Machine vs Human vs Hybrid Website Translation
Choose the translation method by what an error would cost on that page, not by a site-wide rule. Machine translation is fast and cheap enough for low-risk, high-volume content; human translation suits high-stakes or brand-defining pages; hybrid, a machine draft followed by human review, covers most of the middle. Whatever you choose, a glossary, a do-not-translate list and a clear review step matter more than the label.
What are the three approaches?
Machine translation (MT) produces text automatically with no human involvement in the output. It is quick, consistent in speed and easy to re-run when source content changes.
Human translation is done by a person fluent in both languages, often a professional translator, sometimes followed by a second reviewer. It takes longer and costs more per word, but a person can handle tone, idiom, cultural references and ambiguity.
Hybrid translation starts with MT and has people review or edit the result, commonly called post-editing. The depth of review ranges from a light check for clear errors to a full edit equivalent to a human rewrite.
We are deliberately not quoting prices or turnaround figures. They vary by language pair, subject matter, volume, vendor and service level, and we have no sourced benchmark to cite. Get quotes for your own content.
How do the trade-offs compare?
| Factor | Machine | Hybrid | Human |
|---|---|---|---|
| Speed | Fastest | Between | Slowest |
| Cost per word | Lowest | Between | Highest |
| Handles tone and culture | Weakest, varies by language | Good if reviewer is skilled | Strongest |
| Consistency of terminology | Needs glossary to enforce | Glossary plus review | Depends on translator and memory |
| Scales to frequent updates | Easy | Review becomes the bottleneck | Hardest |
| Risk of confident errors | Present | Reduced by review | Lower, not zero |
Quality is not a fixed property of the method. MT quality differs widely across language pairs and subject areas, and human quality depends on the individual and the brief. Treat these rows as tendencies, not guarantees.
Which approach fits which page type?
| Page type | Suggested approach | Reason |
|---|---|---|
| Legal terms, privacy policy | Human, ideally with legal review | Errors can have legal consequences |
| Pricing, checkout, shipping and returns | Hybrid with full review, or human | Wrong numbers or conditions cause real harm |
| Medical, safety, financial advice | Human with subject-matter review | High cost of error |
| Home page, landing pages, brand copy | Human, or hybrid with a strong editor | Tone and persuasion matter |
| Product pages (large catalog) | Hybrid: MT plus sampled review; human for top sellers | Volume is high; see translating ecommerce product pages |
| Blog and help articles | Hybrid, or MT with spot checks | Meaning matters; polish is secondary |
| User reviews, comments | MT, clearly labeled | Volume high, value is in the gist |
| UI strings (buttons, errors) | Human or hybrid with context | Short strings lack context and are easy to mistranslate |
| Internal documents | MT | Low public risk |
This is a starting framework. Adjust it to your market, your regulator and your brand.
What does Google say that affects the choice?
Google's spam policies treat pages generated at scale mainly to manipulate rankings as scaled content abuse, and list automated translation among the transformations that can be used that way. Google's guidance on AI-generated content advises checking output for accuracy before publishing. Neither says machine translation is forbidden. We cover this in detail in does Google penalize machine-translated content. In practice, the lower your review level, the more you should limit what you index.
What does a good review workflow look like?
- Triage by risk. Tag pages high, medium or low using the table above.
- Prepare the source. Clean, unambiguous source text translates better. Remove idioms where you can, and fix typos before translating.
- Apply glossary and rules. Load approved terms and do-not-translate entries before the machine pass.
- Machine draft. Generate the translation.
- Review. Full review for high-risk pages, sampling for medium, spot checks for low. Reviewers should see the page in context, not only isolated strings.
- Record corrections. Add recurring fixes to the glossary or translation memory.
- Publish status. Keep unreviewed pages out of the search index (for example with
noindex) until they pass. - Re-review on change. When source text changes, flag the affected translations rather than silently leaving them stale.
How do you sample for quality?
Pick a number of segments from each batch, have a fluent reviewer score them against a short rubric (accuracy, terminology, fluency, formatting, locale conventions), and set a pass threshold you decide in advance. We do not suggest a universal sample size or pass rate, because none that we can source applies to all sites.
How do glossaries and do-not-translate lists work?
A glossary maps source terms to approved target terms per language, for example your product category names or a feature name that must read the same everywhere. It enforces consistency and cuts review time because the same correction is made once.
A do-not-translate list covers brand names, product names, model numbers, code and anything else that should remain as written. In HTML you can mark such text with the translate attribute; MDN notes translate="no" is respected by automatic translation systems such as Google Translate and may be respected by tools used by human translators:
<p>Install <span translate="no">ResilientLang</span> from your dashboard.</p>
Also keep a style guide per language: formal or informal address, date, number and currency formats, and tone. Languages such as German, Japanese or Spanish have choices (formal "you", regional vocabulary) that a machine cannot infer from your brand.
Keep these assets in a portable format you own, so that changing vendors does not mean losing your accumulated decisions. See avoiding translation vendor lock-in.
What else should you localize beyond the words?
- Titles, meta descriptions, image alt text and structured data.
- URLs and slugs where appropriate.
- Currency, units, dates and phone formats.
- Images that contain text.
- Legal and contact information for each market.
Google's guidance warns that translating only boilerplate while leaving main content in another language creates a poor experience, so aim to translate pages completely or not index them yet.
How should you decide?
A short decision path:
- What happens if this page is wrong? If the answer involves law, money, health or reputation, start at human or full-review hybrid.
- How often does the content change? Frequent changes favor MT plus targeted review.
- How many pages? Large catalogs favor hybrid with tiers.
- Who is the audience, and how much do they expect polish?
- What can you maintain? An unmaintained translation is worse than none.
We are designing ResilientLang to support review states and do-not-translate rules, but it is not released and this article does not depend on any product.
Key takeaways
- Choose per page by cost of error, volume and change frequency, not one method for the whole site.
- Hybrid (machine draft plus human review) fits most mid-risk content; human fits high-stakes and brand copy; machine fits low-risk, high-volume content.
- Do not rely on invented numbers: get quotes and test with your own content and languages.
- Glossaries, do-not-translate lists and style guides improve every approach and should be portable.
- Keep unreviewed pages out of the index until they pass review.
Sources
- Spam policies for Google web search – scaled content abuse and automated translation as a transformation technique
- Google Search's guidance on using generative AI content – checking automatically generated content for accuracy and value
- Managing multi-regional and multilingual sites – warning about translating only boilerplate
- MDN: translate global attribute – marking content that should not be translated
Frequently asked questions
When is machine translation good enough for a website?
It is a reasonable fit for low-risk, high-volume, quickly changing content such as user reviews, support articles or internal documentation, where meaning matters more than polish. For legal, medical, pricing, checkout and brand-defining pages, the cost of an error is higher, so add human review.
What is hybrid translation?
Hybrid translation means a machine produces the first draft and a person reviews or edits it, often called post-editing. Reviewers may check every segment or sample them depending on risk. A glossary and translation memory carry decisions forward so the same corrections are not repeated.
Is human translation always better?
Not automatically. A skilled human translator usually handles tone, culture and ambiguity better, but quality depends on the person, the brief and the review process. Humans are also slower and cost more per word, and they need good source text and context to do well.
What is a do-not-translate list?
It is a list of terms that must stay in their original form, such as brand names, product names, code, SKUs and certain technical terms. Giving it to your translation workflow, and marking such text in markup, prevents the most embarrassing and costly machine errors.