Survey sample size calculator

Find out how many people need to answer your survey before you can trust the results, or check how precise the responses you already have are.

Calculate

How many people you could survey. Leave blank if it's large or unknown.

Confidence level
Confidence level

Optional. Turns completes into invitations.

385completed responses for ±5% at 95% confidence.

How many survey responses do I need?

For most surveys, aim for 385 completed responses. That gets you within 5 percentage points of the true answer, 95 times out of 100.

A smaller audience needs fewer: about 278 from 1,000 customers and 218 from 500. Past about 20,000 people, the number barely changes.

Count completed responses, not invitations. If 1 in 5 people usually reply, invite five times as many. Add your response rate above and the calculator does it for you. To get more people to reply, see how to increase survey response rates.

More responses only help if the right people answer. If only your happiest customers reply, a bigger sample just makes you more sure about the wrong group. Plan a separate sample for each group you want to report on.

How the maths works

The calculator uses Cochran's formula for a proportion: n₀ = z² × p(1 − p) / e². z comes from your confidence level (1.96 for 95%), e is your margin of error, and p is set to 50%, the most cautious choice, because it gives the largest sample.

For a smaller population N, the finite population correction shrinks the answer: n = n₀ / (1 + (n₀ − 1) / N).

Examples at 95% confidence and ±5%: 377 from 20,000 people and 385 from 1,000,000. At ±3%, a large population needs 1,068.

How many interviews are enough?

Plan for 9 to 17 interviews for each type of person you talk to. That's the range where studies of interview research stop hearing new themes.

Stop when new interviews stop teaching you anything new. A narrow question with a similar group of people needs fewer. A broad question across different roles needs more.

Plan per group. Four customer types at 9 to 12 interviews each is 36 to 48 conversations, not 12.

Where these numbers come from

A systematic review of studies that measured saturation found it within 9 to 17 interviews, or 4 to 8 focus groups (Hennink and Kaiser, 2022).

Hearing every theme once isn't the same as understanding it. In one study, themes first appeared by interview 9, but fully understanding them took 16 to 24 (Hennink, Kaiser and Marconi, 2017).

A narrow aim, a specific group and good interview conversations all mean you need fewer people (Malterud, Siersma and Guassora, 2016, on information power).

How many users do you need for usability testing?

Five users per round finds most common problems. Test in small rounds, fix what you find, then test again.

Five isn't enough for rarer problems. A problem that affects 1 in 10 people shows up in a 5-person test about 41% of the time, and with 10 people about 65% of the time.

Chance of seeing a problem at least once

How many people the problem affects
5 users
10 users
20 users
31%5 users84%10 users98%20 users>99%
15%5 users56%10 users80%20 users96%
10%5 users41%10 users65%20 users88%
5%5 users23%10 users40%20 users64%

Need a number, like a task success rate or time on task? That's a measurement, not a hunt for problems, so test with at least 20 people and use the margin of error mode above to see how precise the result is.

Where these numbers come from

The table uses the problem discovery model 1 − (1 − p)ⁿ: the chance that a problem affecting a share p of people shows up at least once in n sessions.

NN/g's case for testing with 5 users assumes a typical problem affects 31% of people and recommends small rounds, with 3 to 4 people per group when you test two distinct groups.

Averages hide unlucky draws. When Faulkner resampled real test data, some groups of 5 found 99% of the problems and others found 55% (Faulkner, 2003).

For measurement studies, NN/g recommends at least 20 participants.

Next step

Get a sample size you can defend

The free Sample size planner Claude skill gives you a sourced range per segment, a saturation check across your interviews, and a paragraph you can paste into the research plan. To run the survey itself, Great Question surveys go straight to the customers in your panel.

Book a demo

No credit card required.

Questions

What does margin of error mean?

It's how far your result could be from the true answer. If 40% of people pick an option and your margin is ±5%, the real figure is probably between 35% and 45%. It covers the luck of who you happened to survey, not bias from who chose to reply or how a question was worded.

Which confidence level should I use?

Use 95% unless you have a reason not to. It's the standard, and what most readers assume. 90% is fine for a product decision you can revisit. Use 99% when a wrong call is expensive, and expect to need a much bigger sample.

Does the size of my audience matter?

Only when it's small. If you're surveying a big share of everyone there is, you need fewer responses. Above about 20,000 people it hardly matters: 377 at 20,000, against 385 for an unlimited audience.

Why does the calculator assume a 50% answer?

Because 50% needs the biggest sample, so it's the safe choice. If you planned for a 10% answer and the real answer was 50%, your results would be less precise than you expected. Planning for 50% means your sample is big enough whatever comes back.

Is five users enough for a usability test?

For finding common problems, usually yes. Five people will spot a problem that affects 31% of users 84% of the time, but a problem that affects 10% only 41% of the time. Test in small rounds, fix what you find, and test each distinct user group.