Key takeaways
- One idea per question. Asking about "speed and accuracy" in one line makes every answer ambiguous. Split it. In our own template library this is the pattern that shows up most often.
- Wording moves the number, not just the mood. In split-sample experiments, adding one clause about casualties moved support for military action from 68 percent to 43 percent. Same topic, same week, different sentence.
- Replace vague quantifiers with dates. "Recently" and "often" mean different things to different people. "In the last 30 days" means one thing.
- Ask only what you will act on. If you cannot name the decision an answer changes, the question is costing you completion for nothing.
- Nothing here is gated. The question bank has a copy button and the checklist has a print button. No form, no email address.
Start with the question you already wrote
Most people arrive here with a draft in hand, not a blank page. So start by finding the flaw rather than reading a philosophy of measurement. Each card below names one failure pattern and the tell that gives it away. Pick the one your question matches.
Read your draft question once and look for these six tells. If more than one applies, fix them in the order listed: a double-barreled question cannot be repaired by better wording until it is split.
- Double-barreledThe word "and" or "or" joins two things you could feel differently about.
- LeadingAn adjective, or a claim about what other people think, sits inside the question.
- LoadedAnswering at all means accepting something you never checked.
- Vague"Often", "recently", "regularly". No date, no count, no anchor.
- Broken optionsRanges overlap, a common answer is missing, or the scale leans one way.
- Double negativeYou have to read it twice to work out which way is agreement.
What makes a survey question work
A good survey question is not a clever one. It is one where every respondent reads the same words and understands the same thing, and where the answer changes what you do next. That is the whole job. Four properties get you there.
- It names one specific thing. "How satisfied are you with our service?" invites six different definitions of service. "How satisfied are you with how quickly your last support request was answered?" invites one.
- It carries no opinion of its own. Any adjective inside the question ("easy", "helpful", "award-winning") tells the respondent what you hope to hear. Let the answer options carry the range instead.
- It survives one read. If someone has to re-parse the sentence, you are measuring reading effort as much as opinion. Aim for one clause and plain words.
- It is attached to a decision. Write down what you would do if the answer came back low, and what you would do if it came back high. If those are the same thing, cut the question.
The format you choose matters less than those four properties, but it is not free. Closed formats are faster to answer and faster to analyse; open formats tell you things you did not think to ask. The table below is a routing guide, and each format has its own guide on this site with worked examples.
| Format | Use it when you need | Example stem | Main cost |
|---|---|---|---|
| Single-select list | To sort people into groups | "Which best describes your role?" | You must know the options in advance |
| Select all that apply | Several answers can be true at once | "Which of these have you used in the last 30 days?" | Order effects; later options get picked less |
| Agreement statements | Attitudes you will track over time | "The tools I use day to day do what I need." | Some people agree with almost anything |
| Numeric scale | A score you will benchmark or trend | "How likely are you to recommend us?" | Scores drift with label wording |
| Short open text | The reason behind a score | "What is the main reason for your score?" | Slower to answer and to code |
| Grid of statements | Many items on one shared scale | "Rate each of these from 1 to 5." | Hard on phones; invites straight-lining |
Going deeper on any one format: see single and multi-select question types, agreement scales and how many points to use, numeric scales and their labels, and open versus closed questions. For age, gender, income and similar items, demographic questions have their own rules about placement and optionality.
Six ways survey questions break, and the fixed version of each
These six patterns cover almost every bad survey question we see. Each one below has the broken version, the rewrite, and a note on why the change matters. The rewrites are not stylistic preferences. In each case the original produces an answer you cannot interpret, and the fix produces one you can.
Double-barreled questions ask two things and accept one answer
A double-barreled question bundles two topics into one sentence. Someone who is delighted with one and annoyed by the other has no honest way to respond, so whatever they pick is noise.
How satisfied are you with the speed and accuracy of your delivery?
How satisfied are you with how quickly your order arrived?
Then ask the second half as its own question. An order that arrives fast and wrong should score high on one and low on the other, and you cannot see that if they share a row.
Was our support team friendly and did they solve your problem?
Was your problem solved?
Outcome and manner are separate results. Teams that are pleasant but ineffective look identical to teams that are effective but brusque when you merge them.
