Translation quality

A translated catalog does not stay translated. The German source changes, products are added, terminology drifts, and eighteen months later your French catalog is a mixture of three vintages. Three reports keep that visible.

The three reports

Read all three per locale and per attribute. The aggregate number is for reporting upward; the per-attribute split is what produces work.

Untranslated

The obvious one, and the one that shrinks as a project runs. What matters is which attributes it contains. A locale at 60% untranslated where the missing 60% is long descriptions on tail articles is in good shape. A locale at 20% untranslated where the missing 20% includes option labels is not — because option labels drive the facet panel of the whole catalog.

Watch for it going back up. New products arrive in the source language only, so every import adds untranslated values. If nobody is assigned to the queue, the number climbs quietly for a year.

Machine-only

Values that were machine-drafted and never reviewed. Whether this is a problem depends on a decision you should have made explicitly:

• Names, safety text, anything with a norm or dimension — machine-only is not acceptable. These are what buyers decide on.
• Long descriptions on low-revenue articles — machine-only is a reasonable deliberate choice, provided it is tracked and provided it is not what you send in a catalog export to a customer who will print it.

The risk of machine-only text is not that it reads badly. It is that it reads well and is wrong: a converted unit, a translated norm designation, a dropped negation in a safety note.

Source changed

The most-ignored report and the one that causes the worst failures. Someone rewrites a German description; the French value keeps its old, reviewed state and now says something the product no longer does.

The mechanism: when a source value changes, every reviewed translation of it is flagged needs review and stays visible on the storefront meanwhile. That is the right default — removing the old translation would leave a gap — but it means the flag is the only thing standing between you and a stale claim.

Terminology drift

The slow failure. Three translators over two years produce three words for Verschraubung, and the catalog reads like three catalogs. Buyers notice, and so does search — a facet whose labels drift becomes several facets.

What keeps it in check:

• One glossary, one owner. See Translate the catalog.
• A term audit twice a year. Take the thirty most frequent nouns in your source catalog and check what they became in each language. Half an hour, and it finds drift nothing else will.
• Reviewer continuity. The same person reviewing a language for two years is worth more than a bigger budget spread over four agencies.
• Search analytics per locale. A rising zero-result rate in one language, with no change in the catalog, is usually terminology drift or missing synonyms — see Search analytics.

The monthly routine

1. Read untranslated per locale. Is it going down? If it is going up, nobody owns the queue.
2. Read source changed. Clear anything on names or safety-relevant fields this week; queue the rest.
3. Read machine-only and confirm it only contains the categories you decided were acceptable.
4. Check completeness per locale for the channels you publish and export to. See Completeness.
5. Check the zero-result rate per locale. It is the only measure of translation quality that comes from buyers rather than from your own reports.

What good looks like

• Untranslated falls month over month, or holds steady while the catalog grows.
• Source-changed is cleared within two weeks for names, longer for descriptions, same week for anything safety-relevant.
• Machine-only exists only where you decided it may.
• The gap between your source-language and target-language zero-result rates is narrowing.

Next

• Translate the catalog
• Data quality in practice — the source-language version of the same discipline.