By the Wavefield Research team · Published Sep 11, 2026
A bad survey question is one that makes some honest answers easier to give than others — through leading wording, smuggled assumptions, merged constructs, or options that leave real respondents homeless. Below are twelve broken questions, why each fails, and the repaired version, because poor survey questions are easy to catalogue and the guides rarely show the fix.
The uncomfortable property they share: biased questions produce clean-looking data. Nothing errors, completion holds, the percentages chart beautifully — and the numbers are wrong in a direction you cannot see from inside the results.
Each card names the failure, shows the broken question, explains the damage, and gives the rewrite. The categories cover what actually appears in circulating questionnaires, not a textbook taxonomy.
Bad How much did you enjoy our award-winning onboarding experience?
The question announces the verdict — award-winning, enjoy — and invites agreement. Respondents follow the path laid for them.
Fixed How would you rate your onboarding experience? (Very poor · Poor · Fair · Good · Excellent)
Bad What do you like most about shopping with us?
Assumes the respondent likes something. People who do not are forced to invent an answer or quit, and either way the data flatters you.
Fixed Which best describes your experience shopping with us? (Mostly positive · Mixed · Mostly negative) — then ask what drove it.
Bad Was the staff friendly and knowledgeable?
Two questions in one costume. A yes could mean either, a no is unreadable, and the fix costs one extra question.
Fixed Was the staff friendly? / Was the staff able to answer your questions? (asked separately)
Bad How satisfied are you with our omnichannel fulfilment experience?
Respondents answer the question they understand, which may not be yours. Internal vocabulary measures exposure to your marketing, not the experience.
Fixed How satisfied are you with how your order was delivered or picked up?
Bad How many times did you visit last month? (1–3 · 3–5 · 5–10)
A respondent with 3 or 5 visits has two homes, and different people resolve the tie differently — unmeasurable noise, invisible in the results.
Fixed How many times did you visit last month? (0 · 1–2 · 3–5 · 6–10 · More than 10)
Bad Rate our service: Excellent · Very good · Good · Fair
Three positive options, zero negative ones. The average is guaranteed flattering, and any reviewer who sees the scale will discount every number it produced.
Fixed Rate our service: Very poor · Poor · Fair · Good · Excellent (two negatives, a midpoint, two positives)
Bad Do you always read nutrition labels before buying?
Always and never turn a tendency question into a purity test. Nearly everyone honestly answers no, and the real variation hides inside it.
Fixed How often do you read nutrition labels before buying? (Never · Rarely · Sometimes · Often · Always)
Bad Do you exercise regularly?
Regularly means three gym sessions a week to one person and a monthly walk to another. Each respondent answers a private version of the question.
Fixed In a typical week, how many days do you exercise for 20 minutes or more? (0 · 1–2 · 3–4 · 5+)
Bad How many customer service interactions have you had in the past two years?
Nobody counts that far back; they estimate, and estimates round to tidy prototypes. Long windows plus frequent events produce confident fiction.
Fixed How many times have you contacted customer service in the past 3 months? (0 · 1 · 2–3 · 4 or more)
Bad Do you support the downtown redevelopment plan? (Yes / No)
On an issue with a real undecided middle, a forced binary manufactures opinion — and the manufactured half is unstable, so the number will not replicate.
Fixed Do you support or oppose the downtown redevelopment plan, or are you not sure yet? (Support · Oppose · Not sure) — probe leaners next.
Bad How satisfied were you with the mobile app? (scale, required)
Respondents who never used the app must fabricate a rating or abandon. Fabricated middles pile up at Neither and quietly flatten your real signal.
Fixed Same question, plus: Have not used the mobile app — or route non-users around it with a usage question first.
Bad Rate 15 statements about our brand on the same 7-point scale.
Somewhere around row seven, attention dies and one column gets ridden to the bottom. The last rows measure fatigue, not attitudes.
Fixed Two grids of 6 with rotated rows — and straightline detection flagging the respondents who gave up anyway.
The twelve failures look diverse, but three tests catch every one of them — which is what makes questionnaire review a discipline rather than a memorized list:
Read the question aloud, once, at conversation speed. If a listener would need it repeated, or could take it two ways, it fails — respondents on phones give a question one pass, and every misreading becomes noise wearing the costume of data.
Walk your options as the least typical respondent you will survey: the person who never used the feature, holds the unpopular view, or does not fit the categories. If any honest respondent has no truthful answer available, the question manufactures its data.
Reword the question to lean the opposite way — award-winning becomes controversial, like most becomes dislike most. If you would expect materially different results, the original was steering. Neutral questions survive their own mirror image.
The academic literature runs far deeper — Choi and Pak's catalogue of questionnaire biases documents 48 distinct species — but the dozen above account for most of the damage in commercial practice. The craft context lives in the neighboring guides: how each format works in survey question types, the option-list rules in closed-ended questions, and balanced scale construction in the Likert guide.
A broken skip logic errors out. A dead survey link gets reported. A biased question does neither: it collects answers on schedule, fills its quota cells, and delivers percentages with respectable base sizes. Every downstream safeguard — weighting, significance testing, quality flags — operates on the answers as given. No statistical machinery can recover what a leading question steered or a double-barreled item blurred, which is why question review happens before fielding or it does not happen at all.
This is also the honest limit of automation, ours included. The agent programs balanced scales, rotated options, and escape choices by default, which removes the mechanical failures — the overlapping ranges, the missing N/A, the three-positive scale. What no software fully removes is the motivated question: the stakeholder who wants the survey to prove the thing they already decided. For that, the three checks above and a reviewer with standing to say no remain the working defense. The review step in our workflow exists precisely so a human reads every question before anything fields.
One property, expressed a dozen ways: the question makes some honest answers easier to give than others. Leading wording pre-announces the verdict, loaded questions smuggle in assumptions, double-barreled items merge two answers into one, overlapping ranges give one respondent two homes, unbalanced scales tilt the average. The twelve patterns above cover what shows up in real questionnaires, and every one has a mechanical fix.
“How much did you enjoy our award-winning onboarding?” leads three ways: enjoy presumes enjoyment, award-winning imports social proof, and the how-much frame makes zero an awkward answer. The neutral version — “How would you rate your onboarding experience?” on a balanced scale — lets every honest answer cost the same effort. The test: reword it to lean the opposite way, and if results would change, it was leading.
A question that asks two things and accepts one answer: “Was the staff friendly and knowledgeable?” A yes cannot be attributed to either quality and a no cannot be diagnosed. Split it — friendliness and competence are separate questions — and the data becomes actionable. The word and inside a question stem is the cheapest warning sign in questionnaire review.
Good, unbiased survey questions share four habits: neutral framing that names both sides (satisfied or dissatisfied, support or oppose), one construct per question, balanced scales with equal positive and negative options, and an honest escape (Not sure, Not applicable, Prefer not to say) wherever a respondent might genuinely need one. Then run the three checks: read it once aloud, walk it as your least typical respondent, and mirror its direction.
Related: survey question types · closed-ended questions · survey screening questions · 50 open-ended survey questions
The agent programs balanced scales, rotated options, and honest escapes by default — and you review every question before it fields. From $99 per study.
Start a study