How useful and easy to understand was the training?
How much of the training could you follow without help?
Clear and useless is a real result. So is confusing and valuable. Merging them hides the exact thing a training team needs to know.
Leading questions tell people the answer you want
A leading question steers through loaded adjectives, or through a claim about what other people already think. The respondent has to push against the sentence to disagree with it, and most will not bother.
How satisfied are you with our award-winning support team?
How would you rate the support you received?
"Award-winning" flatters the team before the respondent has formed a view. It also quietly frames the scale: disagreeing now means contradicting an award.
Most of our users found setup straightforward. How easy did you find it?
How easy or difficult was setting up your account?
Two fixes in one. The social proof is gone, and the stem now names both directions, so "easy" is no longer the default reading.
How much did you enjoy our redesigned checkout?
How would you describe checking out compared with last time?
"How much did you enjoy" has no floor below "a little". A person who hated it cannot answer accurately, so the average comes back positive by construction.
Loaded questions make you accept something to answer at all
A loaded question smuggles in an assumption. Every available answer confirms it, which means the data is wrong before anyone reads it. This is the pattern people most often miss in their own drafts, because the assumption is usually true of the author.
Where do you like to go when you eat out?
In the last 30 days, how many times did you eat at a restaurant?
Ask this first, then show the "where" question only to people who answered anything above "none". People who never eat out otherwise invent an answer.
What problems did you run into during onboarding?
How did onboarding go for you?
The original presumes problems. People who had none will either skip it or invent a small complaint to be helpful, and both outcomes overstate how bad onboarding is.
How has our new pricing helped your team save money?
Since the pricing change, has what your team spends with us gone up, down, or stayed the same?
The original has no room for "it cost us more". Adding "Not sure" matters too: without it, people who genuinely do not know are forced to guess, and guesses look exactly like data.
Vague questions get measured against different yardsticks
Words like "often", "regularly" and "recently" feel precise when you write them and mean nothing shared when they are read. Two people with identical behaviour will give different answers, so the variation you measure is language, not life.
Do you exercise regularly?
In a typical week, on how many days do you exercise for at least 20 minutes?
"Regularly" is doing three jobs at once: how often, how long, and how consistently. The rewrite fixes all three and produces a number you can trend.
Have you contacted support recently?
Have you contacted support in the last 90 days?
Pick a window your respondent can actually recall. Beyond about three months, people compress and reorder events, so a longer window buys reach at the cost of accuracy.
How would you rate the value of our product?
For what you pay, how would you rate what you get?
"Value" can mean price, usefulness, or quality depending on who reads it. Naming the trade explicitly gets everyone answering the same question.
Broken answer options ruin questions that were fine
A well-written stem still fails if the choices overlap, leave gaps, or lean one direction. This is the easiest pattern to catch and the one most often shipped anyway, because people proofread the question and skim the list underneath it.
What is your annual household income? Under $25,000 / $25,000 to $50,000 / $50,000 to $75,000 / $75,000 to $100,000
What is your annual household income before tax?
Three faults fixed. The brackets no longer overlap at $25,000, $50,000 and $75,000; there is now a top bracket; and people can decline instead of abandoning the survey.
How would you rate our service? Good / Great / Excellent / Outstanding
How would you rate our service?
An unbalanced scale guarantees a positive average no matter what people think. A balanced one has the same number of steps either side of a neutral midpoint.
Which of our features do you use most? Dashboards / Reports / Alerts
Which of these have you used in the last 30 days? Select all that apply.
The original forces a pick from people who use none of the three, and hides people who use all three. "None of these" is not padding; it is the difference between a real zero and a fake preference.
Double negatives measure reading speed
Two layers of negation force the respondent to run a small logic puzzle before answering. Most people will not run it. They guess at the direction and move on, which turns the item into a coin flip dressed as an opinion.
Do you disagree that the refund policy is not flexible enough?
How flexible is the refund policy?
Ask the thing itself. Direction belongs in the answer options, where the respondent can see all of it at once rather than holding it in their head.
Should we not remove the option to skip onboarding?
Should people be able to skip onboarding?
