20 AI Prompts for Designing Surveys

TL;DR: These 20 prompts draft survey questions, then catch five ways a professional-looking question quietly produces biased data: leading language, double-barrelled phrasing, unbalanced scales, a missing not-applicable option, and response lists that don't cover the space. None of them fix a bad sample, and none of them generate fake responses to report as real.

What Makes a Survey Question Bad, Even When It Reads Like a Professional One?

A leading, double-barrelled, or unbalanced question can look exactly as polished as a well-built one, run cleanly in your survey tool, and still produce data that's confidently wrong. A model asked to write survey questions will generate this kind of defect by default unless you specifically ask it to check for one. The five patterns below are worth checking every time.

Pew Research Center's own methodology guide names the double-barrelled pattern directly, using a real example from one of its own questions: “How much confidence do you have in President Obama to handle domestic and foreign policy?” Questions built this way "are difficult for respondents to answer and often lead to responses that are difficult to interpret", because there's no way to know which half of the question a given answer is actually about (Pew Research Center, Writing Survey Questions, accessed September 3, 2026).

Sampling Is the Limit No Prompt Can Fix

Pew's own framing of this is blunt: "Accurate random sampling will be wasted if the information gathered is built on a shaky foundation of ambiguous or biased questions" (Pew Research Center, Writing Survey Questions, accessed September 3, 2026). The reverse is just as true and gets said far less often: the most carefully worded questionnaire in the world is wasted if it never reaches a sample that represents who you're trying to learn about. This page can help with the questions. It cannot tell you whether your distribution list is the right one, and nothing below should be read as implying otherwise.

How Do You Draft a Question Bank From a Research Objective?

Most survey-writing time gets spent staring at a blank page trying to turn a vague goal into specific questions. Starting from the objective, not the topic, is what actually produces a usable draft.

1. Turn a vague research objective into a candidate question bank

Role: someone drafting survey questions from a research objective.
Task: given the objective below, draft 8-10 candidate questions that would help answer it,
mixing closed-ended and open-ended formats.
Research objective: [WHAT YOU'RE TRYING TO LEARN]
Audience: [WHO WILL ANSWER THIS]
Format: numbered list, each question tagged with its type (closed-ended, open-ended, or
screener).
Constraints: every question must map back to the stated objective. Flag any question that
feels interesting but doesn't actually serve it, rather than including it anyway.

What to change: state the objective as a decision you'll make with the data, not a general topic. An objective like understanding customer satisfaction produces vague questions; one stated as deciding whether to keep or drop the current onboarding flow produces specific, decision-relevant ones.

2. Draft a screener section that avoids over-collection

Role: someone drafting the screener questions at the start of a survey.
Task: draft the minimum set of screener questions needed to confirm a respondent qualifies
and to allow the analysis below, and nothing beyond that.
Who qualifies: [CRITERIA]
Analysis you plan to run by subgroup, if any: [E.G., "COMPARE BY AGE BAND"]
Format: numbered screener questions, each with a one-line note on why it's needed.
Constraints: exclude any demographic question that isn't tied to a stated analysis need.

What to change: cut anything you added out of habit rather than need. Every screener question is a small tax on completion rate and a small privacy cost, and might be interesting later isn't a justification for collecting it now. If you run the same screener criteria across studies, storing it once as a reusable field is worth setting up; see our guide to global variables worth reusing.

3. Pair an open-ended follow-up to a closed-ended question

Role: someone deciding whether a closed-ended question needs an open-ended follow-up.
Task: given the closed-ended question below, draft a short open-ended follow-up that would
surface the "why" behind the answer, only if the closed-ended answer alone would leave that
why genuinely unclear.
Closed-ended question: [QUESTION]
Format: the follow-up question, or a one-line explanation of why none is needed.

What to change: don't default to adding one. An open-ended follow-up after every closed question inflates completion time and survey fatigue for a why that's often already obvious from the closed answer.

