How to Translate a Website With AI: What 2,800 Pages in 20 Languages Taught Us
How to translate a website with AI in 2026, from a team that did it 2,800 times: why 81% of our blog clicks now come from translated pages, the 10 things that broke, what Google expects from a multilingual site, and a checklist you can use on five pages this week.
This blog publishes every post in 20 languages. That is 143 posts and about 2,800 pages, and every one of them was translated by AI and checked by a process we had to work out as we went.
The result surprised us. In the 90 days to early October 2026, 81% of the clicks our blog earned from Google went to a translated page, not the English original. This guide shows how to translate a website with AI properly, using what those pages taught us: the numbers, the ten things that broke, what Google expects, and a checklist small enough to try on five pages this week.
Want your pages translated and checked, with the layout kept intact? Tell Jobbit which pages and which languages: 100 free credits a day, no card.
How we checked this guide
- Our own Search Console data. The click figures are from this site's Google Search Console for the 90 days to early October 2026, blog pages only.
- Google's published guidance. The rules for multilingual sites and for AI content come from Google Search Central, read on 4 October 2026.
- Named research for the wider picture. Buyer behaviour comes from CSA Research, and language share from W3Techs.
- Honest about scale. We are a young blog and the absolute numbers are small. Treat the ratios as a signal, not a law.
What the numbers say
Translated pages earned 42% of our blog's impressions in Google and 81% of its clicks. Put another way, a searcher who saw one of our pages in German, Polish or Korean was about six times as likely to click as one who saw the English version.
Two things explain it. English results are crowded: every topic has a dozen established sites, and many English impressions now come from AI-generated searches that never produce a click. In most other languages there is far less written about the same subject, so a clear, complete page stands out. Our three strongest languages by clicks were Russian, German and Polish, followed by Japanese and Korean. One language, Portuguese, earned plenty of impressions and no clicks at all, which tells us those titles need work.
The wider research points the same way. CSA Research surveyed 8,709 consumers in 29 countries and found that 76% prefer to buy products with information in their own language, and 40% will never buy from websites in other languages. Yet W3Techs reports that 49.5% of websites are written in English. The gap between what people want to read and what is published is the opportunity.
The 10 things that broke, and how we fixed them
| What went wrong | A real example | The fix |
|---|---|---|
| 1. The keyword stayed in English | A Thai title kept "Plumbing" in English and ranked for English searches at position 60 to 90 | Translate the main keyword in every title and heading |
| 2. A section vanished | On long posts, one FAQ question was dropped in Arabic and Chinese | Count headings, table rows and links in every language |
| 3. Numbers changed meaning | "2,000 credits" in French reads as two, because the comma is a decimal mark | Use the target language's own number format |
| 4. Grammar around numbers | Arabic needs a different noun form after 7 than after 15 | Have a second pass check every sentence with a number in it |
| 5. A phrase became ambiguous | Italian "da Go a £7" reads as "from Go to £7", a range | Read prices and plan names aloud in the target language |
| 6. Punctuation from the wrong language | Chinese titles came back with full-width brackets around English brand names | Set punctuation rules per language |
| 7. Links pointed at English pages | Internal links kept the English address | Rewrite internal links to the same language |
| 8. Leftover instructions | A stray tag from the translation step appeared at the end of a file | Strip and scan for anything that is not content |
| 9. Too-long descriptions | German and French summaries ran past the length search engines show | Give a character limit, not a word limit |
| 10. Brand names got translated | Product names were turned into local words | List every name that must stay as it is |
None of these was caught by reading the translation, because we cannot read 19 languages. They were caught by checks that do not need a human reader, and that is the core of the method.
The method that works
Start with a finished page
Translate only pages you are happy with in English. Every weakness in the original is copied 19 times, and fixing it later means 19 edits. Check the facts, settle the title and stop editing before you translate.
One language at a time, with rules
Give each translation its own instruction and the same set of rules. Ours fit on half a page:
- Keep the structure identical. Same headings in the same order, same table shape, same links.
- Translate the keyword in the title. Keep brand names in their original script.
- Keep every number, price and date. Change only the format the language expects.
- Write for a reader in that country. Explain a UK term once in brackets if they would not know it.
- Add nothing and remove nothing. No translator notes.
Check with counts, not eyes
After each translation, compare it with the English source by counting: the number of headings, table rows, links and questions, and the length compared with English. A page that is half the expected length has lost something. Chinese, Japanese and Korean pages come out much shorter in characters, so set the expectation per language.
