What the Net Promoter Score measures
The Net Promoter Score compresses one survey question — “how likely are you to recommend us to a friend or colleague?”, answered from 0 to 10 — into a single number between −100 and +100. Fred Reichheld introduced it in the Harvard Business Review in December 2003, and its appeal is that it needs one question, produces one figure, and can be tracked over time by anyone.
The banding is deliberately asymmetric. Only 9s and 10s count as promoters. Everything from 0 to 6 — seven of the eleven possible answers — counts as a detractor. The 7s and 8s are passives and are discarded from the numerator entirely, though they stay in the denominator and so dilute the score. Reichheld's argument for that split was behavioural rather than statistical: in the customer data he studied, only the top two answers reliably predicted repeat purchase and referral, while a 7 or 8 predicted neither loyalty nor defection.
Because it is a difference of two shares, the score behaves unlike a percentage. It can be negative. It can stay flat while the underlying mix shifts substantially. And two companies with identical scores can have very different distributions — one polarised, one clustered in the middle — which is why you should always read the promoter, passive and detractor shares alongside the headline figure.
The formula, and why the score has a standard error
The arithmetic is a subtraction of two proportions. Divide promoters by total responses, divide detractors by total responses, subtract the second from the first, and multiply by 100. Passives never enter the numerator.
The more useful way to see it is as an average. Code every response as +1 if it is a promoter, 0 if a passive, and −1 if a detractor. The mean of those coded values is exactly the score divided by 100. That reframing matters, because a mean has a standard error and therefore a confidence interval, which a “difference of two percentages” does not obviously have.
For a variable taking values +1, 0 and −1 with promoter share p and detractor share d, the expected value is p − d and the expected square is p + d, since squaring turns both +1 and −1 into 1. The variance is therefore (p + d) − (p − d)2. Divide by the sample size, take the square root, multiply by 100, and you have the standard error in score points. Multiply by 1.96 for a 95% interval.
Two features of that variance are worth noticing. It reaches its maximum of exactly 1 when the sample splits evenly between promoters and detractors with no passives at all — the noisiest mix there is — and it falls to zero when every response lands in the same band, which is a limitation of the normal approximation rather than real certainty, and the reason a unanimous sample of 20 responses should not be reported with a zero-width interval. Passives cut the variance when the rest of the sample is close to balanced, because a passive contributes a certain zero rather than a coin flip; on a one-sided sample they raise it instead, by introducing disagreement where there was none. Ten promoters out of ten give a variance of 0, while eight promoters and two passives give 0.8 − 0.64 = 0.16.
Worked example: 900 responses, 420 promoters, 310 passives, 170 detractors
Run a quarterly survey, collect 900 usable responses, and sort them into the three bands.
- Total responses. n = 420 + 310 + 170 = 900.
- Promoter share. p = 420 ÷ 900 = 0.466667, or 46.67%.
- Detractor share. d = 170 ÷ 900 = 0.188889, or 18.89%.
- Score. 46.67 − 18.89 = 27.78, reported as +28.
- Variance of one response. (0.466667 + 0.188889) − (0.466667 − 0.188889)2 = 0.655556 − 0.077160 = 0.578395.
- Standard error. √(0.578395 ÷ 900) = 0.025351, so 2.535 score points.
- 95% margin of error. 1.96 × 2.535 = ±4.97 points.
- Interval. 27.78 ± 4.97, so roughly +23 to +33.
The interval is the whole point of the exercise. If last quarter's score on a similar sample was +25, this quarter's +28 is inside the noise: you cannot claim sentiment improved. To detect a three-point move at 95% confidence with this response mix you would need a sample several times larger, which the reference table below quantifies.
How to read the score: bands, trend and industry
Bain & Company, which co-developed the method with Reichheld, publishes interpretation bands: above 0 is good, above 20 is favourable, above 50 is excellent, and above 80 is world class. A score of +28 on the worked example above therefore sits in the favourable band.
Those cut-offs are far less informative than two other comparisons. The first is your own trend, judged against the confidence interval rather than against last quarter's point estimate. The second is your industry: response cultures differ enough that the same underlying satisfaction produces materially different scores across sectors, and Bain's own guidance is to benchmark within your category rather than against an absolute number.
A negative score means detractors outnumber promoters. That is a serious signal, but it is also common in categories where customers have no realistic alternative, so read it against competitors rather than against zero.
