How to Review Machine Translation Before You Publish It

Machine translation is fast, and it can be wrong in ways that read smoothly. Review is how you find those errors before your customers do. The routine below works with any tool: start with the pages where a mistake costs the most, use automatic checks to decide where to look first, have a fluent person read the text in context, and keep unreviewed pages out of search until they pass.

Why review at all?

A machine translation can pick the wrong sense of a word, mishandle names, drop a negative, or sound odd in your industry. Google's guidance on automatically generated content makes the same point in general terms: it says it is critical to manually fact-check and review such content for accuracy before publishing. Translation is not named there, but the principle carries over.

Step 1: sort pages by the cost of an error

This is the same idea as in machine vs human vs hybrid translation: decide per page type, not for the whole site.

Step 2: run the automatic checks first

Tools can find mechanical faults quickly. Typical checks:

• Missing or broken tags. A link or bold word that disappeared, or a sentence whose markup is no longer nested properly.
• Changed placeholders. A customer name, price or {variable} that vanished or was altered.
• Numbers that changed between source and translation.
• Glossary violations. A do-not-translate term that was translated, or a preferred translation that was not used.
• Length oddities. A translation far longer or shorter than the source.
• Untranslated text. Identical to the source (sometimes correct, as for a brand name).
• Wrong script. A translation mostly in the wrong writing system.

LgTranslate runs checks of this kind and shows them as flags in the editor: errors mean the translation should not go live as it is, warnings mean look at it. Be clear about what they are: flags help you decide what to read first. They do not prove a translation is good, and an unflagged translation can still be wrong. The documentation also notes that whether every job in the finished product runs all checks is still marked as open.

Step 3: read in context, not as isolated strings

Short strings are easy to mistranslate because the meaning depends on where they appear. "Book" is a noun on a library page and a verb on a hotel page. Have the reviewer look at the translated page itself, or a preview of it, and check:

1. Meaning. Does it say what the original says, including negatives and conditions?
2. Terminology. Are your product names and key terms right and consistent?
3. Tone. Formal or informal address, politeness, the voice your brand uses.
4. Numbers, dates, currency and units. Formatted the way readers in that locale expect; see dates, numbers, currencies and units.
5. Layout. Text that is longer in the new language can break buttons and menus.
6. Links and forms. They still work and point to the right language version.

Step 4: choose who reviews

The best reviewer is a native speaker who also understands your business: a customer, a local partner, a colleague, or a professional. Give them a short brief: who your visitors are, which tone you want, and which terms are fixed. If you have no reviewer, review the highest-risk pages only and treat the rest as drafts.

Step 5: record what you learn

Every correction teaches the next run. Add recurring fixes to the glossary, mark brand names as do-not-translate, and keep human-edited translations protected so a later machine run does not overwrite them. In LgTranslate a translation a person edited or approved is never replaced by a machine re-run, and the editor shows the state (untranslated, machine, reviewed, manual) as a word, not only as a colour. Our guide to translation memory and glossaries explains how to build these.

Step 6: decide what goes live and what is indexed

Visitors who switch language on your site will see the machine text as soon as you publish it. For search, you have a choice. Google's documentation describes the noindex robots meta tag as the way to keep a page out of search results, and it must not be blocked by robots.txt for Google to see it. Using it for unreviewed drafts is our recommendation, not a Google requirement. Pages with very little translated text are a separate problem: Google's multi-regional guidance warns that translating only the boilerplate of a page while keeping most of the content in one language can create a bad user experience. See machine translation and Google Search for the policy background.

A 30-minute first review

If you only have half an hour for a new language:

1. Read the home page, pricing and contact pages, line by line.
2. Open your three most visited pages and read the headline, first paragraph and button text.
3. Look at every flagged string on those pages.
4. Check a form and one error message.
5. Fix the problems, add two or three terms to the glossary, and note what to review next time.

Next step

Open one machine-translated page in your best-supported language and review it with this list. If you are not using LgTranslate yet, create an account, translate a single page, and compare the flags with what you find by reading; you will quickly see what the flags catch and what they miss.

Key takeaways

• Review by risk, starting with prices, legal text and checkout.
• Use automatic checks to decide where to look, never as a substitute for reading.
• Review in context with a native speaker, and feed corrections back into the glossary.
• Keep unreviewed pages out of the index while they are drafts.

Sources

• Google Search's guidance on using generative AI content – Google Search Central – the advice to manually fact-check and review automatically generated content (checked 2026-11-06)
• Block Search indexing with noindex – Google Search Central – the noindex meta tag and the robots.txt caveat (checked 2026-11-06)
• Managing multi-regional and multilingual sites – Google Search Central – partly translated pages (checked 2026-11-06)