4. Draft an intro blurb that sets context without leading the respondent

Role: someone writing the introductory text respondents see before starting the survey.
Task: draft 2-3 sentences that state the survey's purpose and expected length, without
signalling what answer is desired.
Survey purpose: [PURPOSE]
Expected length: [MINUTES]
Format: 2-3 sentences.
Constraints: state the purpose in neutral terms. Do not include language that implies the
organization hopes for a particular outcome.

What to change: read it back and ask whether it primes a specific answer. An intro that mentions helping us improve our award-winning service invites a more generous response than one that simply states the survey's topic.

How Do You Critique an Existing Survey for Bias?

Running an existing draft back through a bias check catches exactly the five patterns from the table above, and it's most useful on a survey you didn't write yourself, where you have no attachment to the original wording.

5. Leading-question detector

Role: someone auditing a survey draft for leading language.
Task: read the questions below and flag any that contain their own implied answer or use
loaded language, rather than stating the topic neutrally.
Survey draft: [PASTE QUESTIONS]
Format: for each flagged question, quote the leading phrase and suggest a neutral rewrite.
Constraints: only flag genuine leading language. Do not flag a question just for having a
specific or detailed topic.

What to change: run this on someone else's draft before your own if you can. It's much easier to spot a leading question you didn't write than one you did, since you already know what answer you were hoping for.

6. Double-barrelled question detector

Role: someone auditing a survey draft for double-barrelled questions.
Task: read the questions below and flag any that ask respondents to evaluate more than one
concept at once.
Survey draft: [PASTE QUESTIONS]
Format: for each flagged question, name the two concepts being conflated and propose splitting
it into two separate questions.

What to change: watch for the word and joining two different concepts inside one question. It's the single most reliable tell, and it's exactly the pattern Pew's own methodology guide calls out as producing responses that are hard to interpret.

7. Unbalanced-scale detector

Role: someone auditing response scales for balance.
Task: read the scale below and state whether it has an equal number of positive and negative
options around a genuine midpoint.
Response scale: [PASTE THE SCALE]
Format: verdict (balanced or unbalanced), and if unbalanced, a corrected scale with matching
options on each side.

What to change: count the options yourself before trusting the verdict. A scale that reads excellent, good, fair, poor only looks balanced; it has two positive labels and two that read as negative once fair is read as lukewarm, not neutral.

8. Missing-not-applicable detector

Role: someone auditing a survey for questions that force an opinion from people who have none.
Task: read the questions below and flag any that lack a not-applicable, no-opinion, or haven't
used it option, where a meaningful share of respondents plausibly has no basis to answer.
Survey draft: [PASTE QUESTIONS]
Format: for each flagged question, propose the specific option to add.

What to change: think about who's being screened out, not just who's answering. A question with no not-applicable option doesn't just annoy respondents without an opinion; it quietly forces them into a fabricated one, sitting in your data next to genuine opinions with no way to tell them apart.

9. Response-options completeness checker

Role: someone auditing whether a question's response options actually cover realistic answers.
Task: read the question and options below and identify any realistic answer that doesn't fit
any listed option.
Question and options: [PASTE]
Format: list of gaps found, and where an "other, please specify" option would close them versus
where a specific missing option should be added instead.

What to change: think of a genuinely awkward real answer before running this, not just an obviously missing one. The gaps that matter are the ones a real respondent would actually have, not the theoretical edge cases a checklist assumes.

How Do You Write Neutral Response Scales?

The format a question uses to collect an answer shapes the answer as much as the wording does. A few structural choices matter more than most survey writers expect.

10. Convert an agree/disagree statement into a balanced forced-choice pair

Role: someone converting an agree-disagree question into a forced-choice format.
Task: given the statement below, write two contrasting statements representing genuine
alternative positions, and ask the respondent to choose between them instead of agreeing or
disagreeing with one.
Original agree-disagree statement: [STATEMENT]
Format: two contrasting statements and a single selection question.

