What sample size do I need for a survey?

Why 385 keeps appearing

Almost every sample size calculator returns 385 for the default settings, and the reason is arithmetic rather than convention. The formula multiplies the squared z score for the confidence level by the variability of the answer, then divides by the squared margin. At 95 percent confidence the z score is 1.96. The variability of a yes or no answer is largest when the true split is fifty fifty, which gives 0.25, so planning around it is the cautious choice. The margin of 5 points becomes 0.05, squared to 0.0025. That comes to 384.16, rounded up to 385. Notice that the population size does not appear at all. Once a population is large, adding more people to it barely changes the answer, which is why a national poll and a survey of a large city need roughly the same number of completed responses.

Subgroups are where the plan usually breaks

The margin of error in a report applies to the whole sample. The moment results are split, each slice carries its own, larger margin. Pew Research Center's explainer uses a national sample of 1,067 with a margin of 3 points, then shows that a subgroup of around 160 people carries a margin of about 8 points on the same question. A business survey hits the same wall faster. A festival survey with 385 responses sounds solid, but if the plan is to compare Friday, Saturday and Sunday and only 60 people came on Friday, the Friday figure has a margin of roughly 12.7 points at 95 percent confidence. Decide which comparisons the report will make before choosing the number, and size the survey around the smallest group that has to be read on its own.

Completed responses, not people invited

Every figure in the table is a count of finished, usable responses. The invitation list has to be larger by whatever share of people you expect to take part, which is the last step above. A realistic expected rate comes from your own previous surveys to the same audience rather than a published benchmark. Sending more invitations to reach the target is fine; lowering the bar on who counts as a response is not. Be careful about treating a larger list as a cure for a low rate, too. Inviting three times as many people gets you the number, but if the people who answer differ from those who do not, the extra responses make the result more precise without making it more accurate. That is a separate problem, and no sample size calculation touches it.

When the formula does not apply

The calculation assumes that the people who answered were drawn at random from the population, so that every member had a known chance of being included. A link shared on social media, a survey open to whoever visits a page, or a mailing list made up of whoever chose to sign up does not meet that assumption, and a margin of error calculated for it has no real meaning. In those cases the honest move is to report how many people answered and how they were recruited. At the other extreme, a population of a few dozen people is not worth sampling at all. Ask everybody, treat it as a census, and report counts, because the adjustment for small populations quickly shows that you would need most of them anyway.

Tracking the target in a form

Each submission is one row in the responses table, so progress towards the target is a count of rows. If results will be split by a group, ask that group as a single dropdown or radio question near the start, because each field becomes its own column in the CSV export and the option labels are written out as people saw them. That makes checking the size of each subgroup a simple filter while the survey is still open, early enough to send a targeted reminder. Test the form with the demo link, which keeps it working but records nothing, so test submissions never sit in the count.

Questions people ask