Volver al blog
Guía9 min de lectura

How AI Translation Technology Works: Seven Checks Before You Publish

AI translation restates whole-sentence meaning, but quality varies by language and SEO setup is separate. Seven checks to make before you publish.

Equipo Lantis

AI translation, as Lokalise describes it, reads a whole sentence, works out what it means and restates that meaning in the target language, using neural networks trained on very large collections of sentence pairs. Output is strongest in widely resourced languages and weaker for less-resourced languages and non-English source text, so human review still decides the final result. On a website the technology produces the translated text, while URLs, hreflang tags and sitemaps are separate tasks the publisher must check.

  • Neural translation evaluates the whole input sentence at each step instead of splitting it into isolated words or phrases, according to Built In.
  • Quality depends on the data behind the model. Lokalise says models learn from millions, sometimes billions, of sentence pairs, so test the output in your own language pairs.
  • Post-editing can save time, but not always. A Frontiers study on literary text measured productivity gains of 36% for neural post-editing, and that text type differs from product pages.
  • Translating the copy does not finish the job. Verify URLs, hreflang tags, canonicals and sitemaps on every language version.

How does AI translation work?

Older machine translation relied on fixed rules and dictionaries. Lokalise contrasts it with AI tools that use neural-network algorithms. A neural system does not look words up one by one. It processes the full sentence and writes a new one meant to carry the same meaning.

Illustration about “How does AI translation work?”

Three families of tools come up in buying conversations:

ApproachHow it worksWhere the evidence shows limits
Rule-basedApplies fixed rules and dictionariesContrasted by Lokalise with neural tools
Neural machine translation (NMT)Evaluates the whole input sentence at each stepNeeds large volumes of parallel bilingual data, which are not equally available across the world's 7,000+ languages
Large language models (LLMs)General language models used to translateFluent, but weaker than human translators at lexical precision and terminology consistency; quality drops with non-English source text

Seven checks before you rely on AI translation

Each check below meets two conditions: it changes what ends up published, and either published research or Google's documentation speaks to it. Every item follows the same pattern: what the evidence says, then what to verify.

Does quality hold for your language pair?

  • Evidence: A Springer article on LLM-based machine translation reports that GPT-4 had the highest English-to-German quality among the compared setups, with a COMET score of 87.44 and a BLEU score of 35.38. A study cited in an arXiv paper (Zhu et al., 2023) found that GPT-4 still lags commercial systems such as Google Translate, particularly for low-resource languages.
  • What to check: Results change with the study, the pair and the setup. Translate a sample in each pair you need before judging a tool.

Is the source language English?

  • Evidence: According to Acolad's own comparison, a translation company with an interest in the result, a well-trained NMT model outperforms generative AI on most key metrics. Acolad also says LLM quality tapers off quickly when the source language is not English and for less well-resourced languages in general.
  • What to check: If your original site is in Spanish, German or another language, run the test from that original rather than from an English version.

Is the target language well resourced?

  • Evidence: The NLLB paper from Meta, published in PMC, states that high-quality NMT needs large volumes of parallel bilingual data, unevenly available across languages.
  • What to check: A long list of supported languages says what is offered, not how good each one is. Have a native speaker read samples in every language you plan to publish.

Does the terminology stay consistent?

  • Evidence: A 2024 study cited in the ResearchGate paper finds that LLMs write fluently but lack lexical specificity and consistency of terminology compared with human translators.
  • What to check: Build a glossary of product names, legal wording and brand terms, plus a list of text that must stay untranslated. Then scan the output for the same term rendered two different ways.

Does human review save time?

  • Evidence: A survey of post-editing research by Koponen reports that productivity findings are partly conflicting: some studies show notable gains, others no significant difference. It also warns that editing poor machine translation remains an unproductive task. The Frontiers study measured gains of 18% for statistical MT post-editing and 36% for neural MT post-editing against translating from scratch, on literary text.
  • What to check: Literary text is a different genre from product pages or support articles. Time the editing of a sample against translating the same sample from scratch, and decide how much review each page type needs.

Will Google treat translated pages as duplicates?

