NPS Margin of Error: How Many Responses You Need Before a Change Is Real

Calculate your NPS freeDownload the NPS margin of error table (PDF)

Why NPS needs its own margin of error

Any score worked out from a sample of customers is an estimate of what all your customers would say. Ask a different 50 people and you would get a somewhat different score. The margin of error says how different, at a stated level of confidence.

The usual online margin-of-error calculators are built for one percentage, such as "62% said yes". NPS is the difference between two percentages taken from the same people (promoters minus detractors), and its uncertainty behaves differently. MeasuringU's article on confidence intervals for Net Promoter Scores notes that for a long time there was no well-defined method, describes Brendan Rocks' 2016 paper in The American Statistician that tested six methods, and reports that the best performer was an adjusted-Wald method. Its own tests on real survey data agreed.

The calculation, step by step

The method adds a small constant to the counts before working out the spread, which keeps small samples from looking more certain than they are. Here it is with a round of 50 responses: 25 promoters, 15 passives and 10 detractors, an NPS of +30.

1. Count the groups

n = 50, promoters = 25, detractors = 10. The free NPS calculator gives these counts if you paste the raw scores.

2. Adjust the counts

Add 3 to the total and 0.75 to both groups: n' = 53, promoters' = 25.75, detractors' = 10.75.

3. Adjusted shares

p = 25.75 ÷ 53 = 0.486. d = 10.75 ÷ 53 = 0.203. The adjusted NPS is 0.486 − 0.203 = 0.283, or +28.3.

4. Variance and standard error

Variance = p + d − (p − d)² = 0.486 + 0.203 − 0.283² = 0.609. Standard error = √(0.609 ÷ 53) = 0.107.

5. The margin

For 95% confidence multiply by 1.96: 0.107 × 1.96 = 0.210, or ±21 points. The score is +30; the interval is centred on the adjusted +28.3, so the true score for all your customers is plausibly anywhere from about +7 to +49. MeasuringU's worked example uses 90% confidence (1.645 instead of 1.96), which gives narrower intervals; the steps are the same.

The table: margin of error by number of responses

Running the same calculation for the same split (half promoters, a fifth detractors) at different sample sizes gives the 95% margins below. Scores with more detractors and a result near zero have slightly wider margins; scores near +100 have narrower ones.

Two patterns are worth remembering. Four times the responses only halves the margin, so going from 50 to 100 answers helps a lot and going from 400 to 800 helps little. And for a margin of about ±10 you need roughly 230 responses; for about ±15, roughly 105.

Comparing two rounds

The usual question is not "how precise is this score?" but "is this round really different from the last one?". The margin for a difference is wider than either round's own margin, because both rounds carry uncertainty. A quick approximation for two independent rounds: square each margin, add them, and take the square root.

Example: last quarter +32.5 from 40 responses (margin about ±24), this quarter +46 from 50 responses (about ±20). √(24² + 20²) ≈ 31 points. The scores differ by 13.5, well under 31, so the rise may be chance. Had both rounds had 400 responses (about ±8 each), the threshold would be about √(8² + 8²) ≈ 11 points, and a 13.5-point rise would count as real.

What this means for a small business

A business that finishes 30 jobs a month will not collect 230 answers a quarter. That does not make NPS useless; it changes how you read it.

• Do not react to one round. A move of 15 points between two rounds of 40 answers is inside the noise. Our guide to judging an NPS score works through exactly that case.
• Pool rounds. Combine the last two or three quarters into one rolling score with more answers, and compare rolling scores.
• Look for direction over time. Three or four rounds that all move the same way are more convincing than one big jump.
• Read every reason. With small numbers, the written reasons from detractors are often more useful than the score. Ten people naming the same slow reply is a finding at any sample size.
• Raise the response rate, not the mailing list. Asking every customer soon after the job, with a one-question survey, usually adds more honest answers than surveying old customers who barely remember you.

Keeping the count straight is easier in a spreadsheet (our NPS Excel formula guide includes a monthly breakdown with response counts) or in a survey tool that totals it for you; the free form builder's NPS template shows the score and the number of answers in its Summary tab. On paid plans, Fynl shows any survey answer given with an email address in that person's lead panel, so the customer behind a detractor score is one click away when you want to call them.

The NPS calculator sits with the lifetime value calculator, the form builder and the other no-cost tools on our free tools page.

Questions people also ask

Sources

• MeasuringU — Confidence Intervals for Net Promoter Scores
• Net Promoter System (Bain & Company) — Measuring your Net Promoter Score