How to Calculate NPS: The Formula, a Worked Score, and What It Hides

Where the passives hide, and why the score can move without anyone changing their mind

The passives are the strange part of the design. They never appear in the numerator of either percentage, yet they are counted in the denominator of both, so their only effect is to shrink whatever the promoters and detractors would otherwise have contributed. In the worked example, 105 people who answered 7 or 8 dilute a 231-to-84 split from what would have been 73% versus 27% among the decided into 55% versus 20%. Take those 105 out and the score jumps from +35 to +47 without a single respondent revising their answer. That is a real property of the metric, not an artefact, and it means a survey that reaches more lukewarm customers scores lower than one that only reaches people with an opinion.

The second thing the arithmetic hides is that a single score is compatible with many different realities. A base of 55% promoters and 20% detractors gives +35, and so does one of 40% promoters and 5% detractors — the first has nearly four times as many actively unhappy customers as the second, and the score cannot tell them apart. Because of that, the three bucket percentages are worth reporting alongside the score every time. Detractors in particular deserve their own line: they are the only bucket that maps to something you can act on this week, and a rise in detractors masked by a simultaneous rise in promoters will leave the headline number flat while the support queue burns.

A score without a sample size and a response rate is close to meaningless

NPS is a difference between two proportions, so it carries the sampling error of both, and that error is larger than most people assume. The standard error is the square root of the promoter share plus the detractor share minus the square of the score expressed as a fraction, all divided by the number of responses. On the worked example that is the square root of (0.55 + 0.20 − 0.35²) ÷ 420, which is about 3.9 points, so the 95% interval around +35 runs from roughly +27 to +43. Run the same survey on 50 responses instead and the standard error rises to about 11 points and the interval spans +13 to +57. A quarterly report celebrating a move from +31 to +35 on fifty responses is celebrating noise.

The response rate is the other half, and it is the half that can bias the score rather than merely widen it. People with strong feelings answer surveys; the indifferent majority does not. At a 5% response rate you are hearing from the enthusiasts and the aggrieved and almost nobody in between, which typically means both tails are over-represented relative to the customer base. Worse, response rates move on their own — change the email subject line, the timing after purchase, or which segment gets the invite, and the score moves with it while the product stays exactly the same. Report the response rate next to the score, and treat any change in it as a reason to distrust the comparison until you have checked the composition of the two samples.

The 0 to 10 scale does not mean the same thing in every country

Cross-cultural survey research has documented response-style differences for decades, and they hit a metric like this one hard because the cut points are fixed and unforgiving. Some populations use the extremes of a rating scale freely; others treat the top of the scale as reserved and cluster around the upper middle. Under the NPS buckets, a respondent who is perfectly happy but culturally disinclined to award a 10 lands on an 8 and becomes a passive, and one who is mildly satisfied but scores generously lands on a 9 and becomes a promoter. The same underlying satisfaction therefore produces different scores in different markets, and the gap can be larger than any product change you are likely to ship.

The practical consequence for anyone running a survey in more than one language is simple: never rank markets against each other on absolute NPS, and never set one global target. Track each market against its own history instead, and compare the deltas, because a response-style bias that is roughly constant within a market cancels out when you look at the change. If a single company-wide number is unavoidable, report it as a weighted average with the per-market scores visible underneath, so nobody concludes that the German team is failing because their number is lower than the Brazilian team's. It probably is lower, and it probably always will be, and that says more about the scale than about either team.

Frequently asked questions

Is NPS a percentage?

No, and writing it with a percent sign is the most common error in the topic. It is the difference between two percentages, which makes it a plain number on a scale from −100 to +100 — a 201-point range, not a 0-to-100 one. The difference matters when someone tries to average scores or convert one into a share of customers: +35 does not mean 35% of anything, and averaging +35 and +55 into +45 is only valid if the two samples were the same size. Report it as +35, or as 35 points, and keep the promoter, passive and detractor percentages beside it.

What is a good NPS?

Anything positive means promoters outnumber detractors, which is a low bar, and the commonly repeated thresholds — above 50 is excellent, above 70 is world class — come from consultancy benchmarking rather than from any property of the arithmetic. They also travel badly, because the industry, the survey channel, the moment you ask and the language you ask in each move the number by more than most improvement programmes do. A supermarket and an enterprise software vendor cannot be held to the same figure. The only comparison that survives scrutiny is your own score against your own score, measured the same way, in the same market, at the same point in the customer lifecycle.

How many responses do I need before the score means anything?

It depends on how small a change you want to detect, and the arithmetic is unforgiving. At 50 responses the 95% interval around a score of +35 is roughly 22 points wide either side, so nothing short of a collapse is visible. At 420 responses it narrows to about 8 points either side. To reliably detect a 5-point move you need the interval to be narrower than that, which pushes the sample into the low thousands. Two practical adjustments: survey continuously and report a rolling window rather than a quarterly snapshot, and set the alert threshold on the detractor percentage rather than on the score, because a single proportion is estimated more precisely than a difference between two.