Improve a survey before you send it

Example result.

Draft: “Were the booking instructions clear and the equipment ready?” Problem: two questions in one. Someone can say yes to the first and no to the second. Revision: “How clear were the booking instructions?” Then, separately: “Was the equipment ready at the time you booked it?”

1. Write down what the survey must tell you

Write the question the survey is meant to answer. “Understand where the booking process is difficult” is more useful than “prove that people like the new process”.

Then name who can reasonably answer and the period you’re asking about. Someone who hasn’t used the process needs a not-applicable route, not a rating scale that makes them invent an experience.

Why this works: every revision is judged against this purpose. Without it, the AI can only make questions sound nicer, which is not the same as making them measure something.

What you should see: three short lines (purpose, respondents, period) that you can paste straight into the prompt in step 2.

2. Ask for a focused review

Paste the prompt and your questions into a new chat. If you’re drafting in a form builder, the same review works on what its AI produces:

Review these draft survey questions for the stated purpose and respondents.

Purpose: [what we need to learn]
Respondents: [who can answer]
Time period: [relevant period]
Draft questions and response options: [questions]

Identify leading wording, questions asking about more than one thing, unclear terms, unsuitable answer options and unnecessary questions. For each issue, explain the problem and propose a revision that preserves the research purpose. Include a suitable not-applicable or don't-know route where needed, without treating it as a neutral rating. Do not invent respondent feedback, measured validity or survey results.

The first three brackets are your three lines from step 1. [questions] is the full draft, with every answer option written out.

Why this works: naming the usual defects gives the review a checklist, and asking for a reason with each change makes every suggestion something you can accept or reject. A wording pass improves the draft; it can’t tell you how people will actually answer.

What you should see: one entry per problem question, each with the problem, the reason and a revision, plus a not-applicable option placed outside any rating scale.

3. Decide which changes to keep

Don’t accept the list wholesale. Splitting one combined question gives you two useful measures, but it also makes the survey longer and changes how the results read. Remove a question if you can’t explain how its answer will be used.

For frequency, defined counts (“How many times in the last four weeks?”) are usually easier to interpret than “sometimes” or “often”. Keep a one-line record of each change you make; you’ll need it if anyone compares the new results with an older survey.

Why this works: the AI doesn’t know what you’ll do with the answers. You do, and that’s what decides whether a question earns its place.

What you should see: a shorter list than the AI proposed, with a reason beside each change you kept.

Check the result

• Each question asks about one thing, over a stated period.
• You could answer honestly if one part of the experience was good and another was poor.
• The answer options cover the full range, and “I don’t know” or “I didn’t use it” sits outside the scale rather than as its midpoint.
• The introduction and question order don’t signal the answer you’d like. A neutral question after “We’re delighted to launch our new process” isn’t neutral any more.

The usual failure: the revision keeps a quieter steer. The praise moves from the question into an adjective (“the improved process”), or the scale has three positive options and one negative. The assistant may call its own rewrite neutral; that doesn’t make it so. Read the wording and the options together.

This revision still suggests the desired answer through [wording or options]. Rewrite it neutrally for the stated purpose. Keep the time period and ask about one thing. Offer a balanced response range and a separate option for someone who cannot answer. Explain the change in one sentence.

[wording or options] is the exact word or option that gives the preferred answer away.

Fictional example. A team wants to understand colleagues’ experience of a new equipment-booking process over the last four weeks. Their draft:

1. How much did you love our brilliant new booking process?
   Options: loved it / really loved it
2. Were the booking instructions clear and the equipment ready when you arrived?
   Options: yes / no
3. How often did you use it?
   Options: rarely / sometimes / often

The review caught all three problems, but its first revision of question 1 was “How satisfied were you with the improved booking process?” with options from “satisfied” to “extremely satisfied”. That’s the usual failure: “improved” still praises the process and the scale still has no room for dissatisfaction. After the repair prompt, with “improved” and the all-positive scale named as the problem, the change record read:

Three things to notice:

• The repair named the word. “Make it neutral” would have got another polite rewrite. Pointing at “improved” and the scale fixed both in one go.
• Non-use left the scale. Someone who never booked anything isn’t “neither satisfied nor dissatisfied”, and counting them that way would flatten the result.
• The split kept the useful distinction. Somebody may understand the instructions and still find the equipment missing. One yes/no answer would hide which part needs fixing.

Task kit for this job: Download (ZIP) · Reusable instructions

Go deeper

Practise on a second draft

Run the step 2 prompt on these three fictional questions from a staff survey about a new rota, then compare what it finds with the notes underneath:

1. Do you agree that the new rota is fairer and easier to use?
2. How would you rate our SLA adherence?
3. How many hours did you spend on admin last year?

The first is an agree/disagree statement, which tends to pull people towards “agree”, and it asks about two things (fair, easy). The second uses jargon some respondents won’t share. The third asks for a figure nobody can recall across a year; a shorter period, such as last week, gives a more believable answer. If the review misses any of these, ask about that one directly.

Try the wording with intended respondents

Where your organisation allows it, ask a few people from the intended audience to explain, in their own words, how they understand a draft question and choose an answer. Record what they actually say. Don’t replace that with AI pretending to be several respondents and call it user testing.

For sensitive subjects, agree with the right owner whether each question is necessary, how answers stay confidential and how they’ll be used. Don’t add names or identifying details because an AI template includes them. Government Analysis Function’s questionnaire design guidance covers respondent needs, testing and question design in more depth; the page is currently marked as under review.

Keep an existing survey comparable

If a question or scale changes in a survey you run every quarter, record the change before comparing new results with old ones. A clearer question can still break a trend line. Don’t silently treat differently worded questions as the same measure.

When the responses arrive, find the useful themes in survey comments keeps counts connected to the comments behind them.

Sources

• Government Analysis Function: Questionnaire design guidance
• Microsoft Support: Welcome to Copilot in Forms
• Google Docs Editors Help: Create a form with Gemini in Google Forms