How Many Survey Responses Make Results Reliable, and How Many Interviews Are Enough? (A 2026 Sample Size Guide for Business)
Short answer: For a survey meant to estimate a general consumer market, 385 responses give a margin of error of about plus or minus 5 percent at a 95 percent confidence level, the standard used in business work, so in practice teams usually collect 400. In-depth interviews typically reach data saturation at 8 to 15 interviews per target group.
But these numbers only hold when you analyze a single overall group. If you need to break results down by subgroup, every number changes. This article explains the right way to think about sample size, with tables you can actually use as a reference.
Sample Size Calculator
At 95% confidence, results will be within ±5%
How it's calculated
Cochran's formula: n₀ = z² × p(1−p) ÷ e², where z is the standard score for the confidence level (1.645, 1.96, 2.576), p is the expected proportion and e the margin of error. With a known population N, the finite population correction applies: n = n₀ ÷ (1 + (n₀ − 1) ÷ N). This assumes simple random sampling; with non-random or hard-to-reach samples the real requirement may be higher.
How many survey responses make results reliable?
Survey reliability is measured by the margin of error, which tells you how far your result may deviate from the true value.
Margin of error at a 95 percent confidence level, for a large population:
- —50 responses: margin of error about ±13.9 percent
- —100 responses: margin of error about ±9.8 percent
- —200 responses: margin of error about ±6.9 percent
- —300 responses: margin of error about ±5.7 percent
- —385 responses: margin of error about ±5.0 percent (the business standard)
- —600 responses: margin of error about ±4.0 percent
- —800 responses: margin of error about ±3.5 percent
- —1,000 responses: margin of error about ±3.1 percent
- —1,500 responses: margin of error about ±2.5 percent
- —2,000 responses: margin of error about ±2.2 percent
How to read the table: if you collect 400 responses and find that 60 percent of respondents are interested in the product, the true figure in the market is likely between 55 and 65 percent, which is precise enough for most decisions.
A key point people often get wrong: population size has almost no effect on how many responses you need. Whether your market has 50,000 people or 5 million, 385 responses give roughly the same precision. What determines precision is the number of responses, not their share of the population.
When fewer than 385 responses is acceptable
Many projects do not need to reach 385 responses.
When the population is limited. For example, if you have 300 corporate clients in total, 169 responses already give a margin of error of about ±5 percent, because of the finite population correction. In B2B cases where clients number in the hundreds, the number you need is much lower than you might expect.
When you are comparing before and after. For example, measuring satisfaction this quarter against last quarter. If you use the same collection method every time, a sample of 200 is enough to see the trend, because systematic errors cancel each other out. Random sampling error, however, does not cancel out. At 200 per wave, a result has to move by more than about 10 percentage points before you can be confident the change is real; a shift of 3 to 5 points between quarters may be nothing more than sampling fluctuation.
When the difference you care about is large. If you only need to know which group clearly prefers something more, rather than a precise figure, 150 to 200 responses may be enough.
When the results support internal decisions, not publication. Work used for decisions inside the organization can tolerate a higher margin of error, as long as decision-makers know how wide the margin of error on each figure is.
The lowest usable minimum for statistics: 30 responses per group
400 responses is a good target, but in reality many projects fall short: the target group is hard to reach (executives, physicians or corporate clients), budget or time is limited, or response rates are low. So the question is not just how many responses you should collect, but what you can conclude from the number you actually have.
The widely used minimum is at least 30 responses per group you plan to analyze. Roscoe (1975) proposed this rule: when a sample is split into subgroups, each subgroup should have at least 30 cases. The reasoning comes from the Central Limit Theorem: at a size of about 30 or more, the distribution of the sample mean is close enough to a normal distribution to use standard statistics. A common misunderstanding is that 30 responses do not make the data normally distributed, and they do not mean the results are precise enough. At 30 responses the margin of error is still as high as about ±18 percent, so this is only the starting point at which statistics can be used, not a level suited to important decisions.
| Responses per group | Margin of error (95%) | What you can conclude |
|---|---|---|
| 30 responses | ± 17.9% | Minimum for statistics based on the normal distribution; use for a rough sense of direction or a pilot |
| 50 responses | ± 13.9% | Shows very large differences between groups, e.g. 30% vs. 70% |
| 100 responses | ± 9.8% | Recommended minimum per subgroup in a report used for decisions |
| 200 responses | ± 6.9% | Comparing groups, or tracking trends over time |
| 400 responses | ± 4.9% | Estimating the overall market for important decisions |
What to do if you fall short of your target
- —Always report the actual margin of error. Use the calculator above in "already collected" mode and state the margin of error alongside every figure. Decision-makers will then know how far each number can be trusted.