The most actionable output is not the score at all — it is the detractor count and the free-text follow-up question. Reichheld's framework treats the score as the trigger for a closed-loop process: contact detractors, find the cause, fix it. A score tracked without that loop is a dashboard ornament. And if you want to know whether sentiment translates into money, pair the score with your churn rate and retention rate, which move real revenue.
How many responses you need: 95% margin of error by sample size
| Responses (n) | Standard error (pts) | 95% margin of error (pts) | 95% interval around a score of +30 |
|---|---|---|---|
| 50 | 11.0 | ±21.6 | +8 to +52 |
| 100 | 7.8 | ±15.3 | +15 to +45 |
| 200 | 5.5 | ±10.8 | +19 to +41 |
| 400 | 3.9 | ±7.7 | +22 to +38 |
| 800 | 2.8 | ±5.4 | +25 to +35 |
| 1600 | 2.0 | ±3.8 | +26 to +34 |
Halving the margin of error requires four times the responses. A score reported to the nearest point from 100 responses is reporting one significant figure at best.
Response bias usually swamps sampling error
The interval this calculator produces only accounts for random sampling. It says nothing about who chose to answer. If you email a survey to 20,000 customers and 900 reply, the arithmetic above describes uncertainty in those 900 — not the gap between them and the 19,100 who ignored you, who are systematically different.
Two habits reduce the damage. Sample rather than blanket-invite, so response rate is a design choice you can measure. And hold the trigger constant: a score from a post-purchase prompt is not comparable to a score from an annual relationship survey, because the moment you ask changes the answer. Whatever you do, keep the question wording, scale and timing identical between periods, or your trend line measures your survey design instead of your customers.
Mistakes that corrupt an NPS number
- Averaging the 0–10 answers instead of banding them. The mean rating and the score are different statistics; a mean of 8.1 does not translate into a score.
- Excluding passives from the denominator. They belong in n. Dropping them inflates both shares and usually inflates the score.
- Treating 7 as neutral-positive. A 7 is a passive, and a 6 is a detractor. That single-point boundary catches people out constantly.
- Reporting one decimal place. Convention is a whole number, and on samples under a few hundred even the units digit is inside the margin of error.
- Comparing across survey channels. In-app prompts, email surveys and phone interviews produce different score levels from the same customer base.
- Chasing the score with incentives. Paying staff on NPS invites coaching customers toward 9s and 10s, which destroys the measurement without improving anything.
- Reading small moves as signal. Compare the change against the margin of error first. The reference table above shows how large that is at your sample size.
NPS versus CSAT, CES and the academic critique
NPS is one of three widely used customer metrics, and they answer different questions. CSAT asks satisfaction with a specific interaction, usually on a 1–5 scale, and is reported as the percentage choosing the top one or two options — use it to monitor a process. Customer Effort Score asks how easy the company made it to get something done, and predicts repeat contact volume well. NPS asks about willingness to recommend, which is a relationship-level judgement.
The score's original claim — that it is the single best predictor of growth — has not survived replication. Keiningham and colleagues, publishing in the Journal of Marketing in 2007, re-examined the underlying data and found no evidence that the net promoter measure predicts revenue growth better than conventional satisfaction measures. The reasonable position today is that NPS is a serviceable, cheap, comparable loyalty tracker, and not a growth forecast.
Use it accordingly: as one input beside metrics with a direct financial link. Net revenue retention and customer lifetime value tell you what loyalty is worth in cash; ARPU tells you what each account contributes now. And if you are testing whether a product change moved sentiment, the right tool is a designed experiment — see the A/B test significance calculator — not two consecutive survey waves.
Key terms
- Promoter
- A respondent who answers 9 or 10 on the 0–10 recommendation question. Coded +1 in the score.
- Passive
- A respondent who answers 7 or 8. Coded 0: counted in the total but not in the numerator.
- Detractor
- A respondent who answers anything from 0 to 6. Coded −1, so seven of the eleven answer options are detracting.
- Standard error
- The expected sampling variability of the score, in score points. Falls with the square root of the sample size.
- Margin of error
- The standard error multiplied by the z value for your confidence level — 1.96 at 95%. Half the width of the interval.
- Closed loop
- The follow-up process of contacting detractors, diagnosing the cause and fixing it. The part of the Net Promoter System that changes outcomes.