In-Product Survey Translation

Quick answer — Survey translation has to preserve measurement, not just meaning. Rating scales, agreement wording and question order all affect responses, so a loosely translated survey produces numbers that cannot be compared across markets.

Vitra.ai Universe ships product and customer content in every language.

The output is data, not content

Most translated content is judged on whether it reads well. A survey is judged on whether the answers mean the same thing.

A question phrased slightly more positively in one language produces a higher score, and nobody can tell afterwards whether that market is happier or the question was different.

Scales are where it breaks

Negatively worded questions are the classic failure. An item asking whether the respondent would not recommend the product, translated into a language that handles negation differently, can invert the response while the data still looks plausible. Avoid negative phrasing in the source and the problem disappears for every language at once. It is one of the few localization defects cheaper to prevent than to detect, because detecting it means noticing one market's numbers are implausible and working backwards through the instrument to find out why.

NPS is a specific case

The eleven-point scale and the wording of the standard question are what make scores comparable, and there are established localized versions.

Use those rather than translating the question yourself. An in-house translation of the NPS question produces a number that is not comparable with anyone else's, which removes the only reason to use it.

Adapt what is about context, not measurement

Examples, role names and company-size bands can and should be localized — company-size categories that make sense in one market describe almost nobody in another.

The measurement instrument stays fixed. The context around it adapts.

Test before shipping

Run the translated survey past two or three native speakers in the target market and ask what each question is asking, rather than whether it reads well. Divergent answers mean the question moved.

Hold the wording in memory so the same question is asked identically next quarter, and run quality control over the set — comparability across waves is worth as much as comparability across markets. The same rule governs public statistics — census and survey translation.

FAQ

What makes survey translation different? The output is data. A question phrased slightly more positively in one language raises the score, and afterwards nobody can tell whether that market is happier or the question was different.

Why are rating scales the main risk? Because scale labels rarely have exact equivalents. Endpoint wording shifts the distribution, some languages lack a natural neutral midpoint, and frequency words map differently.

Should the NPS question be translated in-house? No. Established localized versions exist, and an in-house translation produces a score that is not comparable with anyone else's, which removes the point of using a standard metric.

How should a translated survey be tested? Ask two or three native speakers what each question is asking, rather than whether it reads well. Divergent answers mean the question moved and the data will not be comparable.