- —Do not break out subgroups with fewer than 30 responses. If you must look at them, report them as qualitative observations, not percentages.
- —Apply the finite population correction when the population is small. For example, with corporate clients numbering in the hundreds, the number you need drops considerably.
- —Supplement with in-depth interviews. When numbers are small, explaining why carries more weight than estimating how much.
- —Watch for non-response bias. People who respond may differ systematically from those who do not. This cannot be fixed by collecting more responses, and it often matters more than sample size.
The most expensive trap: subgroup analysis
This is where research projects waste the most money.
If you collect 400 responses and then want to analyze them by 4 age groups, each group is left with about 100 responses, which carries a margin of error of nearly ±10 percent. That means if group A answers 55 percent and group B answers 48 percent, you cannot conclude that the two groups really differ, because, at about 100 per group, the gap between two groups has to exceed about 14 percentage points to count as a real difference at 95 percent confidence, so a 7-point gap cannot yet be told apart from random sampling error. Checking whether the two groups' margins of error overlap is not a valid test.
Practical rules:
- —Every subgroup you intend to analyze separately should have at least 100 responses, and 150 to 200 if you want to compare groups with confidence.
- —Work backward: if you need to analyze 5 subgroups, you need at least 750 to 1,000 responses, not 400.
- —Decide how many subgroups you will analyze before collecting data. Thinking about it afterward is the number one reason research fails to answer the important questions.
- —If the budget is not enough, analyze only the groups that truly matter to the decision and collect enough for those groups (over-sampling), instead of spreading the sample thinly across every group.
Common sampling methods and how they affect sample size
The formula in the calculator assumes simple random sampling, in which everyone in the population has an equal chance of being selected. In practice, other sampling methods are widely used, and each changes the number you need in a different way, following the principles in Cochran (1977) and Kish (1965).
| Method | Principle | Effect on sample size | Best when |
|---|---|---|---|
| Simple random | Everyone has an equal chance of being selected | The calculator formula applies directly | You have a complete list of the population |
| Systematic | Select every k-th entry from a list | Close to simple random sampling, provided the list is not ordered in a repeating pattern | You have an ordered list, such as a customer database |
| Stratified | Divide the population into strata, such as regions or customer segments, then sample within each stratum | As precise as or more precise than simple random sampling at the same size, and lets you fix the number per subgroup | You need subgroup analysis, or groups differ greatly in size |
| Cluster | Sample groups, such as branches or villages, then collect from people in the selected groups | Usually requires a larger sample, multiplied by the design effect; e.g. if it equals 2, you need twice as many | There is no list of individuals, or you want to save on fieldwork travel costs |
| Quota | Set proportions by gender, age or region, then collect until quotas are filled, without random selection | The formula can serve as a guide, but the calculated margin of error does not cover selection bias | Commercial market research and online panels, the most common case in practice |
| Purposive / Snowball | Select informants who meet criteria, or have informants refer the next person | No statistical formula; use data saturation as the criterion instead | Qualitative research and hard-to-reach groups |
Stratified sampling is the way out of the subgroup trap. If you know from the start that you need to analyze 4 regions, set a sufficient number per region, for example 100 to 150 responses per region, instead of sampling everyone together and hoping each region gets enough. If some regions are collected above their true share (over-sampling), apply weighting back to population proportions when summarizing the overall picture.
How many in-depth interviews are enough?
Qualitative research does not follow the same logic as quantitative research, because it does not aim to estimate values for a population but to understand mechanisms and reasons. The criterion used is data saturation, the point at which additional interviews yield no new themes.
