See the builder itself
One site in up to eight languages, with an AI-assisted, review-before-save translation workflow for your page copy and the interface alike.
An AI website translation review checklist: what to trust, what to check
Machine translation is genuinely good now — good enough that "just translate the whole site" is tempting. Wrong question. The right one: which specific pieces of your site are safe to ship after a light check, and which need a human's full attention every time, however good the tool gets. Here is that line, plus a review workflow that works on any site, with any tool.
Content that's usually safe with a light check
Modern machine translation handles straightforward, literal, low-ambiguity text well. This is generally lower-risk to publish after a spot-check rather than a full professional review:
Descriptive body copy — a blog post, an about-us paragraph, a feature description — where the meaning is direct and there's little room for a mistranslation to change what's actually being said.
Straightforward product descriptions — "cotton crew-neck T-shirt, machine washable" translates cleanly because it's specific and literal, not idiomatic.
FAQ answers that are already written plainly — if the source answer is clear and jargon-free, the translation usually is too.
Content that needs a real review pass
Some content is higher-stakes or more failure-prone in translation even when the tool is good, because the risk isn't fluency — it's a subtle change in meaning, tone, or accuracy that a fluent-sounding sentence can hide:
Calls to action and button copy Short, high-visibility, and often idiomatic ("Get started," "Book now") — a stiff or slightly-off translation here is disproportionately noticeable because it's the exact text a visitor reads right before deciding whether to trust you.
Page titles and meta descriptions These are an SEO problem as much as a translation one: a literal translation of your English keyword phrase is not necessarily what someone actually searches for in that language — this needs a second look for search intent, not just grammar.
Pricing and currency copy Number and currency formatting conventions differ by locale even when every word is translated correctly — a machine translation can leave a decimal or thousands separator in the wrong format for the target market.
Navigation and menu labels Brevity and locally-established conventions ("Sign in" vs. a literal translation of "enter session") matter more here than literal correctness — a menu label that's technically right can still look foreign.
Anything with humor, wordplay, or idiom These are the cases machine translation still handles worst, because the meaning the model has to preserve isn't the literal words.
Content that needs a human, every time
A smaller set of content is high-stakes enough that it should never rely on machine translation without dedicated human review — ideally from someone who understands both the language and the subject matter, not just a bilingual speaker:
Legal and policy pages Terms, privacy policies, returns and withdrawal terms carry real contractual and regulatory weight — see the EU withdrawal-right guide for a sense of how specific the underlying legal requirements can get. A mistranslation here isn't cosmetic; it can misstate what you're actually obligated to do.
Regulated or safety-related claims Health, safety, or financial claims can carry legal consequences that a fluent-sounding but imprecise translation can worsen.
Tax and VAT-specific language Tax terminology is often locale-specific in ways a general-purpose translator doesn't reliably capture.
Brand names, product names, and other do-not-translate terms Left unmanaged, a generic translation tool will sometimes "helpfully" translate a proper noun that should have stayed exactly as written. Build a short glossary of terms that must never be translated before you start, not after you spot the first mistake.
Numbers, units, dates, and addresses Even a perfect translation of the surrounding words can sit next to a date or measurement in the wrong locale format if nothing is checking that separately.
A practical review workflow, for any site or tool
1
Build a do-not-translate glossary first
Brand names, product names, your legal entity name — decide what should never change, before you translate anything, so it doesn't get caught mid-batch.
2
Translate in reviewable batches, never a silent bulk publish
Whatever the tool, avoid a single "translate everything and go live" action with no checkpoint — batches you can actually look at beat one irreversible pass.
3
Flag anything machine-touched as "needs review," not "done"
A translation that's merely produced shouldn't be treated the same as one a human has actually looked at and approved — keep those two states visibly distinct until someone closes the gap.
4
Route by risk, not by convenience
Send legal, pricing and CTA copy to a human reviewer before anything else; let lower-stakes body copy get a lighter spot-check. Reviewing everything to the same depth wastes effort on the safe content and under-serves the risky content.
5
Re-check a translation whenever its source text changes
A stale translation — one that no longer matches an edited source — is arguably worse than a machine one, because it reads as authoritative and reviewed while quietly being wrong.
6Verify placeholders and markup survived, not just the words
A classic silent failure: a variable like {{name}} or a bold tag gets mangled, dropped, or duplicated in translation, and the sentence around it still reads fine — so this needs an explicit check, not just a read-through.
Look at the translated page rendered, not just the string in a spreadsheet or editor
Text that reads correctly in isolation can still break layout, truncate in a button, or clash with an image once it's actually on the page — the real check happens on the real page.
How Olmira handles this
The interface itself — navigation labels, buttons, checkout copy, form fields — has a real AI-assisted workflow behind it in Olmira: it only fills in what is genuinely missing, stale after an edit, or previously rejected, and it never silently overwrites a human translation.
Every AI-produced string lands as pending in a review queue until someone explicitly approves or rejects it. A glossary of do-not-translate terms is honoured automatically, and a parity check verifies that placeholders and formatting markup survived translation intact — if it can't produce a clean result, it leaves the original text rather than shipping something silently broken.