How many responses does your score actually need?

Enter what you have and see the honest range around it, plus what it would take to tighten. This is the same arithmetic UserVane uses to decide whether a score is worth showing at all, so the answer here and the answer in the product agree.

The calculator

Your number, with its interval

Metric

Responses you have Promoters (9 to 10) Detractors (0 to 6)

Which number counts as a response

The number of people who answered, not the number you asked. If 5,000 saw the prompt and 40 answered, this is 40.

What people put instead. Entering the send count, or the number of monthly active users.

Why it matters. The interval is a statement about the answers you hold, so sending more without lifting the response rate moves nothing. This is also the number people quietly inflate when a score looks thin, and it is the one an auditor of your own analytics is most likely to check. Halving the interval takes roughly four times the responses, not twice, which is why chasing a tight margin on a low-traffic product is usually the wrong plan.

Why the split changes the interval

Enter the real counts. For NPS that is promoters (9 to 10) and detractors (0 to 6); passives sit in the middle and still count toward the total.

What people put instead. Assuming a 50/50 split, or reusing last quarter's shape.

Why it matters. A polarised sample is genuinely less certain than a uniform one at the same size, and NPS is a difference between two proportions rather than one, so it carries the error of both. That is the arithmetic reason an NPS needs materially more responses than a CSAT to reach the same interval, and why a team that switched metrics and kept the same target sample is now quoting a looser number than they think.

Limits

What this number is not

A margin is a narrow claim, and reading it as a general accuracy score is the mistake this page exists to prevent.

It does not tell you the score is right: The interval covers sampling noise: the luck of which people answered, among people equally likely to answer. Everything else that can make a score wrong is outside it.
It cannot see who chose not to answer: This is usually the bigger error and no sample size fixes it. If unhappy users churn before the prompt fires, or it only appears once someone reaches step five, you are measuring the people who stayed. A plus or minus 3 on that sample is a precise measurement of the wrong population, and it is more dangerous than a wide interval because it looks authoritative.
It does not make two numbers comparable: Quarter-over-quarter movement is only a change if the intervals do not overlap. Most reported swings of a few points are inside the noise of both samples, and the honest report is that you cannot tell yet.
It assumes the responses are independent: One frustrated team filing eight responses from the same account is one opinion counted eight times. The arithmetic cannot see that, and neither can any calculator.

The arithmetic

How it is calculated

Proportions use the Wilson score interval at 95%, which stays honest when everyone answers the same way. The common alternative reports a margin of zero at 100% agreement, which is obviously false and is why a naive calculator will tell you a five-for-five sample is perfect.

NPS is a difference between two proportions, so its interval is built from the Wilson bounds of the promoter and detractor rates paired at their worst case. That is wider than treating NPS as one proportion, and it is the honest width.

Below 10 responses UserVane suppresses the headline instead of printing a number, and under 100 it marks the score as still settling. Those thresholds are the product's, not this page's.

Every score in UserVane ships with this attached. The margin, the sample size, and where the responses came from, on the dashboard and over the API.

Start free Why a naked score lies