Empirical research gives consistent numbers. Guest, Bunce and Johnson (2006) analyzed 60 interviews and found that almost all themes appeared within the first 12, with the core themes appearing from about 6. Hennink, Kaiser and Marconi (2017) split saturation into two levels: capturing every theme (code saturation) at about 9 interviews, and understanding the meaning of each theme in enough depth (meaning saturation) at about 16 to 24. A systematic review by Hennink and Kaiser (2022) found that most interview studies reached saturation at 9 to 17 interviews when the target group was fairly homogeneous and the research question was clear. For focus groups, Guest, Namey and McKenna (2017) found that 2 to 3 groups covered about 80 percent of themes, and 3 to 6 groups covered about 90 percent.
The right number also depends on how rich each interview is. Malterud, Siersma and Guassora (2016) proposed the concept of information power: the number of informants you need falls when the study aim is narrow and clear, the informants closely match the target group, the study is supported by a theoretical framework, the dialogue is of high quality, which depends directly on the interviewer's skill, and the analysis focuses on in-depth case analysis. So a narrow, clear question may yield enough data from about 6 interviews, the point by which core themes usually appear (Guest et al., 2006), while a complex question or a diverse group may need more than 20. An experienced interviewer may confirm a very narrow point with fewer than 6, but that should be treated only as a preliminary check.
| Situation | Interviews per target group | Reason |
|---|---|---|
| Narrow, clear question, such as confirming a hypothesis or testing a single concept; similar informants; experienced interviewer | About 6 | Core themes usually appear within about 6 interviews (Guest et al., 2006); an experienced interviewer may confirm a very narrow point with fewer, but that should be treated as a preliminary check |
| Typical question, such as finding the pain points or behaviors of one customer group | 9 to 12 | The range in which most themes are captured (Guest et al., 2006; Hennink et al., 2017) |
| Broad exploratory question, a need to understand the underlying reasons in depth, or a diverse group | 16 to 24 | Requires meaning saturation (Hennink et al., 2017) |
The practical conclusion: about 12 interviews per target group is a safe starting point for most questions. You can go lower when the question is narrow and the interviewer is highly experienced, and you should plan for more when the question is complex. Stop when you actually reach saturation, not when you hit the number you set in advance.
Numbers that work in practice:
- —In-depth interviews: 8 to 15 per target group. Most key themes usually emerge within 10 to 12 interviews. If the target group is highly diverse, increase to 15 to 20.
- —Focus groups: 3 to 4 groups per target group, with 6 to 8 people each. Fewer than 3 groups risks capturing the dynamics of one particular group rather than the views of the market.
- —Ethnography (observing real behavior): 6 to 12 participants is usually enough, because the data per participant is much deeper.
- —If there are several target groups, multiply by the number of groups. For example, with 3 target groups you should conduct 30 to 45 interviews in total, not 12 split among them.
The sign that you have reached saturation is when 3 consecutive informants provide no new themes beyond those already recorded. In practice, plan with some buffer and stop at saturation, rather than fixing the number rigidly from the start.
Numbers aren't everything: three things that matter more
In our experience, most research that goes wrong does not fail because the sample was too small, but because of these three things.
1. The sample does not match the target group. 1,000 responses from people who are not your target customers are worth less than 200 from the right group. This problem is common with cheap online panels that screen loosely. What you need to ask the vendor is how respondents are screened and what the incidence rate is.
2. Poorly designed questions. Leading questions, questions that measure politeness instead of purchase intent, or questions that use terms customers do not actually use produce figures that are precise about something meaningless. This is why we always recommend doing qualitative research before designing a questionnaire.
3. No step to explain the numbers. When results come out unexpected, many projects choose to interpret them on their own instead of going back to real people for an explanation. The most valuable findings are often right here.
References
- —Cochran, W. G. (1977). Sampling Techniques (3rd ed.). John Wiley & Sons.
- —Guest, G., Bunce, A., & Johnson, L. (2006). How Many Interviews Are Enough? An Experiment with Data Saturation and Variability. Field Methods, 18(1), 59–82.