"Should we not remove" contains two negations and a policy proposal. Nobody answers that accurately, including the person who wrote it.
I never find the reports unhelpful. Agree or disagree?
How helpful are the reports?
"Never" plus "unhelpful" is a double negative wearing a disguise. Rewriting to a direct item also removes the agree-disagree format, which some people say yes to regardless of content.
How to write unbiased survey questions
Fixing the six patterns removes most of the damage. These four habits catch the rest, and they take about a minute per question.
- Put a date on it. "In the last 30 days" beats "recently" every time, and it is the single highest-yield edit in this list.
- Name both ends in the stem. "How easy or difficult" rather than "how easy". Asking only about the direction you hope for pulls answers that way.
- Give people a truthful exit. "Not applicable", "I do not know", and "Prefer not to say" convert forced guesses into honest blanks. A blank you can see beats a guess you cannot.
- Read it aloud to someone outside the project. Anything they ask you to explain is a question that will be interpreted differently by strangers. Rewrite it, do not clarify it.
Bias also enters through who answers and how they behave once they start, not only through wording. That side of the problem is covered in the guide to response bias, and who ends up in your sample at all is covered in sampling methods.
What 799 real survey questions actually look like
Advice about survey questions is usually written from memory. We had a way to check ours against something real, so we did. Every live template in the SuperSurvey library was pulled from the database and every question in it was classified by machine, then the flagged ones were read by hand. That is 87 surveys and 799 questions. The method is written out at the foot of this page so you can argue with it.
The headline finding is not the one we expected. Of the six failure patterns above, five are close to absent from our own library. One is everywhere.
Rates are of the 636 closed-format questions (scale and single-select). Open text is excluded because a two-part prompt in a text box is not a defect: the respondent can answer both halves. The double-barrelled figure is a manually audited estimate; the raw pattern match caught 170 questions and a hand review of a systematic sample found roughly a quarter of those were fixed phrases such as "role and responsibilities" rather than genuine faults.
See the counts behind the chart
| Pattern | Flagged | Base | Share | Checked by hand? |
|---|---|---|---|---|
| Two topics in one question | about 124 | 636 closed-format | 20% | Yes, sample of 21 of the 145 scale hits |
| Undefined frequency word in an agreement item | 15 | 636 closed-format | 2.4% | Yes, all 15 |
| "How often" with no timeframe in stem or options | 7 | 636 closed-format | 1.1% | Yes, all 7 |
| Absolute wording ("all", "always") | 2 | 636 closed-format | 0.3% | Yes, both |
| Leading adjective in the stem | 1 | 636 closed-format | 0.2% | Yes |
| Double negative | 0 | 799 all questions | 0% | Yes, swept all 799 |
| Stem longer than 25 words | 0 | 799 all questions | 0% | Longest stem in the library is 22 words |
Read that chart again, because it inverts the usual advice. The leading question, which almost every guide on this topic opens with, appears once in 636 closed-format questions. The double-barrelled question, which most guides give a paragraph, is in roughly one in five. If you have limited review time, spend it hunting for the word "and".
- 9Median questions per template. Every one of the 87 sits between 7 and 14.
- 9.3Mean words per question. Not one of the 799 runs past 25 words.
- 37%Of all 799 questions are the same agreement scale, usually with the same two anchors.
- 0Questions use a 0 to 10 scale, despite 23 asking whether you would recommend something.
Two of those deserve a word of caution about what they do and do not mean. The 9-question median is a house rule, not a discovery about surveys in general: we sampled one authoring pipeline, not the world. And the 37 percent figure is a criticism of ourselves. When more than a third of a library is the same widget with the same two anchors, the format is being chosen by habit rather than by what the question needs.
The zero is the most useful number here. Twenty-three questions in the library ask some version of "would you recommend this", and not one of them uses the 0 to 10 scale that the recommendation metric is defined on. If you copy one of those questions into a report and call the result a recommendation score, the number will not mean what your audience thinks it means. Match the scale to the metric you intend to quote.
We checked our own library against our own advice
This page says open with something topical and leave the "who are you" questions until the end. Our library mostly does that, and sometimes does not.