"Localized versions of a page are only considered duplicates if the main content of the page remains untranslated." — Google, in its localized versions documentation

  • Evidence: A fully translated page is not treated as a copy of the original. The risk lies in pages that keep their main content in the source language. Unframed Digital observes that most companies translate the copy and stop, skipping keyword research, hreflang setup and cultural fit, which decide whether a market converts.
  • What to check: Look for untranslated blocks on each language version, and plan keyword research for each market instead of translating the source keywords literally.

Are URLs, hreflang and sitemaps in place?

  • Evidence: Google's documentation says that with three versions of a page, the sitemap has entries for the URLs of each version. Strapi's guide notes that hreflang requires absolute URLs. Squirrelscan's glossary advises keeping each version canonical to itself, not to the primary language, because the latter undoes the hreflang relationship. The x-default value names the fallback page for visitors whose language settings match none of your versions. Google also says IP location analysis is difficult and generally not reliable, and recommends explicit methods such as hreflang, alternate URLs and explicit links. Seek Marketing Partners recommends XML sitemaps as the hreflang method best suited to large sites with many multilingual pages.
  • What to check: Open the source of one translated page and confirm it has its own URL, a self-referencing canonical and absolute hreflang links including x-default, and that it appears in the sitemap. Also check your platform: Blu Mint reports that Squarespace does not natively support hreflang at scale, so the agency built the tags manually across 40 languages on one client site.

How does Lantis handle these checks?

Lantis translates websites on Framer, Webflow, Shopify, WordPress and any other site. It publishes one URL per language and generates hreflang and sitemaps automatically, and installation means pasting one line of code, with no DNS changes and no design edits. Translation runs on Gemini by default, with Claude, DeepL or DeepSeek as alternatives. The visual editor, the per-language glossary and the exclusion of pages or blocks correspond to the terminology and review checks above.

Lantis includes 500,000 words per language and offers more than 60 languages, charging only for active languages: €3/month for one language, €2.50 per language per month for 2 to 9, and €2 per language per month from 10. The language-quality checks above apply to every one of those languages, so the output for each still needs a native-speaker read.

For platform-specific setup, Lantis has a guide to translating a Framer site, a comparison of how Webflow translation tools bill and a Shopify hreflang walkthrough. Its article on language and hreflang setup for Middle East luxury sites shows how the checks apply to one market.

Next steps

  • Pick three page types (a product page, a legal page and a blog post) and have a native speaker mark errors in the machine translation of each, in every language pair you need.
  • Write a glossary of brand, product and legal terms, plus a list of text that must stay untranslated, before generating any translation.
  • Time the post-editing of one sample page against translating it from scratch, and use that result to set review depth per page type.
  • Open the source of one translated page and confirm its own URL, self-referencing canonical, absolute hreflang links with x-default, and sitemap entry.
  • If you want URLs per language, hreflang and sitemaps generated for you, Lantis does it this way: one line of code, one URL per language, and the details in its guide to what Google documents about hreflang.

Frequently asked questions

Is AI translation accurate enough for a commercial website without human review?

It depends on the language pair and the content. Research finds LLMs fluent but weaker than human translators on terminology consistency, and Acolad reports that quality falls for non-English source text and less-resourced languages. Pages with prices, legal wording or product claims warrant a full human check.

Does AI translation hurt SEO, or does Google treat translated pages as duplicate content?

Google states that localized versions count as duplicates only when the main content stays untranslated, so fully translated pages are not treated as copies. The SEO risk comes from skipped steps: Unframed Digital notes that many companies translate the copy and omit keyword research and hreflang setup.

Is machine translation post-editing faster than translating from scratch?

Sometimes. Koponen's survey describes the productivity results as partly conflicting, and the Frontiers study on literary text measured gains of 18% for statistical and 36% for neural post-editing. Editing poor raw output remains unproductive, so time a sample from your own content.

Do I need separate URLs for each language, or can I use a language switcher?

Google says IP location analysis is difficult and generally not reliable, and recommends explicit methods such as hreflang, alternate URLs and explicit links. Any of those methods makes the relationship between language versions clear to search engines, so choose the one your platform supports and check it on every language version.

Sources

Comprobador de SEO multilingüePega la dirección de cualquier página de tu web. La leemos como lo hace Google y revisamos hreflang, lang, canonical y cada versión traducida, con la solución a cada problema.

Pon tu web en varios idiomas hoy.

Una línea de código. SEO real por idioma. Traducción con IA. Sin tarjeta para empezar.

Sigue leyendo