How many respondents do you need? A practical sample size guide

Where does the 385 come from?

The formula links three choices you make before fieldwork:

• Confidence level, expressed as z. At a 95% confidence level, z = 1.96; at 90%, z = 1.645; at 99%, z = 2.576.
• Margin of error, e. This is the half-width of the interval around a result: ±5 points is written as 0.05.
• Expected proportion, p. If you do not know the answer in advance, use 0.5. The term p(1−p) is largest at 0.5, so this choice gives the safest, largest sample.

The arithmetic for the classic case is short: 1.96² = 3.8416; multiplied by 0.5 × 0.5 = 0.9604; divided by 0.05² (0.0025) = 384.16. Sample sizes are always rounded up, which gives 385.

Read the result correctly. If 385 randomly selected respondents give a result of 50%, you can be 95% confident that the true value lies between 45% and 55%. For results further from 50% the margin of error shrinks: at an expected 20%, the same precision needs only 246 interviews.

Sample size by margin of error and confidence level

The table applies the same formula with p = 0.5 for a large population. It shows the trade-off that drives most survey designs.

Precision is expensive at the top end. Halving the margin from ±5 to ±2.5 points roughly quadruples the sample, from 385 to 1,537, because the margin of error falls with the square root of n. Before asking for ±2 points, check whether any decision would actually change between a result of 42% and one of 46%.

Does the size of the population matter?

Much less than most people expect. For populations above a few tens of thousands, the required sample hardly moves. Only when the population is small does the finite population correction make a visible difference:

n = n₀ ÷ (1 + (n₀ − 1) ÷ N), where N is the population size.

Bulgaria had 6,433,302 inhabitants in 2025 and the EU-27 had 451,127,411, according to the World Bank. Both need 385 interviews for ±5 points overall. The EU-wide survey is larger only if you want results for each country, because each country then becomes a subgroup with its own sample.

Small populations: customers, members and B2B buyers

With a small, known population, such as 500 business customers or the members of an association, the correction matters, but so does the response you can realistically get. The chart shows 218 completes for ±5 points out of 500, which means persuading 44% of all customers to answer.

In practice, small populations are usually surveyed as a census: everyone is invited, and the margin of error is calculated on the responses received. If 150 of the 500 customers reply, the margin is about ±6.7 points, once the correction is applied. That is usable for a customer satisfaction study, provided you check that the respondents resemble the non-respondents on size, region and length of relationship.

In B2B research, the unit also matters. If three people in each buying team influence the decision, decide in advance whether you are sizing the sample in companies or in individuals.

Plan for subgroups, not only for the total

The most common sample size mistake is to plan for the total and then report splits. A survey of 1,000 people has a margin of about ±3.1 points overall, but a region with 120 respondents in it has a margin of about ±8.9 points.

Work backwards from the smallest group you must read on its own. If the board needs results for each of four regions at ±5 points, you need about 385 per region, roughly 1,540 in total, and then weighting to restore the national picture.

Comparisons need even more. The margin on the difference between two independent groups is larger than the margin of either group: with 385 respondents in each, a gap must exceed about seven points before it is statistically meaningful. A reported five-point difference between men and women in such a survey may simply be noise. The margin of error calculator shows the interval for any result and base.

Incidence: how many people must you screen?

The sample you need is counted in qualified completes. When the target group is a fraction of the population, the number of people you must screen grows in proportion. The incidence rate is that fraction.

Take a survey of people who bought something online in the last 12 months, drawn from a general-population sample. Eurostat's 2025 figures give the incidence directly:

The same 400 completes require almost twice as many screening interviews in Bulgaria as in the Netherlands. Response rates then multiply the effort again: if one in ten people contacted agrees to take part, an illustrative figure, the Bulgarian study needs about 7,800 contacts. More country indicators that affect incidence are charted in our consumer data section.

For narrow targets, such as owners of a specific appliance or IT managers in mid-sized firms, incidence is often low enough that fieldwork time, not sample size, becomes the binding constraint.

Weighting, quotas and the effective sample

Two further adjustments change the real precision of a survey.

Weighting reduces the effective sample. When some respondents count for more than others, precision drops. The loss is expressed as a design effect: with a design effect of 1.4 (an illustrative value), 1,000 interviews behave like about 714, and the margin widens from ±3.1 to about ±3.7 points. Good reports state the effective base alongside the actual one.

Quota samples are not random samples. Most consumer surveys draw from online panels with quotas on age, gender and region. The margin-of-error formula strictly applies to probability samples, so for quota samples treat it as a guide to precision, not a guarantee. The same applies with more force to AI-generated answers: as our article on market research with AI explains, synthetic respondents have no sampling frame at all, so no margin of error can be calculated for them.

How to choose your sample size in five steps

1. Write down the decision and the smallest subgroup that must be read separately.
2. Choose the margin of error that would change that decision, not the smallest one you can afford.
3. Calculate the sample per subgroup and add the groups together.
4. Divide by incidence and expected response to estimate screening and contacts.
5. Allow for weighting and data cleaning, which remove or discount some interviews.

The sample size calculator handles steps 2 and 3, including the finite population correction. A survey that feeds a market model, such as the penetration and spend inputs in our TAM, SAM and SOM example, should be sized for the precision those inputs need. For designs with several countries, quotas or hard-to-reach targets, our survey research team can propose a sampling plan; ask for a proposal with your target group and deadline.