A survey is a request for unpaid work. The respondent decides in about four seconds whether to spend two minutes on it, and that decision is made from the number of questions, the stated time, and whether the first question is easy. Everything else — the branding, the incentive, the design — matters far less than those three.
The second thing that kills surveys is leading wording. "How much did you love [feature]?" cannot return a negative answer without the respondent contradicting you. "How would you describe your experience with [feature]?" can. When a survey returns uniformly positive results, the usual cause is the question, not the product.
The four surveys worth sending
| Survey | Send when | Questions | What it tells you |
|---|---|---|---|
| Post-purchase | 2–7 days after delivery or onboarding | 3–4 | Whether the experience matched what was promised |
| Churn / cancellation | Immediately after cancelling | 2–3 | The actual reason you lost the customer |
| NPS / relationship | Quarterly, to the whole base | 2 | Trend over time, not a number in isolation |
| Product feedback | After sustained use of a feature | 3–5 | What to build or fix next |
Four surveys is already a lot. If you send more than this to the same person in a quarter, the response rate to each one drops — a pattern we saw repeatedly across our own sends.
1. Post-purchase survey
Use when: the customer has received the thing they paid for and used it at least once.
Q1 (rating, 1–5). Overall, how would you rate your experience with [company]?
Why it is worded this way: "Experience" rather than "satisfaction" because it invites operational comments — delivery, setup, support — not just a mood.
Q2 (open). What nearly stopped you from buying?
Why: This returns more usable material than "What did you like least?" It surfaces friction you can still see, rather than asking the customer to relitigate a decision they have already made.
Q3 (open). Was there anything you expected and did not get?
Why: The gap between expectation and delivery is the source of most complaints, and it rarely appears in support tickets because nobody thinks to file one.
Q4 (multiple choice). How likely are you to buy from us again? Much less likely / Less likely / About the same / More likely / Much more likely
2. Churn survey
Use when: immediately after a cancellation — within 24 hours, while the reason is still current.
Q1 (multiple choice, required). What led you to cancel? Price / Missing a feature I needed / Switched to another provider / No longer needed it / Something else
Q2 (open, conditional on "Something else"). What led you to cancel?
Q3 (open). What would have changed your mind?
Worth knowing: answers here are unreliable as a roadmap, and reliable as a mood read. Roughly a third of respondents write something specific; the rest write nothing or something polite. Treat it as direction, not instruction.
Two things about churn surveys: keep them to three questions or fewer, and do not attach a win-back offer to them. Once the survey becomes a sales touch, future responses skew.
3. NPS / relationship survey
Use when: quarterly, to the active base — and the value is in the trend, not the single score.
Q1. How likely are you to recommend [company] to a colleague or friend? 0 (not at all) to 10 (extremely likely)
Q2 (open). What is the main reason for that score?
The follow-up question is doing the real work. The number alone tells you something changed; the reason tells you what. Do not add demographic questions to make the sample look richer — every extra field costs completion, and most analyses never use them.
4. Product feedback survey
Use when: the respondent has used a specific feature enough times to have an opinion.
Q1 (task success). Were you able to complete [task]? Yes / Yes, but with difficulty / No
Why this matters: "Yes, but with difficulty" is the answer that predicts churn, and a yes/no question hides it.
Q2 (open). What were you trying to do when you got stuck?
Q3 (effort). How much effort did [task] take? Very low / Low / Moderate / High / Very high
Q4 (open). If you could change one thing about [feature], what would it be?
Six questions that bias the result
| Instead of | Write | Why |
|---|---|---|
| How much do you love X? | How would you describe your experience of X? | Removes the assumed positive |
| Do you agree that X is useful? | Did you find X useful? | Removes the requested agreement |
| Would you say our support is excellent? | How would you rate our support? | Stops supplying the adjective |
| Would you recommend us to others? | How likely are you to recommend us? | Yes/no loses the scale |
| How often do you use X? (daily / weekly) | In the past month, how many times did you use X? | Bucketed answers hide real frequency |
| What do you dislike about X? | What would you change about X? | Negative framing raises abandonment |
Where AI helps, and where it does not
Generating thirty candidate questions from a brief takes seconds, and it is genuinely useful for the first pass — you get coverage you would not have written by hand. What it will not do is notice that three of them are leading, or that two ask the same thing in different words. Read the generated set once, in one sitting, specifically looking for duplicates and assumptions. That read takes about four minutes and removes most of the damage.
The other place to be careful is the analysis of open answers. Summarising fifty free-text responses into themes is a reasonable use; taking the summary as the finding is not. If a decision rests on it, open the underlying answers.
Editing time, from our 30-day benchmark
Generating a first set of questions: seconds. Editing it into something worth sending: 6–10 minutes, almost all of it spent removing questions. Every survey we tested got better when we cut it, and none got better when we expanded it. If you take one thing from this page: send four questions, not twelve.
Frequently Asked Questions
How many questions should a survey have?
Three to five for anything sent routinely. Completion falls steadily after that, and the quality of later answers falls faster than the count — respondents who reach question twelve are answering to finish, not to inform.
Should I offer an incentive?
For a short survey to existing customers, usually not necessary — the relationship is the reason to answer. For a longer survey or a cold list, a small incentive raises completion meaningfully. Test it once rather than assuming.
Is NPS still worth measuring?
The trend is useful; the absolute score is hard to compare across businesses because the respondent base differs. If you keep measuring it, keep it to the two questions above and compare yourself against yourself.
Can I generate these questions with an AI writing tool?
Yes for the first draft, and it is a real time saving. Then read it for leading wording and duplicates, which generated sets contain more often than hand-written ones. The four-minute read is the part that matters.
What response rate should I expect?
It varies too widely by list type to give a useful general figure. What holds across our sends: short surveys to recent customers return far more usable answers than long surveys to anyone, and the gap between them is larger than any wording change.