How to Use AI to Translate Your Website and Support Content

Most small businesses leave money on the table in a language they do not speak. You rank well in English, your support articles are clear, your checkout works — and then a prospect in Lyon, Milan or Lisbon lands on your site, reads three lines, and leaves. Fixing that used to mean a translation agency, a quote in the thousands per language, and a six-week wait. In 2026 it costs a fraction of that, and the hard part is no longer the translating.

The hard part is everything around it: choosing which languages are worth the effort, keeping terminology consistent, and keeping the translated version current when the English one changes next month. "Every guide is just surface-level advice" is a fair complaint about most articles on this: they tell you to paste text into a chatbot and stop there. This one walks the whole workflow, with the prompts and the numbers.

What AI translation does well — and where it still fails

Modern language models carry context across a whole page rather than sentence by sentence, and can be told to hold a brand voice. They are strong on marketing and web copy into the major European languages — French, German, Spanish, Italian, Dutch, Portuguese — where output is usually publishable after a light human read, and strongest of all on support articles. Asked to adapt rather than translate, they will give a German version with different sentence lengths and a more direct register, not a word-for-word carbon copy.

Where a human still has to look:

• Legal and regulatory text. Terms, privacy notices, warranty and refund wording. Mistranslate a liability clause and the cheap version becomes the expensive one.
• Prices, units and formats. A model will happily leave "$49" in a French page, or convert 40 square metres to square feet because it has seen that pattern a thousand times.
• Humour, and smaller languages. Your clever headline comes back literal and slightly odd; quality drops noticeably in Finnish, Greek, Hungarian or Catalan — usable, but needing a native reader.
• Invented confidence. If a product name, regulation or job title has no equivalent, the model may invent a plausible-looking one — the failure mode covered in preventing AI hallucinations in client work.

The working rule: AI drafts everything, a human reads anything with a legal, financial or reputational edge.

Step 1: Choose the languages that will actually pay back

The instinct is to do five at once because the marginal cost looks like nothing. Resist it: every language is a permanent maintenance commitment — every price change, new feature and new support article carried across, forever.

Spend twenty minutes pulling three numbers you already have:

1. Where your traffic comes from. Sessions by country and browser language. A market sending 4% of sessions at twice your average bounce rate wants to buy and cannot read you.
2. Where your enquiries come from. Search your inbox and CRM for non-English enquiries, or English ones with obvious non-native phrasing. These people pushed through friction to reach you; others behind them did not.
3. Search volume for your two or three core terms in the candidate language. If nobody searches for your category in Dutch, translating your site will not conjure demand.

Then do the arithmetic honestly. A language worth €3,000 of plausible annual revenue, costing a day of setup and an hour a month, is worth doing. One worth €300 is a hobby. Pick one, ship it properly, measure a quarter, then add the second.

Step 2: Build a glossary before you translate a word

This is the step everyone skips, and the one that separates a professional result from an obviously machine-translated site. Without it your product tier is "Pro" on one page, "Professionnel" on another and "Avancé" on a third, and the reader quietly concludes you are not a serious company.

Build one table: the English term, the agreed translation per language, and "do not translate" where it applies. Cover product and tier names (usually: do not translate), feature and navigation labels, your signature headline phrases, industry terms with more than one plausible translation, and button and form labels, which must stay short in every language.

Add a short style sheet: formal or informal address (vous or tu, Sie or du — for most European B2B, formal), currency and date formats, decimal separators, and which technical terms your market uses in English anyway.

A prompt that drafts it for you:

You are a localisation specialist. Here is the homepage, pricing page and product page of my business [paste text]. Produce a glossary table of every term that must stay consistent across the site: product names, tier names, feature names, navigation labels and signature marketing phrases. For each, propose a French translation, flag anything that should stay in English, and note where a wrong choice would confuse a French B2B buyer. Then propose a short style sheet: form of address, currency format, date format and tone.

Review it — you know your product — then freeze it and include it in every translation request from here on. It belongs in your central business reference; our guide to building an AI knowledge base covers structuring one so every tool and team member works from the same source.

Step 3: Translate the site, page by page

Work page by page, not site at once: one page per request gives the model enough context to be coherent and keeps the output short enough to check.