- —Guest, G., Namey, E., & McKenna, K. (2017). How Many Focus Groups Are Enough? Building an Evidence Base for Nonprobability Sample Sizes. Field Methods, 29(1), 3–22.
- —Hennink, M. M., Kaiser, B. N., & Marconi, V. C. (2017). Code Saturation Versus Meaning Saturation: How Many Interviews Are Enough? Qualitative Health Research, 27(4), 591–608.
- —Hennink, M., & Kaiser, B. N. (2022). Sample Sizes for Saturation in Qualitative Research: A Systematic Review of Empirical Tests. Social Science & Medicine, 292, 114523.
- —Kish, L. (1965). Survey Sampling. John Wiley & Sons.
- —Malterud, K., Siersma, V. D., & Guassora, A. D. (2016). Sample Size in Qualitative Interview Studies: Guided by Information Power. Qualitative Health Research, 26(13), 1753–1760.
- —Roscoe, J. T. (1975). Fundamental Research Statistics for the Behavioral Sciences (2nd ed.). Holt, Rinehart and Winston.
Frequently asked questions
- Are 400 responses enough for nationwide market research?
- Enough for the national overall picture, with a margin of error of about ±5 percent. But not enough if you need to analyze by region or age group, because each subgroup will be left with too few responses. If you need to split into 4 to 6 groups, plan for 800 to 1,200 responses.
- What is the minimum number of survey responses to collect?
- The widely used minimum is at least 30 responses per group you will analyze, following Roscoe (1975) and the Central Limit Theorem, which allows standard statistics to be used. But at 30 responses the margin of error is still about ±18 percent, so it suits a sense of direction or a pilot. If the results will be used for decisions, you should have at least 100 responses per group.
- What is stratified sampling and how does it help?
- It means dividing the population into strata, such as regions or customer segments, and then sampling within each stratum. It ensures each subgroup has enough responses for separate analysis, and at the same sample size it is usually as precise as or more precise than simple random sampling. If some strata are collected above their true share, you must weight them back when summarizing the overall picture.
- How do you calculate sample size?
- The basic formula for a large population is n equals Z squared times p times (1 minus p), divided by e squared, where Z is the value for the confidence level (1.96 at 95 percent), p is the expected proportion (use 0.5 when unknown, because it gives the largest and safest sample), and e is the acceptable margin of error. Plugging in a 5 percent margin of error gives 385 responses. In practice, collect an extra 5 to 10 percent to allow for incomplete questionnaires.
- Why do national polls use only 1,000 to 1,500 respondents when the population is tens of millions?
- Because precision depends on the number of respondents, not their share of the population. At 1,000 respondents the margin of error is about ±3 percent, which is enough for reporting the overall picture. Increasing to 10,000 respondents reduces the margin of error to about 1 percent, which is not worth the tenfold increase in cost for most work.
- How can I present findings from only 10 interviews to the board?
- Present what qualitative research can genuinely claim: mechanisms and reasons, not market proportions. A compelling presentation states how many people were interviewed, from which groups, and which themes recurred in almost every interview, supported by real quotes. A finding that appears in 9 out of 10 interviews carries strong explanatory weight, even though it is not a statistical figure. Research by Guest, Bunce and Johnson (2006) found that almost all themes appeared within the first 12 interviews, which can be cited to support the adequacy of the number. If the question is narrow and the interviewer is experienced, even fewer interviews may be enough, according to the concept of information power (Malterud et al., 2016).
- Do online surveys and face-to-face interviews need different sample sizes?
- The number required statistically is the same, but data quality differs. Online surveys carry the risk of inattentive respondents and looser screening, so they should include attention checks and allow for discarding about 10 to 20 percent of bad data. Face-to-face interviews produce higher-quality data but cost several times more per response.
The margin-of-error figures in this article are calculated at a 95 percent confidence level using an expected proportion of 0.5, the case that requires the largest sample, for a large population and simple random sampling. The actual figures for your project may differ depending on the sampling design and the characteristics of the population. Findings from the cited research are summarized; see the reference list for details. This article was prepared by NXT Consulting Group and translated from the Thai original.
More in this series
Share this article
