On this page
- How QA differs from editing the translation
- Decide what counts as an error, and how bad
- Before translation: check the source
- Checklist 1: the text against its source
- Checklist 2: the page as visitors see it
- Checklist 3: the live site as machines read it
- Checklist 4: after every change
- How much to check
- How LangRelay runs these checks
- Questions people ask about translation QA
- Is translation QA the same as proofreading?
- Who should do translation QA?
- Can AI do translation QA?
Translation quality assurance for a website is a set of checks, done by someone other than the person who translated, that decides whether a translated page is fit to publish. It runs in four places: the text against its source, the page as visitors see it, the live site as browsers and search engines read it, and every later change. Before any of that, agree what counts as an error and how serious each kind is, because that decides what blocks publishing and what can be fixed next week.
How QA differs from editing the translation#
Post-editing is writing: a person improves machine output until it reads well. Our article on machine translation post-editing covers the two levels and a first list of things to fix. QA is checking: someone looks at the finished result against agreed criteria and says pass or fail.
The translation industry builds a second person into this. The MQM Council's terminology, citing ISO 17100, defines a reviser as the "person who revises target language content against source language content", and notes that "the reviser shall be a person other than the translator". It defines a reviewer as someone who "focuses solely on target language content". For a website you want both views: a reviser who compares each line with the source, and a reader who checks the page as a visitor in that market would see it.
Decide what counts as an error, and how bad#
The Multidimensional Quality Metrics framework (MQM) is the shared vocabulary for this. Its full typology has seven top-level dimensions:
MQM's scoring model then gives each error a severity. Its sample scorecard has four levels. Neutral is an acceptable translation that differs from the evaluator's preference. Minor "does not seriously impede the usability, understandability, or reliability of the content". Major "seriously affects the understandability, reliability, or usability", for example through "a significant loss or change in meaning" or because it sits "in a highly visible or important part of the content". Critical "renders the entire content unfit for purpose or poses the risk for serious physical, financial, or reputational harm", and MQM notes that in many systems a single critical error fails the whole text.
Write your own version of this before anyone starts checking. On a shop, a wrong price on a product page is critical. A misspelled word in a three-year-old blog post is minor. Agreeing that up front stops a reviewer from holding back a whole page for a comma, and stops a wrong price from going out because "it was only one word".
Before translation: check the source#
Many translation errors start in the source site. Fix these before the first word is translated, because each one multiplies by the number of languages.
- Terms decided. Brand names, product names and fixed phrases are in a glossary with their translation, or marked to stay as written. See how to build a translation glossary.
- Form of address chosen per language, formal or informal, and written down.
- Text inside images listed, so it is either recreated per language or replaced with real text.
- Sentences assembled from pieces in code ("You have" + number + "items") noted for the developer, because word order changes between languages. Our article on internationalization vs localization explains why.
Checklist 1: the text against its source#
Work through each line with the source beside it. Use the MQM dimensions above as the order of priority: accuracy and terminology first, style last.
- Every number, price, date, unit and name matches the source.
- Nothing is left out or added. Compare paragraph counts, list lengths and table rows.
- Nothing is left in the source language, and nothing is in the wrong target language.
- Glossary terms are used everywhere they occur, including headings and buttons.
- Formats follow the locale. MDN's examples show why this needs checking: US English writes a date as 12/19/2012, British English as 19/12/2012, and German writes the number 123456.789 as 123.456,789.
- One form of address across the page and across the site.
- Links in the text still point where the source links point, ideally at the same page in the same language.
Checklist 2: the page as visitors see it#
A correct translation can still break the page. Check it in the browser, at desktop and phone widths at least.
- Text fits. W3C's article on text size in translation quotes IBM's guidance that short English strings expand most in European languages: up to 10 characters by 200 to 300%, 11 to 20 characters by 180 to 200%, and texts over 70 characters by about 130%. Buttons, menu items, tabs and table headers are where it shows.
- Nothing in the theme stayed in the source language: menus, footer, search box, cookie banner, form labels, placeholder text and error messages.
- Image descriptions (alt text) and accessible labels are translated, since screen readers read them aloud.
- The page title and meta description are translated and read naturally, because they are the search result.
- Forms work: required-field messages, date pickers and confirmation text appear in the page's language.
- The language switcher shows the current language and links to the same page in each other language.
Checklist 3: the live site as machines read it#
These checks are invisible to a reader and decide how browsers, assistive technology and search engines treat the page.
