Sample size for an expert survey is set by the decision it supports, not by a formula built for general populations. Thirty informed responses can carry a directional finding across one segment. The same thirty cannot carry four subgroup cuts, which is where most small expert surveys actually fail.
How many responses do you need?
Enough for the finest cut you intend to report. That is the whole rule, and it reframes the question from statistics to planning: decide the cuts first, then size for them — against what the screen can actually deliver.
Teams that size for the headline and then discover they want the data by region have not made a sampling error so much as a scoping one.
| General population survey | Expert survey | |
|---|---|---|
| Population | Effectively infinite | Frequently in the hundreds |
| Key question | Confidence against the population | Coverage of the population |
| What n=30 supports | Very little | A directional finding on one segment |
Why does the standard formula mislead?
Because it assumes a large population and random sampling, and an expert survey has neither. The relevant population may number in the hundreds, and the sample is drawn from whoever could be found and would respond.
In that setting, coverage of the population matters more than confidence against an infinite one. Thirty out of two hundred qualifying people is a very different claim from thirty out of two million, and the formula cannot tell them apart. Coverage of a small population is the survey equivalent of interview saturation.
What can thirty responses support?
A directional finding across a single segment, stated as such. That is genuinely useful — it can settle whether a market is moving in a direction — and it is not a percentage to one decimal place.
It can also support the qualitative material around it. Thirty structured responses plus the open-text answers are frequently more informative than the closed data alone, and three expert calls on top will usually explain them.
Thirty also happens to be about the point at which the open-text answers become readable as a set. Below that they are anecdotes; above it patterns start to repeat, which is a qualitative form of the same saturation logic that governs interview counts.
What happens when you cut the data?
It stops meaning anything. Three cuts of a thirty-response survey leaves cells of six or seven, and differences between cells that size are noise being presented as signal.
The failure is usually discovered in the readout, when someone asks for the data by segment and the analyst has to explain why the chart should not be drawn.
How should a small sample be reported?
With the n visible on every chart, the population described, and directional language rather than false precision. Reporting 43.3 percent from thirty responses implies a confidence the sample does not carry, which is the fastest way to lose a buy-side reader.
Naming the limitation in the deliverable also protects the finding. A directional result presented honestly survives scrutiny; the same result presented as a measurement does not.
Reporting the population is also what allows a later reader to judge the finding rather than accept it. Thirty responses from a qualifying population of two hundred is a statement about coverage that anyone can assess; thirty responses with no denominator is a number floating free of anything. The denominator is usually knowable from the screening data and almost never published alongside the result.