How many survey responses do you need?

Start with two questions: how precise must the answer be, and how confident must you be in that precision. Those are the margin of error and the confidence level, and once you fix them, the sample size falls out.

Margin of error is the ± band around your result. If 60% pick option A with a 5% margin, the true figure is somewhere between 55% and 65%. Confidence level is how often that band would contain the truth if you repeated the survey many times; 95% is the near-universal default. Tighten either one and you need more people.

For a large population at 95% confidence, these are approximate targets for a proportion using the conservative 50% assumption:

Notice the shape: halving the margin of error roughly quadruples the sample. Going from ±5% to ±3% is cheap; chasing ±1% is a different budget entirely. The 385-response target is a common starting point, subject to the sampling assumptions above.

The instinct is that a bigger population needs a bigger sample. It mostly does not. Sampling theory cares about the absolute number of responses, not the fraction of the whole, once the whole is large. The 385 that gives ±5% works for a city of 80,000 or a country of 80 million.

The exception is small, closed groups. If you are surveying a 300-person company, sampling 385 is impossible and unnecessary; you would aim for roughly 170 for the same precision, because each response now covers a bigger slice of everyone. When your population is under a few thousand, use a finite-population calculator. Above that, the round numbers in the previous section hold.

Every number so far is completed responses, not invitations sent. The gap between the two is your response rate, and it is almost always where sample plans go wrong. Decide the responses you need, then divide by the rate you honestly expect:

Big samples buy precise numbers. Plenty of good survey work does not need precise numbers. If you are running a usability study, sanity- checking a concept, or gathering open-text feedback, you are looking for themes and problems. A smaller sample can be useful for that purpose, though it cannot support precise estimates for the whole population.

The trap is subgroups. A sample of 400 feels comfortable until you slice it by four regions and two age bands and each cell holds a dozen people. If a comparison between segments is the point, size for the smallest cell you care about, not the total. That is usually what pushes a sample from hundreds into thousands.