Translate the following webpage into French for a French B2B audience. Rules: use the attached glossary exactly; formal address (vous); keep all HTML tags, links and attributes unchanged; keep prices in euros and convert no numbers or units; keep product names in English; adapt headlines to read naturally rather than translating literally; stay within 10% of the original length so the layout holds. Then list separately: (1) any phrase you adapted rather than translated, and why, (2) anything you were unsure about, (3) anything resembling a legal or pricing claim a human should check.

Those three closing lists are the whole trick. They turn an opaque block of text you cannot evaluate into a short review queue you can: five minutes checking ten flagged items instead of thirty re-reading a language you may not speak.

Do not forget the invisible text

Page titles, meta descriptions, image alt text, form validation messages, email templates, PDF invoices, the 404 page. Half of every half-finished multilingual site is the bits nobody saw in the design. Ask for these explicitly, one batch per language.

Get the technical setup right

Translated content with broken technical signals ranks nowhere. Three non-negotiables: each language on its own stable URL (/fr/ subdirectories are simplest for a small site), hreflang tags linking the versions to each other and to an x-default, and the right lang attribute on each page's <html> element. Never auto-redirect by IP; show a language switcher and let people choose. Our AI for SEO guide covers how to brief and verify these checks.

Have a native speaker read the money pages

Budget two to three hours of a native freelancer's time — €60 to €150 — for your homepage, pricing page and checkout flow. Not to retranslate: to catch the three sentences that are technically correct and subtly wrong. Highest-return money in the project.

Step 4: Translate support content and keep it current

Support content is where AI translation earns its keep: high volume, low risk, factual, and translating it cuts tickets in markets you never served properly.

Translate in batches of related articles so terminology stays consistent within a topic, and order them by what gets read — your ten most-viewed help articles and ten most common ticket subjects first. Twenty articles usually cover 70–80% of volume. If you are also automating replies, AI customer service automation for SMBs covers how that translated set feeds a multilingual support flow.

Then solve the part that kills most multilingual projects: drift. Six months on, the English help centre has twelve new articles and nine edits, and the French one is a museum. Three habits fix it:

• Make translation part of publishing, not a project. An article is not done until every language version exists — one extra prompt, two minutes.
• Keep a source-of-truth list of every page and article with the date each language version was last updated. A spreadsheet is fine. What you cannot see, you will not maintain.
• Run a monthly fifteen-minute sweep. Diff the list, translate what changed, update the dates.

What it costs — and the five mistakes that cost more

For a typical small business site — 12 to 15 pages plus 25 support articles — one language costs roughly: nothing extra for the AI, since a €20–€25 monthly subscription covers this volume; €60–€150 once for native review of the money pages; one focused day of your time, then about an hour a month; and half a day of technical work for URLs, hreflang and the switcher.

Call it €200 and a day and a half for the first language, and under half that for the second, since the glossary and scaffolding already exist. An agency quote for the same scope typically runs €1,500 to €4,000 per language — and that gap is the reason to spend part of the saving on the native review rather than pocketing it all.

The five mistakes that reliably cost more than they save:

1. A translation widget instead of real pages. Client-side widgets produce no indexable content: all the maintenance of a multilingual site, none of the search benefit.
2. Launching five languages at once. Five half-maintained languages read worse than one good one, and you will abandon all five by spring.
3. Translating without a glossary. Inconsistent product and feature names are the clearest possible signal that nobody is minding the shop.
4. Translating legal text unchecked. Have a human who knows the local rules read anything about liability, refunds, data or warranties.
5. Treating it as a one-off project. Translation is the easy 20%. Keeping it current is the other 80%, and where the value either compounds or evaporates.

If translation is what you sell rather than how you market, the tooling question differs — see AI tools for translators in 2026.

The bottom line

Going multilingual used to be a strategic decision with an agency invoice attached. Now it is an operational habit with a glossary attached. The businesses that get it right in 2026 are not the ones with the best model — everyone has the same models. They are the ones who picked one language on evidence rather than enthusiasm, froze their terminology first, paid a native speaker to read the three pages that take money, and built the monthly fifteen minutes that stops the whole thing rotting. Do that once and the second language costs an afternoon.