How Many Survey Responses Do You Need? (Survey Sample Size Explained)

"How many responses do I need?" has a real, calculable answer. It isn't a guess, and it isn't "as many as possible." Your survey sample size is the number of completed responses that makes your results trustworthy, and this post walks through the actual statistics behind it: the formula, the two numbers that drive it, a worked example, and what to do when you can't reach the target. There's no rounding corners on the math here. Undersampling a survey and presenting it with false confidence is the mistake this whole topic exists to prevent.

By the end you'll be able to calculate your own target sample size, you'll understand what you're trading when you adjust it, and you'll know what to do honestly if you fall short.

The two numbers that decide your sample size

Before you touch the formula, you need two decisions, and both are judgment calls about what the survey is for, not statistical constants you can look up and forget.

Confidence level is how often this method would produce a correct result if you repeated the survey many times. 95% is the standard default: run the same survey process 100 times, and about 95 of those times the true population value falls inside your stated margin. 90% is common for faster, lower-stakes exploratory work. 99% is for situations where a wrong call is expensive, regulatory or safety-related research, mainly, and it costs you meaningfully more responses for the same precision.

Margin of error is how wide your result is allowed to be. A result of "60% satisfied, ±5%" means the true figure plausibly sits anywhere from 55% to 65%. Tighter margins need dramatically more responses, because the relationship isn't linear, precision only improves with the square root of sample size. Halving your margin of error roughly quadruples the required sample.

The formula

The standard starting point is Cochran's formula, which assumes you're drawing from an effectively unlimited population:

n₀ = Z² × p(1 − p) ÷ e²

• n₀, the sample size you're solving for
• Z, the z-score for your chosen confidence level (1.96 at 95%, 1.645 at 90%, 2.576 at 99%)
• p, the expected proportion answering a certain way, usually set to 0.5 when you don't know
• e, your margin of error, as a decimal (0.05 for ±5%)

The term p(1 − p) is largest when p = 0.5, which is why 0.5 is the default: it produces the biggest, and therefore safest, sample size estimate. If you have genuine prior data on how the population splits, using the real proportion will lower the required sample, but guessing at a non-0.5 value to shrink your sample artificially defeats the purpose.

The finite population correction

Cochran's formula assumes an infinite pool. Most real surveys draw from a countable group, 3,000 customers, 400 employees, and in that case you need fewer responses than the raw formula suggests, because each response represents a larger share of the whole:

n = n₀ ÷ (1 + (n₀ − 1) ÷ N)

Where N is your actual population size. This correction is why a 300-person company needs nowhere near 385 responses to hit ±5% at 95% confidence, it needs closer to 168.

A worked example

Say you're running an engagement survey at a company with 1,500 employees, and you want 95% confidence with a ±5% margin of error. You don't have prior data on how answers will split, so p stays at 0.5.

Step 1. Confidence of 95% gives Z = 1.96.

Step 2. Apply Cochran's formula: n₀ equals 1.96² × 0.5 × 0.5 ÷ 0.05², which is 384.16.

Step 3. Correct for the finite population of 1,500: n equals 384.16 ÷ (1 + 383.16 ÷ 1,500), which comes to 306.0.

Step 4. Round up: you need 306 completed responses, not 306 invitations.

If your historical survey response rate for this group runs around 40%, that means sending the survey to roughly 765 employees to land 306 completions. Sizing the send from the completion target, not the other way around, is the step most teams skip, and it's the most common reason a survey closes underpowered.

How confidence level and margin of error trade off against sample size

It's worth seeing a few combinations side by side, for an effectively unlimited population, so the cost of tighter precision is concrete rather than abstract:

Two things stand out here. First, going from ±10% to ±5% roughly quadruples the sample, and going to ±1% takes you into the thousands. Precision gets expensive fast, especially past ±3%. Second, moving from 95% to 99% confidence at the same margin costs you about 73% more responses, and that's rarely worth it outside regulated or safety-critical research.

What to do when you can't hit the number

Sometimes the honest number is out of reach. Your list is small, or your response rate is weak, and you won't hit the target no matter how long you keep the survey open. You've got a few legitimate ways to handle that.

You can widen the margin of error and say so: recalculate at ±8% or ±10% and report that wider range honestly. It's a narrower claim, not a false one. You can lower the confidence level and say so, too. Dropping from 95% to 90% cuts the required sample by roughly 30%, and stating "90% confidence" in your write-up keeps the claim honest. Or, better than either, you can improve the response rate instead of lowering your standards. A shorter survey, mobile-friendly formatting, a clear reason to respond and one well-timed reminder move response rates more reliably than almost anything else.

What you can't do honestly is collect 60 responses, report them as if they represent the full population, and say nothing about precision. That's how organizations end up making confident decisions based on noise. If you're short on responses, say so in the write-up. A result with a stated ±9% margin is still useful. A result presented with false precision isn't.

Common sample size mistakes

The most common one is sizing for the overall survey and then slicing the results by segment anyway. 385 responses gives you ±5% for the whole group, but split that into four segments and each one has under 100 responses with a margin closer to ±10%. Size against the smallest segment you intend to report on, not the total.

A close second is treating sample size and response rate as the same thing. A 70% response rate from a list of 80 people is a smaller, less reliable sample than a 15% response rate from a list of 4,000. What matters is the number of completions, not the percentage who replied.

Then there's the assumption that a bigger sample fixes a biased one. It doesn't, because sample size only controls random error. If the responses all come from your most engaged customers, a bigger sample just measures the wrong group more precisely. People also skip the finite population correction for a defined group, a single company, a known customer list, which means asking for more responses than they actually need. And plenty of teams report a number with no margin attached at all. "68% are satisfied" is incomplete on its own. "68%, ±5% at 95% confidence, n=390" is a result someone else can actually evaluate.

Try it

Rather than working the formula by hand every time, use the sample size calculator, enter your population, confidence level, and margin of error, and it applies the finite population correction automatically. It's worth knowing that this exact math is what sits underneath the calculator, so the number it gives you is the same one you'd get working it out on paper.

Once you've collected responses, you can run the reverse calculation with the margin of error calculator to see the actual precision your completed sample achieved.

Figure out your target sample size before you send, not after you've collected responses, it's the difference between a result you can defend and one you're hoping nobody questions too closely.