Machine translation post-editing (MTPE) for website copy

Machine translation post-editing, or MTPE, means a person who reads both languages corrects machine output against the source until it reaches an agreed quality. For website copy there are two levels. Light post-editing fixes meaning and leaves style alone, which suits help articles and older posts. Full post-editing brings the text up to the standard of a human translation, which is what your home, product and pricing pages need. On a website, the check also has to happen on the page itself, because a correct translation can still overflow a button or leave the page title in English.

Light and full post-editing#

The guidelines TAUS published with CNGL in 2010 describe two target qualities.

"Good enough" is reached by light post-editing. TAUS defines it as comprehensible and accurate, "but as not being stylistically compelling. The text may sound like it was generated by a computer, syntax might be somewhat unusual, grammar may not be perfect but the message is accurate." The editor makes sure "no information has been accidentally added or omitted" and makes no corrections "that are of a stylistic nature only."

Quality "similar or equal to human translation" is reached by full post-editing. The text must be accurate and "stylistically fine, though the style may not be as good as that achieved by a native-speaker human translator." Grammar and punctuation are correct, terminology is checked, and the editor will "ensure that formatting is correct."

There is also an international standard. ISO 18587:2017 is titled "Translation services: Post-editing of machine translation output: Requirements." The guidelines also say something worth remembering when you budget: "light postediting of really poor MT output may not bring the output up to publishable quality standards." If the raw output is weak in a language, light editing will not rescue it.

Which level each page needs#

Smartling's 2026 pricing guide lists its own post-edited service at $0.12 per word against $0.15 to $0.30 per word for professional human translation in the United States, which gives a sense of the saving vendors price in. Our comparison of machine and human translation works through a full cost model, and how accurate Google Translate is shows how much raw quality varies by language.

What to check on a web page#

A post-editor working on a document reads sentences. On a website, the same text appears as headings, buttons, menu items, image descriptions and search snippets, and each has its own failure. Work through this list for every page at full level, and the first four items at light level.

1. Meaning. Compare each line with its source. Numbers, prices, dates, names and negations come first, because a mistranslated "no setup fee" reads naturally and costs you money.
2. Omissions and additions. Machine output sometimes drops half a sentence or adds a clause that is not in the source.
3. Untranslated text. Look for lines still in the source language, and for text in the wrong target language.
4. Terminology. Names your glossary keeps as written must still be as written, and fixed terms must use their fixed translation. Our guide to building a translation glossary covers what goes on that list, and why a chat prompt is a poor place to keep it (ChatGPT for website translation).
5. Form of address. In languages with a formal and informal "you", the whole page should use one, and the site should be consistent.
6. Links and formatting. Every link should still be there and go where the source link goes. Bold, lists and line breaks should match.
7. Fit. Read the page at mobile width. A translation can run longer than the source, and a label that wraps onto two lines inside a button is a layout bug.
8. Search text. The page title and meta description are what searchers see first, so they get the same care as the headline.

If you need a vocabulary for logging problems, the MQM typology groups translation errors under seven headings: terminology, accuracy, linguistic conventions, style, locale conventions, audience appropriateness, and design and markup. For a website, a shorter list is enough, as long as your team agrees which problems block publishing and which are notes for later.

Running post-editing as a review queue#

Post-editing a whole site in spreadsheets breaks down quickly. You lose track of which lines were checked, a fix on one page does not reach the same sentence on another, and the reviewer never sees the text in place. LangRelay organizes the same work around the site itself.

The translation table#

The Translations table has one row per piece of source text and one column per language. A sentence that appears on several pages is one row, listing the pages it is on, and an edit in that cell applies to every one of them. Each cell has a state you can filter by: Not translated, Machine translated for machine output nobody has approved, Needs review for text a person entered or imported, Approved, and Approval outdated for a line whose approval covers an earlier version. Filtering a language to Machine translated gives a post-editor the queue. Selected rows can be approved in one action, sent for a fresh machine translation with Machine retranslate, or sent to Relay for review.

The page itself#

The visual editor shows the translated page as visitors will see it, at desktop, tablet and mobile widths. You click text on the page to edit it, which is where checks 6, 7 and 8 above are easiest: you see the button that wrapped and the heading that no longer fits.

A first pass by Relay#

Review with Relay sends pages to Relay, which compares each translation with its source the way a language expert would, and tells you the cost in AI credits before it starts. It sorts what it finds into major and minor. Major findings include the first three checks above: a changed number, price, date or name, content left out or added, and text untranslated or in the wrong language. Minor findings are wording that is correct but awkward, punctuation and capitalisation. The editor sidebar then shows "Relay reviewed: looks good", or a count of findings that need you, each pinned to its line, with the major ones visited first. Let Relay fix these opens Relay with the findings already named, so it can draft corrections, which you then check like any other edit.

The person who decides#

Whatever a Relay review says, Relay cannot approve or publish a page on its own. On a page that passed, Go live in the visual editor approves and publishes it in one click. On a page that did not, you work through the findings first. If you want someone else to redo the work, Request changes records your reason on the page, with your name and the date, where the next editor sees it.

Each approval is recorded against the exact text that goes live. The translation quality page describes these controls together. To try it on your own site, book a demo.

Common questions#

Is post-editing the same as proofreading?#

No. A proofreader reads the translation on its own. A post-editor compares it with the source, which is the only way to catch a dropped clause or a changed number that still reads smoothly. The TAUS guidelines assume bilingual post-editing for this reason.

How long does post-editing take?#

It depends on the raw output and the level you need. TAUS names those two factors as what determines the effort: "the quality of the MT raw output" and "the expected end quality of the content." Measure it on a few of your own pages in each language before you plan a launch.

Can AI do the post-editing?#

An AI reviewer can do the first pass and find many of the problems on the checklist, which is what Relay's review is for. The decision to publish should stay with a person who reads the language, or who trusts someone who does.