AI translation of product content: glossaries, review and tests

Is AI translation of product content ready for production?

AI translation of product content is ready for production with a glossary and native post-editing. PIMvendors rates it as production-ready with review in the AI in PIM maturity table. The reason is that errors are visible: a native speaker reading the output spots most of them quickly.

The business case is straightforward for European companies. The European Union has 24 official languages. A brand selling in the Benelux, DACH, France and the Nordics needs at least seven languages. Each channel then needs its own texts in each language. Human translation of every text at that scale is slow and expensive.

AI translation in a PIM takes two forms. Some PIMs connect to a machine translation service and send attribute text through it. Others use a large language model that can also apply tone and channel rules. Both forms need the same controls: a glossary, protected terms, a draft status and a reviewer.

Three conditions decide success:

• Closed value lists are translated once, as master data, outside the AI flow.
• Product terminology sits in a glossary that the translation step must apply.
• Safety and legal texts are excluded from AI translation.

Which product content suits AI translation?

Descriptive and marketing text suits AI translation; safety, legal and regulatory text does not. The table shows how PIMvendors classifies typical PIM content.

Closed value lists cause the most avoidable errors. If a model translates "anthracite" separately for each of 3,000 products, the catalogue ends up with several terms for one colour. Webshop filters then split one value into many. Translate the list once and let every product reference it.

Units belong to formatting rules. A PIM should store a canonical value and unit, then format decimals and separators per locale on output.

How do glossaries and locked terms work in a PIM?

A glossary forces the translation step to use approved terms, and locked terms stop it from translating names at all. Both should be configurable per language in the PIM or the connected translation service.

A practical glossary for product content contains four kinds of entries:

1. Brand and product line names that must stay untranslated.
2. Technical terms with one approved translation per language, such as product types.
3. Terms that differ between markets with the same language, such as Dutch in the Netherlands and Belgium.
4. Terms the company has decided never to use, for example a competitor's trademark used as a generic word.

The glossary needs an owner. Without an owner, reviewers fix the same term in every text and the model repeats the error next time. With an owner, each correction becomes a glossary entry and the edit rate falls over time.

Ask vendors whether glossary corrections from reviewers feed back into the glossary automatically or need manual entry. Ask too whether the glossary applies to every AI feature or to translation only. A description generator that ignores the glossary undoes the translation controls.

How do you measure AI translation quality?

Measure AI translation quality per language with human edit rate, terminology adherence and critical errors. A single quality score across languages hides the weak ones.

Post-editing deserves a defined standard. ISO 18587:2017 sets requirements for full human post-editing of machine translation output and for post-editor competences. ISO lists the standard as under revision, so check the current status before you write it into a contract. Even without certification, its distinction between raw output and full post-editing is useful for briefing reviewers.

Use the same reviewers and instructions for every vendor in a test. Different reviewers produce different edit rates on the same text.

How does AI translation fit the PIM workflow?

AI translation fits the PIM workflow as a step between approved source text and approved localised text. The PIM tracks status per attribute and per locale, so nothing publishes untranslated or unreviewed.

A workable sequence looks like this:

1. The source language text is approved.
2. The PIM triggers translation for the locales each channel requires.
3. Output lands in draft status per locale.
4. A native reviewer post-edits and approves.
5. Channel completeness rules check every required locale before export.

Change handling matters as much as first translation. When the source text changes, the PIM should flag every dependent translation as outdated. Without that flag, markets publish old information next to new specifications.

Translation memory reduces cost and edit effort. If the PIM or the connected service stores approved segments, repeated sentences across product variants are reused. Ask vendors where translation memory sits and whether you can export it when you leave. The process requirements page covers the workflow and status requirements in more detail.

What should you ask PIM vendors about AI translation?

Ask vendors to translate a sample of your products into your languages, then measure edit rate and terminology adherence. Language lists in brochures say nothing about quality in your product category.

Requirements to include:

• Translation status per attribute and per locale, with outdated flags on source change.
• Glossary and do-not-translate lists per language, applied to all AI features.
• Attribute groups that can be excluded from AI translation.
• Integration with an external translation service or agency, if you use one.
• Export of translation memory and glossaries in a standard format.
• Stated price per item, per language and per regeneration.

Combine the translation test with AI channel-specific descriptions, because many teams generate in one language and translate the result. Score both tasks in the AI Claims Lab protocol, which includes translation as one of four tasks on 200 buyer-provided products. Choose at least two target languages: one major market language and one smaller language where model quality tends to vary more.