Across the 87 templates, 70 percent open with a genuine topical question, and 84 percent of the demographic questions sit in the final quarter of their survey. But 13 templates still open by asking who you are, and five open with an administrative field. One asks for your full name before it asks you anything about the topic. Those are on our list to fix, and naming them here is cheaper than pretending the number is 100 percent.
The average position tells the same story more precisely. On a scale where 0 is the first question and 1 is the last, rating scales average 0.33, single-select questions 0.61, open text 0.68, and demographics 0.78. That is the right shape: easy and topical first, effortful later, personal last.
How much does wording actually move the answer?
Almost every guide on this topic tells you to avoid leading questions. Almost none tells you what it costs if you do not. That gap matters, because "this wording is slightly better" and "this wording moves the result by fourteen points" call for very different amounts of your attention.
The table below collects the wording experiments we could verify, with the measured effect and a confidence grade. Where the evidence is contested, it says so.
| Change | Measured effect | Grade | Source |
|---|---|---|---|
| "Forbid" instead of "allow" | Answers shift by 14 points on average across every published experiment since 1940. The spread is wide: for most questions the gap falls somewhere between minus 6 and plus 34 points. | A | Holleman, 19991 |
| Adding a clause about consequences | Support for military action fell from 68 percent to 43 percent when the question added "even if it meant U.S. forces might suffer thousands of casualties". | B | Pew Research Center2 |
| Adding one adjective | "Are there plenty of jobs available" got 60 percent yes. "Plenty of good jobs" got 48 percent. Same week, same sample design. | B | Pew Research Center, 20193 |
| Agreement scales instead of direct questions | People agree more than they disagree with the same idea stated either way, a built-in yes bias of about 10 points, summarised across 25 studies. | A | Krosnick and Presser, 20104 |
| Rewriting an agreement item as a direct question | Measurement quality rose to about 0.75 for direct questions against 0.18 to 0.40 for the matched agreement version, across 14 countries. | A | Saris and others, 20105 |
| Splitting a two-topic question in two | Of 8 two-topic items tested, 6 gave significantly different results once each half was asked on its own. | B | Menold, 20206 |
| Listing an answer instead of leaving it open | 58 percent picked "the economy" when it was offered in a list. Only 35 percent volunteered it when the same question was open. Open answers also fell outside the offered list 43 percent of the time. | B | Pew Research Center2 |
| "Select all that apply" instead of yes or no per item | Select-all undercounts by about 8 points on average, and up to 16 points on a single item, across 4,581 respondents. | B | Pew Research Center, 20197 |
| Adding scale points | Reliability climbs steeply from 2 points to about 7, then flattens. People use only about half the points you offer, topping out near 9. | A | Krosnick and Presser, 20104 |
| Offering "no opinion" | The share taking it rose from 15 percent to 38 percent, but the balance of opinion among everyone else did not change at all. | A | Presser and Schuman, 19758 |
| Where an option sits in the list | Eye tracking found only 54 percent of people ever look at the last option on a 5-point scale, and 60 percent picked from the first half of a 12-item list. | B | Galesic and others, 20089 |
| Hiding a definition behind a hover | 10 percent of people read a definition that needed a mouse-over. 78 percent read the same definition when it was always visible. | B | Galesic and others, 20089 |
| Adding 56 questions to a questionnaire | Response rate fell 2.7 points, and the whole drop was on paper. Online respondents showed no difference, and answer quality did not fall either. | B | Sandelin, 202210 |
| Numbering a scale 0 to 10 rather than minus 5 to plus 5 | A 1991 study found a 21-point shift. A 2025 cross-national replication found the effect running the other way. Treat this one as unsettled. | C | Schwarz and others, 1991; Bottoni and Aizpurua, 202511 |
- A meta-analysis, or replicated across studies
- B one strong study
- C well known but contested or dated
Three things fall out of that table that are worth acting on today.
First, the agreement scale is the default in almost every survey tool, including ours, and it is the weakest format on the list. Rewriting "The reports are useful" with an agree-to-disagree scale into "How useful are the reports?" with a not-at-all-to-extremely scale costs you nothing and buys a large amount of measurement quality.5 If you keep the agreement format, keep it at five points.12
Second, the last option on your scale is not being read by roughly half your respondents.9 That is an argument for short lists, for putting the option you most want measured accurately away from the very end, and for randomising order where the list has no natural sequence.
