Create a survey — AI picks the right questions

The most common question before every survey: "how many responses do I need?" Short answer: significantly fewer than most think. With a population of 100,000 people, 384 responses suffice for a 95-percent confidence interval at ±5 percent accuracy — at 1 million people this number barely changes.

The formula is manageable: n = (z²·p·(1-p)) / e². With z=1.96 (95% confidence), p=0.5 (worst case) and e=0.05 (±5% tolerance), the result is around 384. For more accuracy (±3%) you need around 1067. For internal trends, 100 responses are often enough — perfect for quick decisions.

More important than the raw number is sample quality. 384 responses from random newsletter subscribers carry less weight than 100 responses from a stratified sample by region, age and customer-since date. A live confidence interval in the dashboard helps interpretation — when the confidence width stabilizes, the survey is statistically robust even without hitting the threshold.

The choice of question type decides whether answers are analyzable. Asking everything open gives rich answers — but means reading thousands of texts. Pre-defining everything yields fast quotas — but risks missing the essential.

For measuring frequencies and comparisons, single-choice (exactly one answer, classic for demographics) and multiple-choice (multiple options, e.g. for used features) work well. For evaluations, Likert scales are ideal: 5 or 7 points, always odd-numbered, with clear endpoints ("not at all" to "fully agree"). Sliders suit estimates, NPS suits loyalty.

Use free text sparingly — at most 1-2 open questions per survey, ideally at the end. More open questions reduce response rate and shorten answers. A modern platform clusters free-text answers via AI automatically — making 500 open answers readable in minutes. Matrix questions (multiple items with the same scale) save space but should contain at most 7 items — otherwise care drops.

The first question of a survey has two jobs: signal to respondents what it is about — and be easy enough not to scare them off. A too-hard first question massively raises abandonment.

Classic example: "how satisfied are you with our service?" as the first question forces immediate judgment — without warm-up. Better is a low-barrier filter question ("which product do you use?") or an open entry question ("what was your last contact with us?"). The hard rating comes at position 2 or 3.

Order effects go deeper. Answers to later questions are influenced by earlier ones (priming, assimilation effect). Whoever is first asked about "wait times" and then general "satisfaction" rates worse — the wait time is salient. Solution: place critical rating questions before specific detail questions, or randomize order (especially for matrix items). Demographic questions belong at the end — not the start. Whoever is immediately asked age and income often abandons.

Translating a survey into multiple languages costs time and care. The question: from when does the effort pay off? Rule of thumb: if more than 10 percent of the target group speaks a different language, response rate without translation drops noticeably.

In international B2C this is almost always the case — even within the DACH region there are Swiss-German-leaning respondents, Austria uses different terms than Germany. In B2B with English-speaking stakeholders, an English variant alongside the local language is often enough. AI-assisted translation with human review per language is the standard today — effort per additional language: 30 to 60 minutes of review.

More important than translation is cultural adaptation. Scales are interpreted differently across cultures — Asian respondents avoid extremes more than Western ones, distorting comparisons. Solution: automatically pick target language by browser language, but filter results by language during analysis. This shows whether a "bad" NPS in a market is cultural or product-related.

Exporting a 500-row CSV to Excel and wrestling with pivot tables is the most common analysis method — and at the same time the most inefficient. Modern platforms deliver five things Excel cannot.

First: automatic visualization per question — bars for single-choice, word clouds for free text, heatmaps for matrix. Second: cross-tabulation by click — "how do existing vs. new customers rate?" without pivot magic. Third: live confidence intervals showing whether a difference is statistically significant (or just noise).

Fourth: AI clustering of free-text answers in minutes, instead of manually reading 500 texts. Fifth: comparison over time — how did NPS shift this quarter vs. last, with trend display. Excel remains useful for special analyses or reporting into existing structures — but as the primary analysis environment it costs more time than it saves. A modern platform combines both worlds: live dashboard for daily work, CSV/JSON export for special cases, and API endpoint for BI tools like Looker or Metabase.