the websites of companies registered in Ukraine, a Ukrainian version is mandatory: since 16 July 2021 it must load by default, and in volume and content it cannot be smaller than the versions in other languages.

Note the wording: the law requires completeness, not quality. AI translation formally lets you meet the requirement fast — the whole page will be in Ukrainian. But the calques and interface errors will be seen by a customer, not by an inspector. More on the legal requirements for websites and the penalties is in the article “Translating a website into Ukrainian”.

Frequently asked questions (FAQ)