Third, and least comfortable: length matters less than most advice claims. Adding 56 questions cost under three points of response rate, and nothing at all among people answering online.10 Cut questions because they have no decision behind them, not because you read that a long survey will destroy your response rate.
Numbers we deliberately did not use
Four figures circulate constantly in writing about surveys. We chased each to its origin and left all four out. If you have seen them cited elsewhere, this is why.
- "The average attention span is 8 seconds, shorter than a goldfish."
- Usually credited to Microsoft, 2015. Microsoft did not measure it. Their report attributes the figure to a third-party statistics site and dates it to 2013, and their own study measured detection counts and reaction times, which cannot produce a number in seconds. Journalists who contacted the two organisations named as the underlying sources found neither had any record of the research.13 As of August 2026 the site that supplied the number no longer resolves at all. Fish memory, meanwhile, has been studied since 1908.
- "People can only hold 7 plus or minus 2 things in memory, so cap answer options at seven."
- Miller's 1956 paper is real and important, and it says the opposite of this. It describes three different limits and warns in its own text that treating them as one process is "a fundamental mistake". Miller called the number itself a coincidence. More to the point, answer options are visible on screen while the respondent chooses, so nothing has to be held in memory at all. How many scale points to use is a measurement question, and the measurement answer is in the table above.14
- "Survey response rates have fallen to 3 percent."
- There is no single response rate to fall. Rates differ by more than an order of magnitude across contact methods. What genuinely collapsed is telephone polling: Pew's own series runs from 36 percent in 1997 to 6 percent in 2018.15 Government face-to-face surveys declined far more gently over the same period. Quote a response rate only with the method attached.
- "Surveys over 7 to 8 minutes lose 5 to 20 percent of completions."
- This traces to two undated vendor blog posts by the same author. One carries the percentage range with no date range, no selection rule and no dataset. The other carries the methodology and no percentages. The two get merged in retelling. No peer-reviewed version of either analysis exists. The measured effect from a randomised experiment is 2.7 points for 56 extra questions, and zero online.10
The pattern behind all four is worth more than the corrections. Two of them trace to web pages that no longer exist, and both looked properly sourced right up until someone followed the link. The rest are vendor pages citing other vendor pages. Where peer-reviewed work does exist on these questions, it is consistently less dramatic than the folklore.
70 survey question examples you can copy
Every question below follows the four properties above: one topic, no built-in opinion, a stated timeframe where one is needed, and a decision behind it. Pick a category, then use the copy button to take the whole list as plain text. Nothing is gated and there is no form.
70 questions across 6 categories. The answer format under each one is a suggestion, not a rule. Copy takes the visible tab, one question per line.
- How likely are you to recommend us to a friend or colleague?
- What is the main reason for your score?
- How easy or difficult was it to get what you needed today?
- How satisfied are you with how quickly your request was handled?
- Was your problem solved?
- How many times did you have to get in touch about this one issue?
- Compared with the last time you bought from us, how was this time?
- Which part of the experience took the longest?
- In the last 30 days, how many times have you used us?
- For what you pay, how would you rate what you get?
- How well did what arrived match what you expected?
- If we changed one thing before your next order, what should it be?
- How likely are you to buy from us again?
- Is there anything we did that you would not want us to change?
- How clear is it to you what you are meant to deliver this quarter?
- In a typical week, how many hours go to work you would call low value?
- The tools I use day to day do what I need.
- When you raise a problem, how often does something change?
- How comfortable would you be disagreeing with your manager in a meeting?
- In the last 3 months, have you had a conversation about your development?
- How manageable is your current workload?
- What is one thing that would make next week easier?
- How well do you understand how your work connects to company goals?
- How likely are you to still be here in 12 months?
- Which of these got in your way in the last 30 days? Select all that apply.
- Is there anything you would want leadership to hear that this survey did not ask?
- Which of these have you used in the last 30 days? Select all that apply.
- How disappointed would you be if you could no longer use this?
- What were you trying to do when you opened the app today?
