How to Design a Survey or Questionnaire

A survey is only as good as its questions — a single leading or ambiguous item can quietly wreck your data before you collect a single response.

A questionnaire is a research instrument, and designing one well is a skill in itself. Bad questions produce bad data, and no amount of clever analysis can rescue answers to a question people misunderstood. The goal is to gather accurate, relevant information while asking as little of your respondents as possible, because every extra question costs you attention and honesty.

Start from your research objective

Before writing any question, be clear about what you need to find out. Every item on the survey should earn its place by mapping to your research question — if you can’t say what a question is for, cut it. Long surveys full of “might be interesting” questions lower response rates and dilute your data. A quick exercise: for each question, write down which part of your analysis will use it. Anything without an answer gets removed.

Write clear, single-barrelled questions

Each question should ask exactly one thing in plain language. Avoid the common traps:

• Double-barrelled: “How satisfied are you with the price and quality?” — split price and quality into two questions.
• Leading: “How much did you enjoy our excellent service?” — the wording pushes an answer.
• Ambiguous: “Do you exercise regularly?” — define “regularly,” or ask how many times per week.
• Jargon or double negatives — anything that makes the respondent work to decode the question.

Write for your least-expert respondent. If a question could be read two ways by anyone, it will be, and you’ll never know which reading a given answer reflects.

Choose the right response format

Match the answer format to the question. Closed questions (multiple choice, yes/no, rating scales) are easy to analyze and best for measurable data; open questions capture nuance but take effort to code. For attitudes, a balanced scale — for example, a five-point Likert scale from “strongly disagree” to “strongly agree” — works well. Make sure options don’t overlap and cover all reasonable answers, including “not applicable” or “prefer not to say” where needed, so respondents are never forced into an answer that isn’t true.

Order and structure the survey

Open with easy, engaging questions to build momentum, group related items together, and place sensitive or demographic questions near the end. A logical flow keeps respondents moving and reduces drop-off. Watch for order effects, too — an earlier question can prime how people answer a later one — and keep the whole thing as short as your objective allows, because respect for the respondent’s time is also good methodology.

Pilot before you launch

Test the survey on a small group before sending it out for real. A pilot reveals confusing wording, broken skip logic, missing answer options, and questions that everyone answers the same way (and so tell you nothing useful). Ask your testers not just to answer but to say where they hesitated or felt unsure. Fixing these problems after you’ve collected hundreds of responses is impossible; fixing them in a pilot takes an afternoon.

Mind ethics and consent

If your survey is for research, participants should know its purpose, that participation is voluntary, and how their data will be used and protected — usually through a brief consent statement at the start. Anonymity and clear data handling aren’t just ethical requirements; they also make people more willing to answer honestly, which improves the quality of the very data you’re trying to collect.

Start from what you need to know

Write the research questions before writing any items, then map each item to the question it serves. Any item that does not map to one should be cut, however interesting. Long surveys with unmapped questions are the commonest cause of poor completion rates and of data nobody analyses.

This mapping also protects you at the analysis stage. Knowing in advance which item answers which question tells you what analysis you will run, and reveals when a question has no item capable of answering it.

Choose response formats deliberately

The format determines what analysis is possible, so it should be chosen with the analysis in mind rather than afterwards.

• Closed categorical — fast to answer and analyse, but only captures options you anticipated.
• Likert scales — good for attitudes; decide deliberately whether to include a midpoint.
• Numeric — precise, but people estimate poorly for frequencies and durations.
• Open text — rich, but costly to analyse and often left blank.

Mixing a few open items into a mostly closed instrument usually gives the best return: the closed items carry the analysis, the open ones explain anomalies.

Write items that mean one thing

Several faults recur often enough to check for specifically. Double-barrelled items ask two things at once — “the service was quick and friendly” cannot be answered by someone who found it quick and rude. Leading items suggest an answer. Loaded terms carry judgement. Negations invite misreading, and double negations guarantee it.

Assume less knowledge than you think. Technical terms, internal jargon and acronyms all reduce response quality, and respondents rarely admit they did not understand — they guess, and the guess enters your dataset indistinguishable from a real answer.

Make response options exhaustive and exclusive

Every respondent must find exactly one option that fits. Overlapping age bands are the classic failure: someone aged 30 offered “20–30” and “30–40” chooses arbitrarily, and the resulting data is unusable. Provide “other”, “prefer not to say” or “not applicable” where genuine, and be aware that omitting them forces false answers rather than producing complete data.

Order questions to keep people answering

Open with easy, clearly relevant items to establish that the survey is worth completing. Group related items together. Place sensitive questions — income, health, anything potentially embarrassing — late, once some commitment has been built. Demographics conventionally go at the end for the same reason.

Watch for order effects: an early question can prime later answers. Where that risk is real, randomising item order within a block is a standard remedy.

Pilot before you launch

Pilot with five to ten people resembling your intended respondents, and ask them to think aloud as they answer. This surfaces problems no amount of rereading will: items understood differently from how you intended, options that do not fit, and the point at which people begin to lose patience. Time the completion, and treat your estimate as optimistic.

Fixing an item after launch is usually impossible without discarding the responses already collected, which makes the pilot the highest-value hour in the whole process.