"We need an English version for our European customers." "Can we run this for our APAC team too?" Demand for multilingual surveys keeps growing as teams expand globally.
But "just translate it" produces unreliable data. This post covers the design process and the cultural considerations that make multilingual survey results actually comparable.
Translation ≠ localization
| Translation | Localization | |
|---|---|---|
| Scope | Words | Culture and context |
| Process | Convert to another language | Adapt to another culture |
| Output | Same meaning, different words | Same intent, naturally received in the new culture |
Example:
Source (English): Are you satisfied with our service?
× Translation (French): Êtes-vous satisfait de notre service ?
○ Localization (French): Dans quelle mesure êtes-vous satisfait de notre service ?
(To what extent are you satisfied…)
In some cultures, asking a yes/no question about satisfaction reads as confrontational. Asking about degree lands more naturally and produces more honest answers.
The multilingual survey workflow
Step 1: Nail the source language first
Pick one base language (usually English) and get that version right before anything else.
Lock down:
- Question count and structure
- Likert scale (3-point, 5-point, 7-point)
- Required vs. optional fields
- Order and conditional logic
Sloppiness in the source propagates to every translation. Fix it once, here.
Step 2: Translate
Three quality tiers:
Tier 1: Machine translation (DeepL, Google Translate)
- Pros: fast, cheap
- Cons: misses nuance, no cultural awareness, awkward phrasing
- Use case: internal-only, low-stakes surveys — and even then, expect to lose some signal
Tier 2: Machine translation + human review
- Machine draft, then a native-speaking team member edits
- The pragmatic middle ground for most B2B teams
- Reasonable cost, dramatically better quality than tier 1
Tier 3: Professional translator
- Specialized services (Gengo, Lionbridge, TransPerfect, etc.)
- Pick someone with marketing-translation experience, not legal or technical
- Cost: typically a few hundred USD per 1,000 words
Customer-facing research deserves tier 2 minimum.
Step 3: Back-translation
The standard quality check for cross-cultural research:
Source (English)
↓ Translator A: EN → FR
French version
↓ Translator B (different person): FR → EN
Back-translated English
↓
Compare original to back-translation. Drift = problem.
If the back-translation says something subtly different, the question's intent has shifted. Rewrite and try again.
Step 4: In-market pilot
Have 5–10 native speakers in the target market take it for real.
What you're listening for:
- Does the phrasing feel natural?
- Does the question's intent come through?
- Are answer options culturally appropriate?
- Anything inadvertently offensive or confusing?
Fix issues here, before broad distribution.
Step 5: Launch
Now you can send it. Analyze each language separately by default — see below for why.
The pitfalls that catch most teams
Pitfall 1: Shipping raw machine translation
Source (English): "Did you have any problems with our response?"
Machine to German: "Hatten Sie Probleme mit unserer Antwort?"
What a native would write: "Gab es etwas, womit Sie unzufrieden waren?"
The literal translation is grammatically fine but tonally off. Respondents notice, trust drops, response quality drops with it.
Pitfall 2: Likert scales don't mean the same thing across cultures
5-point satisfaction scales carry well-documented cultural skew:
| Cultural cluster | Tendency |
|---|---|
| East Asia | Central tendency bias — clusters around the midpoint |
| US, Northern Europe | Extreme-response bias — picks ends of scale |
| Middle East, Southern Europe | Acquiescence — skews high |
| Northern Europe (Nordic specifically) | Modesty — skews lower than peers |
A 25-point NPS in Japan and a 25-point NPS in the US do not mean the same thing about your product.
Mitigations:
- Analyze by language/region separately — never aggregate raw scores across cultures
- Maintain culture-specific benchmarks
- Use behavioral questions where you can ("How many times in the past 3 months did you use feature X?") — they're less culturally elastic than attitudinal Likert items
Pitfall 3: Answer options that don't apply locally
Question: What region are you in?
English options: Northeast, Midwest, South, West ← meaningless to a UK respondent
Geography, job titles, industry categorization — all need to be localized per market, not just translated.
Pitfall 4: Privacy law varies by jurisdiction
US: collecting name + email is mostly fine (with notice)
EU: GDPR — explicit, granular consent required
California: CCPA — disclosure and opt-out rights
China: PIPL — explicit consent + cross-border transfer restrictions
Brazil: LGPD — similar to GDPR
The moment you offer the survey in another language, you've potentially extended its reach into a different regulatory regime.
Mitigations:
- Get legal review per jurisdiction
- Maintain a localized privacy notice at the top of the survey
- Be intentional about data residency (where the data is stored, who has access)
Pitfall 5: Date, time, and currency formats
US: 05/11/2026, 12:00 PM, $10
EU: 11/05/2026, 12:00, €10
Japan: 2026/05/11, 12:00, ¥1,000
If the displayed format doesn't match the respondent's expectation, you'll get errors and confusion.
Mitigations:
- Dates: spell out the month or use ISO format (2026-05-11)
- Times: be explicit about AM/PM or use 24-hour
- Currencies: symbol + ISO code (USD, EUR, JPY)
Pitfall 6: Mobile font rendering
Chinese, Arabic, Thai, and other scripts need fonts the default system might not load cleanly. Tofu boxes, broken kerning, and unreadable text are common with naive setups.
Mitigations:
- Load script-specific webfonts dynamically based on locale
- Test on real iOS and Android devices, not just emulators
- For RTL languages (Arabic, Hebrew), the entire layout direction must flip — not just the text
Analyzing multilingual responses
The rule: separate by language by default
× Combine all languages, report a single NPS of 25
○ Report Japanese NPS 22, English NPS 31, Chinese NPS 18 separately
A combined number averages away the cultural and market signal you actually need.
Open text across languages
When free-text responses come in across languages:
1. Cluster themes within each language first
2. Compare the top themes across languages
3. Document language-specific concerns separately
Machine-translating open text before clustering loses the nuance that made the response valuable in the first place. If budget allows, have native speakers do the thematic pass.
Pick the right tier of effort
You don't need to fully localize every survey. Match the rigor to the stakes:
Tier A: Lightweight (internal, low budget)
- Machine translation only
- 4–5 major languages
- Analyze separately
Tier B: Standard (customer-facing, mid budget)
- Machine + human review
- Pilot in each language
- Skip back-translation
Tier C: Rigorous (global flagship research)
- Professional translators
- Back-translation required
- In-market pilots in every language
- Culture-aware question redesign, not just translation
How Repoan helps
Repoan supports multilingual workflows out of the box:
- AI translation assist — generate first-draft translations from your base language
- Per-locale URL routing — automatic language switching by URL parameter
- Per-language dashboards — responses are split by locale by default
- Multi-script font loading — major scripts handled automatically
- Locale-aware formatting — dates, times, currencies adapt to respondent locale
- Region-aware privacy templates — GDPR, CCPA, and PIPL notice patterns built in
Summary
Multilingual surveys require:
- Localization, not just translation
- Machine translation alone is rarely enough for customer-facing research
- Back-translation as your quality check
- Cultural awareness in scale design — never naïvely aggregate
- Per-jurisdiction privacy compliance
- Per-language analysis as the default
The credibility of global research rests less on translation accuracy and more on cultural fit. The distinction between "the words match" and "the intent transferred" is the entire game.