- How well does this feature do the job you needed?
- Where did you get stuck, if anywhere?
- How much of the setup could you finish without help?
- What did you use before this?
- How often do you use this feature?
- If this feature disappeared tomorrow, what would you do instead?
- How would you rate how fast the app feels?
- Which one of these should we work on first?
- What almost stopped you from signing up?
- How much of the session could you follow without help?
- How useful will this be in your actual work?
- How confident are you doing this on your own now?
- Was the session the right length?
- Which part was most useful to you?
- Which part would you cut?
- How well did the room and equipment work?
- What will you do differently next week because of this?
- How did you hear about this event?
- How likely are you to attend another session?
- What did you expect that you did not get?
- What did you come to this page to do?
- Were you able to do it?
- How easy or difficult was it to find what you needed?
- How clear was the information on this page?
- What is missing from this page?
- How many pages did you look at before this one?
- Did anything on this page confuse you?
- Which device are you using right now?
- How did you get to this page?
- If you did not find what you needed, where will you look next?
- How would you rate the checkout you just completed?
- What is the main reason for your score?
- What was the single most frustrating part?
- What nearly stopped you from finishing?
- What would have made this a 5 out of 5?
- What did you expect to happen that did not?
- If you could change one thing, what would it be?
- What should we have asked you that we did not?
- Who else was involved in this decision?
- What were you doing right before this went wrong?
- Is there anything else you want us to know?
If you want a whole questionnaire rather than single questions, there are ready-made sets for customer satisfaction, employee research, marketing, training, technology and business, or you can browse the whole library. They are the 87 templates measured further up this page, so you can see for yourself where they follow the advice and where they do not.
The order you ask in changes the answers you get
Question order is not just housekeeping. An earlier question can prime the frame someone uses for a later one, which is why the same item can score differently depending on what preceded it. Five rules cover most of it.
- Open with something easy and on-topic. The first question sets expectations for the rest. A quick single-select about what the person was doing works. A text box does not.
- Keep one topic together before moving on. Jumping between billing, product and support makes people re-orient every few items, and re-orienting is where they quit.
- Score first, explain second. Ask the rating, then ask why. Reversed, the act of writing an explanation changes the score the person then gives.
- Put general before specific, unless you are testing the effect. A general satisfaction item asked after five detailed complaints will absorb those complaints. Ask it while the frame is still wide.
- Leave anything sensitive to the end, and make it optional. Age, income, identity and pay belong last. If someone abandons at that point you still keep everything before it.
Length and format decide who finishes
Two things drive abandonment more than wording does: how long the survey is, and how much effort each item takes. Open text is the expensive format. It is worth paying for once or twice, and rarely worth paying for five times.
One closed question for the number, one optional open follow-up for the reason. You get something you can trend and something you can quote, and you have only spent one text box on it.
Length also interacts with how many responses you actually need. If you are trying to work out how many completed responses will support a claim, that is a separate calculation: see how to work out how many responses you need.
Run this before you send
Ten checks, about a minute per question. The first four catch most problems.
- One idea only. If "and" or "or" joins two things you would act on separately, split the question.
- No adjectives in the stem. Delete every one. If the question still means the same thing, leave it deleted.
- A real timeframe. Replace "recently", "often" and "regularly" with a date range or a count.
- No assumption. Nobody should have to accept a claim about themselves in order to answer. Add a filter question if they would.
- Options that fit everyone, once. No gaps, no overlaps, no missing top or bottom bracket.
- A balanced scale. The same number of steps above and below the midpoint, with matching label strength either side.
- An honest way out. "Not applicable", "I do not know" or "Prefer not to say" wherever a forced answer would be a guess.
- No double negatives. Read the stem together with the most negative option. If it needs a second pass, rewrite it positively.
- A decision behind it. Write down what you do if the answer is low and what you do if it is high. Different answers, or cut the question.
- Tested on a phone, by a stranger. Send the draft to someone outside the project and watch where they hesitate. Hesitation is the bug report.
Frequently asked questions
What makes a good survey question?
A good survey question asks about one specific thing, carries no opinion of its own, can be understood on a single read, and is attached to a decision you will actually make. If you cannot say what you would do differently based on the answer, the question is not earning its place.
