Sample size is the number of completed responses you need for your study to produce reliable estimates. Get it wrong and you're either spending too much money on respondents you don't need, or producing numbers stakeholders won't trust.
The math is well-understood. Here's the practical version.
Skip ahead: if you just want a number, use the free sample size calculator.
The four inputs
To calculate sample size, you need:
- Population size — total number of people you could theoretically survey (your customer list, the US adult population, etc.)
- Confidence level — usually 95%. The probability that the true population value lies within your reported range.
- Margin of error — usually ±5%. How precise the estimate needs to be.
- Expected proportion — usually 0.5 if unknown (most conservative). The expected percentage that will answer in a particular way.
Common rules of thumb
| Use case | Typical sample |
|---|---|
| National-population survey | 800–1,200 |
| B2B subscriber base, statistically meaningful | 300–400 |
| Concept test / ad test | 200–300 |
| Conjoint / MaxDiff | 250–400 |
| Qualitative interviews | 8–15 per segment until saturation |
| NPS tracker | 200+/wave for stable trend |
What changes the answer most
- Subgroups. If you need to report on segments (e.g., "results among 18–34 women in the West"), each segment needs its own meaningful sample. A 1,000-person national study cut into 12 segments leaves you with 80 per segment — too thin for confident segment claims.
- Margin of error. Cutting margin of error in half quadruples the sample size. ±3% is much more expensive than ±5%.
- Population size only matters when small. For populations over ~100,000, the population size barely affects the math. For populations under 1,000, it does.
The math (for the curious)
For a large population, the required sample size at 95% confidence and ±5% margin of error with p = 0.5:
n = (z² × p × (1 − p)) / e²
n = (1.96² × 0.5 × 0.5) / 0.05²
n = 384.16
Round up — 385 completes. For a finite population N, apply the finite population correction:
n_adj = n / (1 + (n − 1) / N)
Our calculator handles both forms.
Practical guidance
- Always plan for completes, not invites. Account for incidence rate (% who qualify) and completion rate (% who finish once they start). For a typical online study targeting a niche audience, you might invite 5,000 to get 400 completes.
- Pad for sub-segments you'll report on. If "results among new customers" is going to be a stakeholder slide, make sure the new-customer subsample is meaningful on its own.
- Don't over-target margin of error. ±3% sounds rigorous but the cost is ~3× the sample. Most business decisions don't need that precision.
Statistical power vs sample size
For studies that test specific hypotheses (e.g., "does this concept beat the control by ≥5 points?"), sample size depends on statistical power — the probability of detecting an effect if it's real. Standard target: 80% power.
Conjoint and MaxDiff have their own sizing rules — see those guides for specifics.
Related: Sample size calculator · Cross-tab analysis · Survey weighting