How to Pick a Sample Size for a Survey That Actually Holds Up

Every survey I have ever run started with the same question from someone holding the budget: "How many people do we actually need?" The honest answer is that it depends on three numbers, and once you know them the arithmetic is fixed. Most people overshoot because nobody told them that a sample of 385 can describe a city of two million almost as well as it describes a town of twenty thousand. This guide walks through the three inputs that decide a survey sample size, shows the math on a real example, and explains why throwing more responses at the problem usually buys you nothing.

If you want to skip the theory and just get a number, the sample size calculator takes your confidence level, margin of error, expected proportion, and population, and hands back the count. The rest of this post is about understanding what that number means so you can defend it.

The Three Inputs That Set Your Sample Size

There are only three dials that matter, and they are easy to confuse.

Confidence level is how often your interval would capture the true value if you repeated the whole survey many times. A 95% confidence level means that across repeated samples, 19 out of 20 of the intervals you build would contain the real answer. The common settings are 90%, 95%, and 99%, each with its own critical z-value: about 1.645, 1.96, and 2.576 respectively.

Margin of error is the half-width of your result band. If 60% of respondents pick option A and your margin is 5%, the true figure sits somewhere between 55% and 65%. Confidence and margin are independent: raising confidence from 95% to 99% does not tighten your margin, it forces a bigger sample to hold the same margin steady.

Expected proportion (p) is your best guess at the split you are measuring. When you have no prior data, use 0.5. The term p(1 − p) in the formula peaks exactly at 0.5, where it equals 0.25, so that choice produces the largest, safest sample. If you already know a figure runs near 20%, entering p = 0.2 drops the requirement, because 0.2 × 0.8 = 0.16 is smaller than 0.25.

The core formula is n0 = z² · p(1 − p) / e². For 95% confidence and a 5% margin with p = 0.5, that is 1.96² · 0.25 / 0.05² = 3.8416 · 0.25 / 0.0025 = 384.16. You always round a sample size up, never down, so the answer is 385. That single figure is why almost every national poll you read quotes a sample of roughly 1,000 to 1,200 (a 3% margin) or about 385 (a 5% margin).

A Worked Example: Surveying a 4,000-Customer Base

Theory is cleaner with real numbers, so here is one I run often. Say you operate a SaaS product with 4,000 customers and you want to know what fraction would recommend it, accurate to within 5 points at 95% confidence.

Start with the infinite-population figure: 385, as computed above. But 4,000 is a known, countable group, so the finite-population correction applies:

n = n0 / (1 + (n0 − 1) / N)

Plugging in n0 = 385 and N = 4,000 gives n = 385 / (1 + 384/4000) = 385 / 1.096 ≈ 351. So you need 351 completed responses, not 385.

Now work backward to a campaign plan. If your survey completion rate runs around 20%, you need to email roughly 351 / 0.20 ≈ 1,755 customers to land those 351 answers. That is a defensible number to put in front of a stakeholder, and it is the kind of arithmetic the statistics calculator is handy for once the raw responses come back and you want the mean, standard deviation, and spread.

Notice what just happened: knowing the population was finite shaved 34 responses off the requirement. The smaller the population relative to your sample, the more the correction helps. Survey 500 employees and the savings are dramatic. Survey a whole country and they vanish.

Why Bigger Is Almost Never Better

The most expensive mistake in survey design is assuming that more responses always mean more accuracy. Two facts kill that assumption.

First, sample size barely moves once the population is large. The finite-population correction approaches n0 as N grows, because (n0 − 1)/N shrinks toward zero. At N = 10,000 a 95% / 5% survey needs 370; at N = 100,000 it needs 383; at N = 1,000,000 it needs about 384. The jump from a hundred thousand people to a million costs you one extra response. This is the counter-intuitive heart of sampling: you do not need a bigger sample to poll a bigger country.

Second, precision is expensive and the cost scales with the square of the margin. Sample size grows with 1/e², so halving the margin from 5% to 2.5% does not double the sample, it roughly quadruples it: 385 becomes about 1,537. Push to a 1% margin and you need close to 9,600. That is the single biggest cost driver in survey work, and it is exactly why commercial polls settle on a 3% to 5% margin instead of chasing 1%.

So the move is not "collect as many responses as possible." It is "decide the margin you actually need for the decision, then collect just enough." A product roadmap call that hinges on a 50/50 versus 55/45 split needs tighter precision than a sanity check on whether most users are happy. Match the sample to the decision, not to your anxiety.

Sizing an A/B Test or a Market Research Read

Surveys are not the only use. The same engine sizes an A/B test conversion sample. Suppose your landing page converts around 12% and you want to know whether a redesign moves it. Enter p = 0.12 and a margin tight enough to matter for the decision, and you get the per-variant visitor count. Because p sits well below 0.5, the required sample is smaller than the worst-case 385, so you can call the test sooner than the default estimate would suggest.

Reverse mode is the other half of this. A reader once objected that my 600-person poll "could not possibly" represent a city of two million. Feeding 600 responses at 95% confidence with p = 0.5 back through the formula returns a margin of error of about 4%. With the math attached, you can say the poll is accurate to plus or minus four points 19 times out of 20, regardless of how big the city is. That single reframe ends most arguments about whether a sample is "big enough."

For market research, the takeaway is to treat sample size as a budget conversation. Each tightening of the margin is a real line-item cost in recruiting, incentives, and time. Knowing that 1% precision costs 25 times what 5% does lets you have an honest discussion about whether the extra confidence is worth the spend.

A Quick Checklist Before You Field the Survey

Before you send anything, run through these:

Pick confidence first. 95% is the default for a reason; reserve 99% for high-stakes reads where being wrong is costly.
Set the margin from the decision. What is the smallest difference you need to detect? That sets e, not the other way around.
Use p = 0.5 unless you have real prior data. It is the safe, conservative choice and it guarantees your margin will be met.
Enter the population only when it is finite and known. A national audience or open web form should leave it blank; a 50,000,000 entry is wasted effort.
Round up, always. n0 = 384.16 becomes 385. Truncating to 384 leaves your true margin a hair above the 5% you promised.

Plug your three numbers into the sample size calculator, copy the result, and you have a figure you can stand behind in any meeting. The whole thing runs in your browser, the inputs live in the URL, and a shared link reopens the exact calculation for whoever questions it.

Made by Toolora · Updated 2026-06-13