AI website translation: glossary, quality review and SEO
AI website translation can produce language versions in a few hours that would take a translator weeks. Speed, however, is not the same as quality. Without a glossary, clear rules and human review, you end up with pages that use the wrong terms, wrong units, unnatural titles and promises you never made in the original.
So should you translate your website with AI? Yes, if you use AI as a first draft inside a controlled process. This article explains which content suits AI translation, what Google’s rules say, a practical example, an implementation process, an error typology for measuring quality and a pre-publication checklist.
When AI translation makes sense and when it does not
Not every text carries the same risk. A wrong product specification confuses customers, an error in your terms and conditions can have legal consequences, and a literally translated slogan looks cheap. Decide by content type, not for the whole website at once.
| Content type | Suitability for AI translation | Human review required |
|---|---|---|
| Product specifications and short descriptions | Suitable if you use a glossary | Sample review and checks of numbers, units and names |
| Blog posts and expert articles | Suitable as a first draft | Full review by a native speaker with subject knowledge |
| Homepage, service pages and slogans | Draft only, often needs rewriting | Copywriter in the target language |
| Terms and conditions, returns, privacy policy | Not suitable without expert review | Lawyer or specialist translator for the market |
| Interface texts, cart and emails | Suitable if variables are preserved | Testing in the actual interface and in a sent email |
| SEO titles, meta descriptions and alt texts | Suitable as a draft | Check length, meaning and the keyword for the market |
What Google says about automatically translated content
Google does not ban AI-assisted content. Its spam policies, however, describe scaled content abuse: generating many pages primarily to manipulate search rankings without adding value for users, no matter how they are created. The examples include using generative AI tools to generate many pages without adding value and pages built from scraped content through automated transformations, including translation (Google: Spam policies).
In its guidance on generative AI content, Google notes that AI outputs may contain inaccuracies and recommends manually fact-checking and reviewing all AI-generated content before publishing. The review also covers the title element, meta description, structured data and image alt text (Google: Using generative AI content).
For multilingual websites, Google recommends separate URLs for each language version instead of switching the language through cookies or browser settings, hreflang annotations and a single language for both content and navigation on a page. It determines page language from the visible content, not from the lang attribute or the URL. It also warns that translating only the surrounding page elements while the main content stays in the original language can create a bad user experience, and advises against automatically redirecting visitors based on their assumed language (Google: Managing multilingual sites).
Example: an online shop expands its catalogue to Czechia and Germany
Illustrative example: a Slovak tool shop has 400 products, 20 categories, terms and conditions and a blog. It wants to sell in Czechia and Germany. Instead of one bulk translation, it splits content by risk. Products and categories go through AI translation with a glossary that fixes the translation of technical terms, plus a do-not-translate list for items such as brands and model designations.
A native speaker reviews the best-selling products and a random sample of the rest. Errors are logged by type, and recurring problems are fixed by updating the glossary rather than correcting each product separately. The terms and conditions are checked by an expert for the market and the homepage is written by a copywriter. The German version first publishes the categories with the highest demand and only then the full catalogue.
Step-by-step implementation of AI translation
1. Content inventory and priorities
List page types, the number of texts and their business importance. Assign each type a review level based on the table above. Publish first the sections that actually support sales or enquiries in the new market.
2. Glossary and do-not-translate list
A glossary is a table of terms and their mandatory translations for each language pair. The DeepL API documentation states that glossaries let you specify how particular words and phrases should be translated and that they suit standardising technical terms and product information (DeepL: Glossaries in the Real World). The same principle works with language models if you include the glossary in the instructions and check the output automatically.
3. Style guide for each language
Define the form of address (formal or informal), tone and formats for dates, currencies, decimal numbers and units of measurement. Without these rules, every translation batch will sound different.
4. Translation with protected formatting
Translate structured content, not text copied from the page. Template variables, HTML tags, links, product codes and prices must stay untouched. After translation, automatically compare the number of headings, links and variables with the original.
