Machine vs human translation for websites: how to choose

For a website, the choice between machine and human translation is best made page by page. Machine translation costs around $10 to $20 per million characters and is fast enough to keep up with every edit, but it makes mistakes a reader in the target language will catch. A professional translator costs cents per word and takes longer, and is more reliable on names and legal meaning. The setup we recommend is a mix: machine output everywhere, a person reviewing it, and a translator for the pages where an error would be expensive. So decide which of your pages belong in which group.

What you pay for each#

Prices below were checked on 2026-10-06. Rates vary by language pair and subject, so treat them as reference points for a budget rather than quotes.

Machine translation#

Google's Cloud Translation API charges by character. Its standard neural model translates the first 500,000 characters each month free (applied as a $10 credit), then costs $20 per million characters. Its translation LLM is listed at $10 per million input characters and $10 per million output characters. Smartling's 2026 guide puts machine translation in general at about $10 to $20 per million characters.

Human translation#

Smartling's guide, published 2026-09-15, says that "in the United States, professional translators usually charge between $0.15 and $0.30 per word." Agencies that publish price lists can come in lower. TheWordPoint lists English to Spanish at $0.08 per word for general content, $0.10 for business and marketing, and $0.12 for technical or medical text, with German at $0.10, $0.12 and $0.15.

Machine output with human editing#

Between the two sits post-editing: a person corrects machine output against the source. Smartling lists its own service of that kind at $0.12 per word. The TAUS post-editing guidelines describe two levels: light post-editing for "good enough" text, and full post-editing for quality comparable to human translation. Our article on post-editing for website copy explains both.

A worked cost model#

Take a business site with 40 pages of about 500 words each, so 20,000 words, going into Spanish, French and German. That is 60,000 words of translation.

For the API line we assumed about six characters per English word including spaces, which makes the site roughly 360,000 characters across the three languages. That fits inside Google's 500,000 free characters. Billed in full it would be about $7. The API price leaves out the work of extracting text, rebuilding pages and checking the result, which is where most of the cost of a do-it-yourself approach goes.

The first translation is only part of the bill, because websites change. If a tenth of the site is edited each month, that is 2,000 source words, or 6,000 words across three languages, each month. At $0.10 a word that is $600 a month of translation, plus the time someone spends finding what changed. A machine-first workflow moves that monthly cost from translating to reviewing.

LangRelay's own plans include a monthly word allowance, and the current numbers are on the pricing page.

Decide by content type#

The accuracy studies we cite elsewhere agree on one point: results vary by language, and fluent output can still change the meaning. A 2021 study of Google Translate on hospital discharge instructions found meaning kept in 94% of Spanish translations and 55% of Armenian ones, and its authors said it "should not be relied on for patient instructions." (More on those studies in is Google Translate accurate.) For a website, that suggests sorting content by what a mistake would cost.

Traffic and language can each move a page up the list. For traffic, your most visited pages deserve more review than the long tail. For language, if your team has nobody who reads it, budget for an outside reviewer, because nobody in-house will catch a problem.

Where the person goes in the workflow#

Translating from scratch and reviewing machine output are different jobs with different prices. Reviewing is quicker, but only if the reviewer has the source next to the translation and can see the text where it appears on the page. Without that, review turns into reading a spreadsheet of sentences with no context, which misses the button that no longer fits and the term that changed between pages.

The other requirement is that a person decides what goes live. In LangRelay, Relay can flag problems but cannot approve or publish on its own, and you can choose Ask me first to approve every page yourself.

How pages go live in LangRelay#

LangRelay uses AI translation that follows your Glossary and each language's formality setting. By default a translation goes live as soon as it is ready, and free checks hold a page with untranslated text, the wrong language or a changed price. To approve every page yourself, choose Ask me first.

Approval belongs to a person with permission to approve translations, and it is recorded against the exact version of the page in that language. If the text changes afterwards, through a source edit or someone's correction, the earlier approval does not carry over to the new text, and the translation table marks it Approval outdated until someone looks again. A reviewer who disagrees uses Request changes and writes why. That comment shows on the page it was given on, with the reviewer's name and the date, so the next person sees what was wrong.

You can also ask Relay for a review first. 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 page with no major finding shows "Relay reviewed: looks good". Major findings are things like a changed number, price or name, a missing or added sentence, or text left in the wrong language, and they hold the page until a person looks at it. Minor findings, such as wording that is correct but stiff, appear as notes. Relay never approves. A person does, with one click on Go live in the visual editor when the review passed, or after working through the findings when it did not.

To see it on your own pages, book a demo, or read about the quality controls.

Common questions#

Is human translation always better than machine translation?#

A good translator is more reliable on names, tone, legal meaning and anything cultural. On plain informational text in widely spoken languages, machine output can be close enough that a review costs less than a translation. Which one is better for a page depends on what an error there would cost.

Will AI replace website translators?#

In the workflow we describe here, the person's job moves from translating every sentence to reviewing machine output and owning the glossary. Legal, regulated and brand-defining text still goes to a translator.

Can I mix the two on one site?#

Yes. Use machine translation with review for the bulk of pages and a translator for the few where accuracy carries legal or financial weight. If you are tempted to skip the tooling and paste pages into a chat, read what ChatGPT leaves out first.