How do you write survey questions?
Start from the decision, not the topic. Write down what you will do with the answer, then write the shortest question that gives you it. Use plain words, one idea, a stated timeframe, and answer options that cover everyone without overlapping. Then read it aloud to someone outside the project and rewrite anything you have to explain.
What is a leading question in a survey?
A leading question steers people toward one answer, usually through an adjective inside the question or a claim about what other people think. "How satisfied are you with our award-winning support?" is leading, because disagreeing now means contradicting the award. Remove the adjective and let the answer options carry the range.
What is a double-barreled question?
A double-barreled question asks about two things at once and only accepts one answer. "How satisfied are you with the speed and accuracy of delivery?" cannot be answered honestly by someone whose order arrived fast and wrong. Split it into two questions. The tell is an "and" or "or" joining two things you would act on separately.
What are the main types of survey questions?
Six formats cover nearly everything: single-select lists, select-all lists, agreement statements, numeric scales, short open text, and grids of statements on a shared scale. Closed formats are quicker to answer and to analyse; open text tells you things you did not think to ask.
Format guides: single and multi-select questions, agreement scales, numeric scales, open versus closed questions.
How do you write unbiased survey questions?
Strip adjectives from the stem, name both directions ("how easy or difficult"), give a concrete timeframe, balance the answer scale so there are as many positive steps as negative ones, and always offer a truthful way out such as "I do not know". Then check that answering does not require accepting an assumption about the respondent.
Bias that comes from who answers rather than how you asked is covered in the guide to response bias.
How many questions should a survey have?
Fewer than you want to ask. A transactional check after a purchase or support ticket works with one to three questions. A periodic feedback survey usually lands between five and twelve. The honest rule is not a number, it is a test: every question needs a decision behind it, and the ones that fail that test are the ones to cut first.
For a reference point: across the 87 survey templates in our own library, the median is 9 questions and none goes above 14. That is a house rule rather than a law of nature, but it is a workable target.
How long should a survey be?
Count effort, not items. A closed question takes a few seconds; an open text box takes far longer and costs you people. Tell respondents the real length up front, and make the estimate true. Two surveys with the same question count can feel completely different depending on how many text boxes and grids they contain.
What should a 3-question survey ask?
One outcome score, one reason, one filter. For example: "How easy or difficult was it to get what you needed today?" on a 1 to 5 scale; "What is the main reason for your score?" as optional short text; and one single-select that tells you which group the person is in, so you can compare segments later. Three questions is enough to spot a problem and enough to know who has it.
What are some examples of confusing survey questions?
The three that trip people most are double negatives ("Do you disagree that the policy is not flexible enough?"), vague quantifiers ("Do you exercise regularly?"), and overlapping answer brackets where $25,000 appears in two ranges. Each one is rewritten side by side in the six failure patterns above.
What is a survey question?
A survey question is one item in a questionnaire: a stem that asks something, plus the set of answers a respondent can choose from or write. Both halves are part of the question. A well-written stem with broken answer options produces data you cannot use, which is why the options deserve the same review as the wording.
What is the difference between a survey and a questionnaire?
The questionnaire is the instrument: the list of questions and answer options. The survey is the whole exercise, including who you ask, how you reach them, and what you do with the results. You can write a flawless questionnaire and still run a poor survey if you send it to the wrong people. Survey research methods covers the wider process.
Methods, limits, and sources
What this page is based on
Two things: published survey-methodology research, and a measurement of our own template library.
For the research, we prioritised split-ballot and randomised experiments over commentary, and we graded every claim. A means a meta-analysis or an effect replicated across independent studies. B means one strong study. C means well known but contested or dated. Every source in the list below was opened and checked against the number attributed to it before it was cited. Where we could not open a paper, we did not use its numbers.