What to change: make sure both statements are genuinely defensible positions, not one real option and one straw man. A forced choice between a real view and an obviously weaker one isn't neutral; it's a leading question with extra steps.

11. Build a Likert-type scale with a genuine midpoint

Role: someone building a Likert-type agreement or satisfaction scale.
Task: produce a 5-point scale with a genuine, meaningfully neutral midpoint and a separate
not-applicable option outside the scale itself.
Topic: [WHAT'S BEING RATED]
Format: five labeled points plus a separate not-applicable option, labels stated in full
(not just "3" for the midpoint).
Constraints: the midpoint must read as genuinely neutral, not as a mild version of either
extreme.

What to change: read the midpoint label out loud. If it reads as somewhat good rather than truly neutral, the scale is silently unbalanced even though it has five points.

12. Build frequency response options that avoid vague middle terms

Role: someone building response options for a frequency question.
Task: produce options with specific, non-overlapping frequency ranges instead of vague terms.
Behavior being measured: [BEHAVIOR]
Typical range of frequency, if known: [RANGE]
Format: 4-5 options with specific ranges (e.g., a defined number of times per week or month),
not vague terms like "sometimes" or "often."

What to change: anchor the ranges in something the respondent can actually count. Sometimes means something different to every respondent; a specific weekly range means roughly the same thing to all of them.

13. Ranked-choice versus rating-scale decision helper

Role: someone deciding whether a question should use ranked choice or independent rating
scales.
Task: given the list of items below, recommend ranked choice or independent ratings, and
explain the tradeoff for this specific case.
Items to be evaluated: [LIST]
What you need from the data: [E.G., "WHICH ONE OPTION WINS" VS. "HOW EACH ITEM RATES ON ITS
OWN"]
Format: recommendation plus a short explanation of the tradeoff.

What to change: be honest about what you actually need. Ranking forces tradeoffs between items that might all rate highly on their own; independent ratings lose the forced tradeoff but let every item succeed or fail on its own terms.

How Do You Turn Open-Text Answers Into Themes Without Inventing Them?

Open-text responses are where the richest and messiest data lives, and it's also where an assistant is most likely to overstate what it found.

14. Cluster open-text responses into themes

Role: someone analyzing open-text survey responses.
Task: read the responses below and group them into themes, each theme labeled and given a
short description.
Raw open-text responses: [PASTE RESPONSES]
Format: numbered themes, each with a label, a one-line description, and a count of how many
responses fit it.
Constraints: only create a theme that genuinely recurs across multiple responses. Do not
create a theme for a single unusual answer.

What to change: read the actual theme list skeptically before trusting it. Models tend to over-cluster, finding more distinct themes than the data really supports; merge two themes yourself if they're clearly saying the same thing in different words.

15. Pull representative verbatim quotes per theme

Role: someone selecting representative quotes to illustrate themes found in open-text
responses.
Task: for each theme below, select one or two verbatim quotes from the raw responses that
best represent it.
Themes: [LIST]
Raw open-text responses: [PASTE RESPONSES]
Format: theme, followed by the exact verbatim quote(s), unedited.
Constraints: quotes must be copied exactly from the raw responses given. Never paraphrase a
quote and present it as verbatim, and never invent a quote that sounds representative.

What to change: check every selected quote against the original response yourself before publishing it. A verbatim quote that's been quietly smoothed for readability is no longer verbatim, and readers of your report have no way to tell.

16. Summarize open-text sentiment without inventing a percentage

Role: someone summarizing the overall tone of open-text responses.
Task: describe the general sentiment pattern across the responses below in qualitative terms.
Raw open-text responses: [PASTE RESPONSES]
Format: 2-3 sentences describing the pattern (e.g., predominantly positive with a recurring
complaint about X), with no invented percentage or count unless you provide the actual tally.
Constraints: do not state a percentage or proportion unless it was computed from an actual
count you supply. A qualitative description is preferable to a fabricated number.

