What is an NPS score, and how is it calculated?
NPS is the percentage of promoters minus the percentage of detractors, from one survey question. The formula, a worked example, and what the number cannot tell you.
What is NPS in one line: it is the percentage of promoters minus the percentage of detractors, from one survey question on how likely you are to recommend a company.
The question behind the Net Promoter Score: how likely is it that you would recommend this company to a friend or colleague, answered on a scale of 0 to 10. Respondents who answer 9 or 10 are promoters, those who answer 7 or 8 are passives, and those who answer 0 to 6 are detractors. These three bands are fixed. The score is a figure in points between −100 and +100.
How to calculate NPS
NPS = % promoters − % detractors.
Passives are counted in neither percentage, but they are part of the total both percentages are divided by, so they still move the score.
500 responses: 225 promoters, 200 passives, 75 detractors. Promoters 225 ÷ 500 = 45%. Detractors 75 ÷ 500 = 15%. NPS = 45 − 15 = 30.
Two things go wrong from here.
Calling it a percentage
It is points on a scale of −100 to +100, a difference between two proportions rather than a proportion. Why NPS is not a percentage covers this.
Dropping passives out of the total
Leave the 200 passives out: 225 ÷ 300 = 75%, 75 ÷ 300 = 25%. The score reads 50 instead of 30.
Same responses, twenty points of difference, entirely manufactured.
The number on its own is not enough
NPS was introduced as the one number that predicted company growth. The metric's own creator no longer asks the score to carry that argument. In 2021 Reichheld, writing with Darnell and Burns in "Net Promoter 3.0", Harvard Business Review, November–December 2021, moved it off the score and onto a measure taken from audited accounts rather than from a survey.
Two companies run the same survey.
Company A: 5% promoters, 90% passives, 5% detractors. NPS = 0. Company B: 50% promoters, 0% passives, 50% detractors. NPS = 0.
Company A's customers are almost all in the middle. Company B's are split in half and feel opposite things. Same number, opposite businesses, opposite things to do about it. The example is from Grisaffe (2007), "Questions about the Ultimate Question", Journal of Consumer Satisfaction, Dissatisfaction and Complaining Behavior, volume 20.
The bands hide two more things. A band is not a personality: a detractor is anyone from 0 to 6, so the label puts a customer who was mildly unimpressed alongside one who was furious. And passives are the least passive of the three: Fisher and Kordupleski (2019), in "Good and bad market research: A critical review of Net Promoter Score", argue the 7s and 8s are not indifferent but actively comparing alternatives.
The score cannot tell you what to do on its own, so it has to be set against something the customer did.
Use it like this
Numr CXM does this one way:
- Take one touchpoint, a single point of contact such as a service visit or a renewal, where a survey was answered.
- Split the respondents by how they rated it: favourable at the top of the scale, unfavourable at the bottom, the middle left out.
- Watch a window of time after the survey, its length set per touchpoint. Count who came back and did the next thing: renewed, bought again, came back for the next service.
- Divide the favourable group's return rate by the unfavourable group's. That ratio is the multiple.
30% of the favourable group came back, against 3% of the unfavourable group. The multiple is 10. Those who rated the touchpoint well were ten times more likely to do the next thing.
Three things make this worth doing.
It is the sharpest contrast you can get. The favourable group had the best experience of that touchpoint and the unfavourable group had the worst, and the middle of the scale is left out on purpose so that nothing blurs the two ends together. Because the middle sits outside the comparison, the multiple is larger than a comparison of the favourable group against everyone else. Read it as the gap between the best and the worst, not as an average across everyone who answered.
It is a count, not a model. Nothing is fitted, forecast or estimated. The platform counts what people did.
Any rating question works, and it needs no extra data from you. An effort or satisfaction question does the same job, so the two groups are not promoters and detractors; they are the top and the bottom of whatever scale you asked on. And the transaction that shows a customer came back is the same one that triggers the next survey, so the platform already holds everything the count needs.
Four limits.
- "More likely to", never "causes", "drives" or "proves". The two groups may differ in ways beyond their rating, so this is an observed difference in behaviour, not a controlled result.
- Roughly 385 respondents in each group, separately: that is the sample for a margin of error of about five percentage points at 95 per cent confidence. If one respondent in five is unfavourable, that is about 1,900 responses in all; at one in ten, nearer 3,800.
- Empty group, no multiple. Nothing is estimated, capped or filled in.
- Not available on Numr CXM Lite, which holds survey responses without the transactional record.
The multiple tells you a touchpoint matters. It does not tell you what to change there.
How to read your score
Calculate NPS the same way every time, and read the movement in your own score rather than its level. The bands are crude, but they have not changed in twenty years, so this year's score can be set against your own from last year. Then connect it to something the customer actually did.