Then a second reader in the same language
Counting catches what is missing. It does not catch what is wrong. For that, ask a second AI, acting as a native proofreader, to check meaning, grammar and number formats against the English. When we did this on 1,140 short edits in 19 languages, the reviewer corrected 60 of them, about 5%. Most were awkward phrasing. Five were real errors that would have gone live.
For pages that carry legal or medical weight, or your prices, pay a human native speaker to read them. For a blog post, the two-pass method has been good enough.
What Google expects from a multilingual site
Google does not penalise translated or AI-assisted content for being translated or AI-assisted. Its guidance says to focus on accuracy, quality and relevance, and to fact-check AI output before publishing. What it does penalise is scale without value: its spam policy names automated translation among the techniques that count as abuse "where little value is provided to users".
Four technical points come straight from Google's guidance on multilingual sites:
- Give every language its own address. A folder per language, such as /de/ and /pl/, is the simplest structure. Do not swap the language on one address using cookies.
- Tell Google the versions belong together. Hreflang tags link each page to its translations so the right one shows in each country.
- Do not redirect people automatically. Let visitors choose with a visible language switcher instead of forcing a version on them.
- Translate the whole page. A page with a translated menu and English content is a poor experience and can appear as duplicates.
If you are working on search more broadly, the 10-step method for SEO and AI search covers the rest.
Which pages and which languages first
You do not need 20 languages. You need the two or three your customers already speak.
- Look at who visits. Search Console and your analytics show the countries and languages you already reach. Start there.
- Look at who buys. If a quarter of your enquiries come from Polish speakers in your own town, Polish comes first.
- Start with the money pages. Home page, services, prices and contact. Blog posts come later.
- Be honest about support. If you cannot answer an email in German, say on the German page that replies will be in English.
A plumber in Peterborough may get more from one well-translated services page in Polish than from 50 posts in 20 languages.
The five-page test for this week
- Pick five pages that already bring enquiries.
- Pick one language from your own visitor data.
- Translate with the rules above, one page at a time.
- Run the counts and the second-reader check. Fix what they find.
- Publish at a separate address with a language switcher, then watch Search Console for four weeks.
If the translated pages earn clicks, add a second language. If they do not, you have spent an afternoon, not a budget.
How an AI agent helps you translate a site
ChatGPT, Claude, Gemini, Microsoft Copilot, Manus and Meta's Muse will all translate a paragraph you paste in. A website is hundreds of paragraphs with links, tables and settings between them, and that is where an agent earns its place.
- Translates whole pages and keeps the layout. Headings, tables and links come back in the same shape.
- Runs the checks. Counts against the original and a second-reader pass, in the same job.
- Builds the language versions. For a site it hosts, Jobbit can add the language folders and the switcher too.
- Lets you choose the model. You can pick the model per task, which helps when one handles a language better than another.
- Keeps translations in step. When the English page changes, ask for the same change in every language.
The free plan gives 100 credits a day. Paid plans start with Go at £7 a month, which includes hosting for one small site on your own domain. If you do not have a site yet, the guide to getting a free website for your business compares ten ways to start.
Mistakes to avoid
- Translating everything at once. You multiply the errors before you have found them.
- Trusting a translation you cannot read. Use counts and a second reader.
- Leaving the title in English. It is the one line every searcher sees.
- Forgetting the small text. Buttons, form labels and error messages need translating too.
- Publishing and walking away. Check Search Console after a month and rewrite the titles that get seen but not clicked.
Frequently asked questions
Can I translate my website with AI?
Yes. Modern AI translation is good enough for most business pages, provided you keep the structure, translate the keyword in each title and check every page. For legal text or anything where a mistake is costly, add a human native reader.
Does Google penalise AI-translated content?
Not for being translated by AI. Google's guidance asks for accuracy, quality and relevance. Its spam policy targets automated translation at scale where the pages give little value to users, so publish complete, useful pages.
How accurate is AI translation?
In our own check of 1,140 short edits across 19 languages, a second AI reviewer corrected about 5%, and five corrections were real errors. Expect most pages to be right and a few to need fixing, which is why the checking step matters.
Which languages should I translate my website into?
The ones your visitors and customers already use. Check the countries and languages in Search Console and your enquiries, and start with one. In our own data Russian, German and Polish brought the most clicks, but your audience will differ.
What is hreflang and do I need it?
Hreflang is a tag that tells Google which language each version of a page is for, so the German page shows in Germany and the English one in the UK. If you publish the same page in more than one language, you need it.
What is the quickest way to try this?
Translate five pages into one language and measure the result for a month. Start free on Jobbit with 100 credits a day and no card.