- The
htmlelement carries the rightlangvalue. WCAG 2.2 makes this success criterion 3.1.1, Language of Page, at level A, and W3C advises always putting the attribute "on the html tag", since it "is inherited by all other elements". - A passage left in another language on purpose, such as a quote, is marked with its own
lang. That is WCAG 3.1.2, Language of Parts, at level AA. - Hreflang lists every version, and every version links back. Google's documentation is direct: "If two pages don't both point to each other, the tags will be ignored." The hreflang guide has the full rules.
- Each version has a canonical that points at itself.
- The robots rules are the ones you intended. A
noindexleft on from staging hides a whole language. - Internal links on the translated page go to translated pages, not back to the original language.
Checklist 4: after every change#
QA does not end at launch. A changed price, a new paragraph or a corrected sentence in the source creates new text in every language, and each piece needs the same checks as the first translation. Keep a list of what changed, check those lines first, and look again at the pages that show them. Untranslated sentences in the middle of an otherwise translated page usually come from exactly these later changes.
How much to check#
You do not have to read every word with the same care. MQM's scoring page notes that evaluators "frequently work with samples in the range of 500 to 20,000 words". A reasonable split for a website: check every line of the pages that sell and the pages with legal weight, and sample the long tail of older articles and help pages. Its sample threshold, given as an example, is "10 minor errors OR 2 major errors per 1000 words". Set your own threshold per page type, and fail any page with a critical error outright.
How LangRelay runs these checks#
LangRelay builds the review into the place where translations are managed, so most of this checklist happens in one queue.
Relay reviews any page you ask it to, the way a language expert would, and tells you the cost in AI credits before it starts. A Relay review reports each problem as major or minor, close to MQM's middle two levels. Major means the meaning changed (a number, price, date, name, product or quantity differs, or a negation was lost or added), content was left out or added, text is untranslated or in the wrong language, or a tag in the text is missing or broken. Minor means correct but could read better: phrasing, style, a line in a different form of address from the rest of the page, punctuation. It recommends pass, fail or uncertain, and it never approves or publishes anything.
A person then works through the translations table, which filters by page, language and five translation states and searches source and target text, or through the Visual editor, which shows the rendered page at 1440, 768 and 390 pixels wide, so checklist 2 happens on the real layout. The editor's SEO tab lists the title and meta tags. The glossary flags terms with no translation for a language yet. You can invite a reviser or reviewer by email with a preset that lets them suggest changes or approve translations, without a WordPress login.
By default a translation goes live as soon as it is ready, and free checks hold any page that looks wrong: text left untranslated, the wrong language, a translation far longer than its source, or a changed number or price. To approve every page yourself, choose Ask me first. After each activation, a delivery check reads the live page and confirms the translated text was applied and that lang, the canonical, hreflang and the robots rules are what they should be, which covers most of checklist 3. When the source changes, the plugin reports the edit about a minute later, only the new text is translated, and it goes live like any other translation: right away by default, unless the free checks hold it. If you want to see the review queue on your own site, book a demo.
Questions people ask about translation QA#
Is translation QA the same as proofreading?#
Proofreading is one step: a final read for typos and mechanical errors. QA covers the whole result, from meaning and terms to layout and the markup search engines read, and it ends in a decision to publish or not.
Who should do translation QA?#
Someone other than the translator, who reads both languages, for the comparison with the source; and if possible a reader in the target market for the page as a whole. ISO 17100, as MQM's terminology cites it, requires the reviser to be a different person from the translator.
Can AI do translation QA?#
AI can do a first pass that finds candidates, especially changed numbers, omissions and untranslated text, and that is what a Relay review does. Its report is evidence for a person's decision: Relay cannot approve anything.
Sources
The pages this one cites, with the day each was read.
- 01MQM Council, The MQM Full Typologythemqm.org, checked 7 October 2026
- 02MQM Council, The MQM Scoring Modelsthemqm.org, checked 7 October 2026
- 03MQM Council, Multidimensional Quality Metrics Terminology (2026-01-13, citing ISO 17100:2015)themqm.org, checked 7 October 2026
- 04W3C Internationalization, Text size in translationw3.org, checked 7 October 2026
- 05W3C, Web Content Accessibility Guidelines (WCAG) 2.2w3.org, checked 7 October 2026
- 06W3C Internationalization, Declaring language in HTMLw3.org, checked 7 October 2026
- 07Tell Google about localized versions of your page (Google Search Central, last updated 2026-09-21)developers.google.com, checked 7 October 2026
- 08MDN Web Docs, Intl.DateTimeFormatdeveloper.mozilla.org, checked 7 October 2026
- 09MDN Web Docs, Intl.NumberFormatdeveloper.mozilla.org, checked 7 October 2026