For the library measurement, on 21 August 2026 we exported every live survey template on this site: 87 templates containing 799 questions. Question type, wording and position came from the structured record that the "use this template" flow loads into the editor, not from the rendered page, because that is the version a reader actually receives. Extraction succeeded on 87 of 87. Each failure pattern was detected by an explicit rule, then a sample of the matches was read by hand to estimate how often the rule was wrong. Where the hand check found a meaningful error rate, the published figure is the corrected one and the correction is stated. The one pattern with real judgment in it is the two-topic question, where the hand check found roughly a quarter of machine matches were fixed phrases such as "role and responsibilities" rather than genuine faults, so that figure is reported as approximately one in five rather than to a decimal place.
What this page does not tell you
Our 87 templates are one authoring pipeline, not a sample of the world's surveys. The 9-question median is our house style; it is not evidence about what length surveys generally are. The failure-pattern rates describe our library specifically, and a library written by different people would produce different rates. Treat them as a worked example of auditing a question set, not as a benchmark.
The published effect sizes come mostly from public-opinion research on general population samples, much of it by telephone or face to face. Whether a 14-point wording swing on a political attitude transfers cleanly to a customer feedback question on a phone screen is genuinely unknown. The direction of these effects is well established. The exact size in your context is not.
Two claims here are actively contested and are marked C in the table. Numeric scale labelling in particular has a 2025 cross-national replication that found the effect running the opposite way to the 1991 original, so treat it as unsettled rather than as guidance.
Disclosure
SuperSurvey makes survey software. Nothing on this page is for sale, there is no signup wall on any of it, and the question bank copies to your clipboard as plain text you can paste into any tool. The library we measured is our own, which is why the measurement section reports the results that do not flatter us alongside the ones that do.
References
- Holleman, B. (1999). Wording effects in survey research: using meta-analysis to explain the forbid/allow asymmetry. Journal of Quantitative Linguistics, 6(1), 29-40. doi:10.1076/jqul.6.1.29.4145
- Pew Research Center. Writing survey questions. Methods reference.
- Pew Research Center (2019). Good jobs vs jobs: survey experiments can measure the effects of question wording and more.
- Krosnick, J. A., and Presser, S. (2010). Question and questionnaire design. In Handbook of Survey Research (2nd ed.), 263-313. Full text.
- Saris, W. E., Revilla, M., Krosnick, J. A., and Shaeffer, E. M. (2010). Comparing questions with agree/disagree response options to questions with item-specific response options. Survey Research Methods, 4(1), 61-79. doi:10.18148/srm/2010.v4i1.2682
- Menold, N. (2020). Double barreled questions: an analysis of the similarity of elements and effects on measurement quality. Journal of Official Statistics, 36(4), 855-886. doi:10.2478/jos-2020-0041
- Pew Research Center (2019). Comparing forced-choice and select-all online survey responses.
- Presser, S., and Schuman, H. (1975). Question wording as an independent variable in survey analysis: a first report. Proceedings of the American Statistical Association, Social Statistics Section, 16-25. Cited without a link: the only copy we found online is served without a valid certificate.
- Galesic, M., Tourangeau, R., Couper, M. P., and Conrad, F. G. (2008). Eye-tracking data: new insights on response order effects and other cognitive shortcuts in survey responding. Public Opinion Quarterly, 72(5), 892-913. doi:10.1093/poq/nfn059. Open copy.
- Sandelin, F. (2022). The effects of questionnaire length on response rate, non-response bias, and data quality. SOM Institute, University of Gothenburg.
- Bottoni, G., and Aizpurua, E. (2025). Between two minds: the influence of numerical labels on survey responses in a cross-national study. Quality and Quantity, 60(2), 4595-4613. doi:10.1007/s11135-025-02436-9. Replicates and contradicts Schwarz, N., and others (1991), Public Opinion Quarterly, 55(4), 570-582.
- Revilla, M. A., Saris, W. E., and Krosnick, J. A. (2014). Choosing the number of categories in agree-disagree scales. Sociological Methods and Research, 43(1), 73-97. doi:10.1177/0049124113509605. We read the abstract and the corroborating result in reference 5; the full text was not reachable.
- Maybin, S. (2017). Busting the attention span myth. BBC, More or Less.
- Miller, G. A. (1956). The magical number seven, plus or minus two. Psychological Review, 63(2), 81-97. Full text. doi:10.1037/h0043158
- Pew Research Center (2019). Response rates in telephone surveys have resumed their decline.