UX Survey Design

When To Use

A "quick survey" is about to go out with questions that start "don't you agree". Run it before anything goes into the survey tool. It answers: what one decision does this survey feed, and will each question measure what people do and think rather than what they will politely agree to?

When Not To Use

If you need to know why people do something, a survey will not tell you; use User Interview Guide. To analyse responses already in, use Survey Results Analysis; for SUS or SEQ inside test sessions, use Usability Benchmark Study.

Inputs

• The decision the survey serves, who decides, and who will be asked (how they are reached)
• Your draft questions, if any If you have none of this, I start from the decision and draft no more questions than it needs, marked as a first draft.

Approach

Question writing as the Pew Research Center sets it out in Writing Survey Questions: one idea per question, neutral wording, balanced options, attention to question order, and pretesting. One change for product surveys: plan the analysis first, so every question has a known use. The failure it prevents: "How satisfied are you with the speed and reliability of the app?", one number, two questions, and nobody can tell which one moved.

Workflow

1. Ask up to three questions: the one decision this survey informs, who will answer and how they are reached, and what you plan to compare (groups, before and after).
2. Plan the analysis first: for each question, which comparison or chart it feeds. A question with no planned use is cut. If there are several decisions, that is several surveys or another method.
3. One construct per question: split double-barrelled items ("fast and reliable" is two questions). Ask behaviour as frequency ranges with a clear time frame.
4. Neutral wording: replace agree-disagree statements with a choice between alternatives, since agree-disagree invites agreement (Pew). Balanced scales with the midpoint included or left out on purpose; options randomised where their order could bias.
5. Open or closed: closed when the options are known and complete, one open question at most. The format changes the answers, so do not switch it between rounds.
6. Order: screen-outs first, behaviour before attitude, general before specific, sensitive items late. Check where an earlier question could colour a later one.
7. Pilot with [n set by you] people: where did they hesitate or guess? Fix, then field. Write what the sample can say: a convenience sample has no true margin of error, so results are hints about those who answered.

Output Format

# UX Survey Questionnaire
**Decision served:** [decision, owner, date] | **Who is asked:** [population, how reached] | **Length cap:** [n questions]
## Analysis plan
| # | Question | Construct | Feeds which comparison | Keep or cut |
|---|---|---|---|---|
| 1 | [question] | [one construct] | [comparison] | [keep / cut: no use] |
## Questionnaire in order
| # | Section | Wording | Type | Options |
|---|---|---|---|---|
| S1 | Screen-out | [behaviour question] | [closed] | [options; screen out if] |
| 1 | Behaviour | [question with time frame] | [frequency range] | [options] |
| 2 | Attitude | [choice between alternatives] | [balanced scale] | [options, randomised: yes / no] |
## Wording fixes
| Original | Problem | Rewrite |
|---|---|---|
| [original] | [double-barrelled / leading / agree-disagree] | [rewrite] |
## Pilot and sample
[Pilot: who, date, what changed] [Sample: who answers, what it can and cannot say]
## Decision
[Decision owner] approves the final questionnaire and fielding dates by [date].

Done When

• Every question maps to a planned comparison, and every unmapped question is cut
• No question is double-barrelled, leading or agree-disagree
• The order runs screen-outs, behaviour, attitude, sensitive last, and a pilot is planned
• The sample note says what the results can and cannot claim

Quality Bar

• No open question invites naming colleagues or customers
• Demographic items only where the decision needs them, never by default
• No expected response rates or sample sizes are supplied; you set them
• Claude writes the questions; only real responses become results, and none are simulated to preview them

Next

Run uxr-survey-analysis (Survey Results Analysis) to analyse the responses honestly.

About the makers

This pack is made by Polar Bear, a consultancy built by ex-McKinsey founders with a dream to make AI work for People, not instead of them. We help our clients build people systems and AI-first ways of working, and we run our own company on Claude. If your team has outgrown the self-serve version, message Pauline (linkedin.com/in/paulinebertry).