5. Native speaker review
The reviewer checks meaning against the original, not just fluency. A fluent translation can still be wrong. Errors are logged in a simple table by type and severity.
6. Localising SEO and metadata
Translating a keyword is not enough; you need to check how people in that market actually search. Localise titles, meta descriptions, slugs, alt texts and link texts. Internal links must point to pages in the same language.
7. Publishing and maintenance
Set up separate URLs, hreflang and a visible language switcher. When the original changes, for example a service price or a technical specification, the system must flag outdated translations. Otherwise the language versions gradually drift apart.
How to measure translation quality
A feeling that „it sounds fine“ is not a metric. The MQM (Multidimensional Quality Metrics) framework groups translation errors into seven high-level dimensions and recommends choosing the level of detail that fits your environment (MQM Council: The MQM Error Typology):
- Terminology – a term does not match the company or domain glossary.
- Accuracy – added, omitted or mistranslated content; MQM also places machine translation hallucination here.
- Linguistic conventions – grammar, spelling and punctuation.
- Style – the text is grammatically correct but breaks the style guide.
- Locale conventions – formats of dates, numbers, currency or addresses.
- Audience appropriateness – content is invalid or inappropriate for the target market.
- Design and markup – broken formatting, overflowing text or damaged tags.
For websites, we recommend tracking terminology, accuracy and locale conventions in particular. If the same type of error keeps recurring, fix the cause in the glossary, instructions or template.
Benefits and tradeoffs
- Speed and cost: AI shortens the first draft, but review and SEO localisation remain human work.
- Consistency: a glossary unifies terms across thousands of texts, but a wrong glossary entry spreads everywhere just as quickly.
- Scope: you can translate the whole website at once, but a smaller, well-reviewed scope is worth more than hundreds of weak pages.
- Maintenance: every language version is more content that has to be updated whenever the original changes.
Common mistakes
- Publishing translations without review, especially for prices, technical specifications and legal texts.
- Literally translated keywords that nobody in the target market searches for.
- Translated menu and footer while the main content remains in the original language.
- Automatic redirects based on browser language that block access to other versions.
- Untranslated metadata, alt texts or order emails.
- No rule for what happens when the original changes.
Pre-publication checklist
- Every content type has an assigned review level.
- The glossary and do-not-translate list are approved for each language pair.
- Numbers, units, prices and product codes match the original.
- A native speaker has checked meaning, not just fluency.
- Titles, meta descriptions, slugs and alt texts are localised.
- Each language version has its own URL, hreflang and links in the same language.
- Legal texts have passed expert review for the market.
- It is clear who updates translations, and when, after the original changes.
Frequently asked questions
Does Google penalise an automatically translated website?
Using AI is not a policy violation in itself. The problem is content produced at scale without adding value for users. A reviewed, localised translation that helps customers in a given market is a different situation from thousands of unchecked pages.
Is another AI model enough to check a translation?
For automated checks, yes, for example comparing numbers, variables and glossary terms. For important content, meaning, tone and market fit should be confirmed by a person who speaks the language natively.
Should URLs be translated as well?
Google allows localised words in URLs. For readability and consistency, we recommend localised slugs without diacritics. If you change existing URLs, however, do not forget redirects.
How often should translations be updated?
Whenever a fact changes in the original: a price, a specification, a condition or a contact detail. The best approach is tracking that automatically flags related translations for review when the original is edited.
Next step
Start with a content inventory and a glossary for your most important market. If you want AI translation built directly into your website or online shop, including the glossary, checks and approvals, see our custom AI agents and process automation services. You will find more practical guides on our blog. Write to us via the contact form and we will propose an approach based on your languages and content volume.
Sources
- Google Search Central – Spam policies for Google web search
- Google Search Central – Google Search’s guidance on using generative AI content on your website
- Google Search Central – Managing multi-regional and multilingual sites
- MQM Council – The MQM Error Typology
- DeepL Documentation – Glossaries in the Real World