"How many respondents is enough for a survey?" — the question every research project starts with.
Search the web and you'll find "minimum 400" and "30% of employees" and a dozen other numbers floating around without justification. This article covers the real way to calculate sample size, plus the operational reality that "statistically sufficient" and "decision-sufficient" are not the same number.
The answer first — benchmarks by use case
| Research purpose | Recommended sample size | Reasoning |
|---|---|---|
| Board / exec decisions | 400+ | Large population, ±5% margin |
| Internal team decisions | 100–300 | Mid-scale, ±5–8% margin |
| Hypothesis directional check | 30–100 | Trend awareness |
| Individual deep-dive (with interviews) | 5–30 | Qualitative emphasis |
| Early prototype testing | 5–10 | UX validation |
"400 every time" is wrong. Size varies by purpose — that's the substance.
The basic formula
Strict statistical sample size is determined by three inputs:
n = (Z² × p × (1-p)) / e²
- n — required sample size
- Z — Z-value for the chosen confidence level (95% → 1.96, 99% → 2.58)
- p — population proportion (use 0.5 if unknown — produces the conservative maximum)
- e — acceptable margin of error (±3% → 0.03, ±5% → 0.05)
Common combinations
| Confidence | Margin | Required n |
|---|---|---|
| 95% | ±3% | ~1,067 |
| 95% | ±5% | ~385 |
| 95% | ±10% | ~97 |
| 90% | ±5% | ~271 |
| 99% | ±5% | ~664 |
The widely-cited "at least 400" comes from 95% confidence × ±5% margin.
Finite population correction
The formula assumes infinite population. For finite populations (e.g., 500 internal customers), apply the correction:
n_adjusted = n / (1 + (n - 1) / N)
- N — population size
Example: population 500, target n=385:
n_adjusted = 385 / (1 + (385 - 1) / 500) = 217
So for a population of 500, n=217 delivers the same precision as n=385 against an infinite population.
Smaller populations need smaller samples.
The hard part — "statistically sufficient" vs. "decision-sufficient"
The textbook calculation guarantees statistical sufficiency. In practice:
- Statistically insufficient is decision-sufficient sometimes
- Statistically sufficient is decision-insufficient other times
Case 1: Statistically insufficient, decision-OK
Situation: n=30 survey shows 80% dissatisfaction with feature A
Statistics: ±15% margin (95% confidence) — technically uncertain
Decision: 80% is so dominant → feature A clearly needs improvement
n=30 is statistically weak, but when an overwhelming majority is dissatisfied, you can act. You don't need n=400.
Case 2: Statistically sufficient, decision-insufficient
Situation: n=400 survey shows "overall satisfaction 4.0"
Statistics: Statistically robust, ±2% margin
Decision: "OK, so what do we change?" — invisible
n=400 is statistically clean, but a number that doesn't point to a next action is useless. Big samples can't fix sloppy question design.
Decision tree
Q1: What's the research for?
- Exec / public communication → strict stats → use the formula
- Internal directional → 100–300 is plenty
- Insight exploration → 5–30 with deep follow-up
Q2: Population size?
- Hundreds → with correction, ~50% of population is plenty
- Thousands to tens of thousands → use the formula
- Unknown → p=0.5, 95% / ±5% → ~400
Q3: How precise does the decision need to be?
- High (±3%) → 1,000+
- Medium (±5%) → ~400
- Low (directional) → under 100
Don't get hypnotized by sample size
Caution 1: "Everyone responds" is not the goal
It's tempting to chase 100% response, but the more you push for responses, the lower the quality drops.
Pulling in casual answerers degrades the data; getting a sufficient size from people who answer earnestly is more valuable.
Caution 2: "Big sample = correct" is wrong
n=10,000 with biased question design still produces wrong conclusions. Sample size is about noise reduction; it doesn't address question quality.
Caution 3: Segment cuts shrink samples fast
Total: n=400
Cut by industry (5): ~80 per industry
Then cut by size (3): ~27 per cell
If you want segment-level patterns, the total sample has to be sized for that — otherwise post-cut you have nothing meaningful.
Counters:
- Focus on a few high-priority segments
- Run dedicated segment-specific surveys
- Use segment cuts as "directional" rather than statistical
Caution 4: Open-text capacity is a separate axis
"I need 100 open-text responses" — even if your selected-question n=400 is met, you may still fall short. Open-text fill rates are typically 20–40%, so 100 open texts requires n=300–500 on the selected questions.
Working benchmarks by survey type
| Survey type | Practical sample size |
|---|---|
| Customer satisfaction (NPS, etc.) | 200–500 |
| Employee engagement | 80%+ of all staff |
| Consumer market research | 500–2,000 |
| A/B test | Calculate from statistical power |
| Usability test | 5–10 (qualitative emphasis) |
| Depth interview | 5–15 |
| New feature reaction | 50–100 |
| Candidate trend awareness | 30–100 |
These are working numbers, not strict statistical floors.
What matters more than sample size
In practice, 80% of the real problems aren't sample size — they're question design and response rate:
- Leading questions → biased regardless of n
- Response rates so low the formula stops applying
- Open text never used → utilization problem, not sample-size problem
Memorizing the formula matters less than running the question design → response rate → utilization cycle well.
Related reading:
Where Repoan fits
Repoan supports operating at an appropriate sample size:
- Auto response cap — easily cap at "first 300"
- Distribution targeting — segment-specific distribution for size control
- Response rate monitoring — auto-displays past distribution response rates
- AI sample size suggestions — describe the research; get an appropriate size recommendation
- Continuous-survey dashboard — month-over-month size management automated
Summary
Sample size:
- Formula: n = (Z² × p × (1-p)) / e²
- Exec decisions → 400+; team decisions → 100–300; insight exploration → 5–30
- Finite-population correction shrinks the requirement
- "Statistically sufficient" ≠ "decision-sufficient"
- Watch sample size when cutting by segment
- Question design and response rate quality matter more than the raw n
The right answer to "how many?" is "depends on what for." Back-calculate from purpose, population, and acceptable margin — that's the discipline.