What to change: if you want an actual percentage, count it yourself first and feed the count into the prompt. An assistant asked for a sentiment breakdown without real counts will produce a plausible-sounding number that isn't grounded in anything.

How Do You Pilot-Test a Survey Before Fielding It?

A pilot catches confusing wording and broken skip logic before either one costs you real respondents. None of this substitutes for a genuine pilot with real people reading the actual questions.

17. Cognitive-interview probe script for pilot testing

Role: someone preparing a cognitive interview to pilot-test a survey question.
Task: write 2-3 probe questions to ask a pilot tester immediately after they answer the
question below, to check whether they understood it as intended.
Question being tested: [QUESTION]
What you intend it to measure: [INTENT]
Format: 2-3 probe questions, e.g. asking the tester to restate the question in their own words
or explain what they thought it was asking.

What to change: run this with a real person, not with the AI answering its own probe questions. The entire value of a cognitive interview is discovering a gap between your intent and a real reader's interpretation, which the model can't discover about itself.

18. Synthetic pilot responses for testing instrument logic, not for reporting

Role: someone testing a survey's skip logic and question flow before fielding it.
Task: generate 5-10 synthetic respondent answer sets that exercise different branches of the
skip logic below, clearly labeled as test data.
Survey structure and skip logic: [DESCRIBE OR PASTE]
Format: numbered synthetic response sets, each labeled "SYNTHETIC TEST DATA - NOT A REAL
RESPONSE" at the top.
Constraints: this output is for verifying that skip logic and branching work correctly. It
must never be reported, analyzed, or presented as real respondent data under any circumstance.

What to change: delete or clearly quarantine this output once you've confirmed the logic works. Synthetic test data sitting in the same folder as real response exports is exactly how it ends up in a chart by accident.

19. Time-to-complete estimate check

Role: someone estimating how long a survey draft will take to complete.
Task: given the question count and types below, estimate a rough completion time range and
flag any section that looks disproportionately long relative to the rest.
Number and types of questions: [LIST OR PASTE THE DRAFT]
Format: a rough time range and a flag on any section that seems long relative to the others.
Constraints: state this as a rough estimate, not a precise figure, since actual completion
time depends on your specific audience and platform.

What to change: validate the estimate against an actual pilot timing, not just the model's guess. This prompt is useful for catching an obviously bloated section before you field anything; it isn't a substitute for timing real testers.

20. Pre-field checklist run against a finished draft

Role: someone doing a final check on a survey draft before fielding it.
Task: run the finished draft below against a checklist: any leading questions, any
double-barrelled questions, any unbalanced scales, any missing not-applicable options, and any
non-exhaustive response lists.
Finished survey draft: [PASTE FULL DRAFT]
Format: a pass/fail note per checklist item, with the specific question flagged for anything
that fails.
Constraints: flag every instance found, even if it means the draft isn't ready to field yet.

What to change: run this as a genuinely final step, after content changes are done. Running it mid-draft and fixing issues one at a time as you go is fine too, but always run it once more against the truly final version, since a late wording change can reintroduce a problem an earlier pass already cleared. For a broader pre-send check beyond bias alone, our prompt hygiene checklist covers the general habits worth running on any important prompt, not just a survey draft.

Does the Free Plan Cover This?

Yes, for occasional use. The Free plan includes 5 prompt enhancements a day, forever, no card required, per the FAQ page, which is enough to draft or bias-check a handful of questions in one sitting. What Free doesn't include: saving these as reusable templates, or a personal context library for storing your standard screener questions and research objective once instead of retyping them into every prompt. Those are Pro-plan features, priced at the time of writing on /pricing; check the current figure there, since it changes.

If your research question itself is still vague, start one step earlier with our research question generator before drafting the survey questions above. And if you need a live discussion rather than written responses, see our companion guide on running better meetings with AI.