The percentage of promoters minus the percentage of detractors, on a 0-10 question. It measures stated intent, never a referral actually made.
In practice
NPS measures stated intent, never a referral actually made — the distinction that matters most here, because it is the one most often lost. Someone can answer 10 and never recommend anyone: the act needs an occasion, a mechanism and a reason, and a product can score well while providing none of the three. Two further limits are worth knowing before steering by it. The score crushes eleven answers into three buckets, so two opposite distributions can produce the same figure — 45/35/20 and 30/65/5 both give +25, and call for contrary decisions. And it is computed on those who answer, while disengaged users answer least: a score on 10% of the base describes the 10% still paying attention. Hence the rule, which costs nothing: publish the response rate and the three percentages next to the number.
The formula
NPS = (% of promoters, who answered 9-10) − (% of detractors, who answered 0-6). Passives, at 7-8, count for nothing
- The question
"How likely are you to recommend us, from 0 to 10?" — one question, always the same wording, always at the same moment in the lifecycle. Change any of the three and the series before and after are not comparable.
- The buckets
Three groups from eleven answers, which throws away most of the information on purpose. A 7 and a 0 are both "not a promoter", and only one of them is a person about to leave.
- The response rate
The number that has to be published beside the score and almost never is. A score computed on 10% of your users describes the 10% who felt strongly enough to answer.
This is a measure of stated intent, not of referrals made. Someone can answer 10 and never refer a single person; the viral coefficient counts what actually happened. Treating this score as a referral metric is the most common way it gets misused.
Worked example
The same score, two opposite companies
- Company A: 45% promoters, 35% passives, 20% detractors. NPS = 45 − 20 = +25.
- Company B: 30% promoters, 65% passives, 5% detractors. NPS = 30 − 5 = +25. Identical score, and nothing else about them is.
- A has real enthusiasm and a real problem: a fifth of respondents are unhappy enough to say so, and they are worth interviewing this week. B has neither — two thirds of people are indifferent, which is a harder situation and produces no complaint to act on.
- And the trap underneath both: 200 answers out of 2,000 users surveyed is a 10% response rate, and the 1,800 who ignored it are not neutral — disengaged users answer surveys least. The score describes the people still paying attention.
The score is a summary of a distribution you already have, and the distribution is where the decisions are. Keep the three percentages and the response rate next to the number, and it becomes useful; publish the number alone and it becomes a target to optimise.
Orders of magnitude
- Industry tables of "good NPS" are widely quoted and worth very little, because the score depends heavily on how and when you ask. Two companies surveying at different lifecycle moments are comparing survey designs as much as products.
- The defensible use is your own trend, with the question, the timing and the channel all held constant. Change any one of them and you have started a new series, whatever the chart says.
- A response rate under about 10% makes the score mostly a measure of who answers. That is not a reason to drop the survey — it is a reason to publish the rate, so the number is read with the weight it deserves.
How to improve it
- Publish the response rate and the three percentages next to the score, every time. This is the cheapest change on the list and it prevents the two failure modes above at once.
- Read the free-text answers and treat the score as the index, not the finding. The sentence a detractor wrote is worth more than the fact that they were one.
- Ask at a consistent moment in the lifecycle — thirty days after activation, say — rather than to everyone at once. A campaign sent to your whole base mixes people who signed up yesterday with people who have used the product for three years.
- Do not use it as a referral metric. If what you want to know is whether people recommend you, count the referrals: the viral coefficient measures the act, this measures the intention, and the gap between the two is often the whole story.
In the Tour
If it exists, is it actually used by customers?
- Yes, meaningfully20 pts
- A little, rarely7 pts
- Doesn't exist or never used0 pts
The Tour asks whether your sharing or referral mechanism is actually used by customers — 20 points for yes and measured, 7 for a little, 0 for not really. That question and this score are not the same thing, and the difference is the point: a high NPS says people would recommend you, while this asks whether they did. A team quoting a strong score and answering 0 here has measured an intention and never built the path for it to become an act, which is a fixable problem and a very common one.
Questions people ask
What is a good NPS?
The question has no answer that survives leaving your own survey, because the score moves with the wording, the timing and the channel far more than most people expect. The version worth asking is whether this quarter beats the last one with all three held constant — and whether the detractors are saying the same thing they said then.
Does NPS predict referrals?
Weakly, and it is worth saying plainly. It measures whether someone says they would recommend you, which is a different act from recommending you — the second needs an occasion, a mechanism and a reason, and a product can score well and provide none of the three. Measure the referrals themselves if that is the question.
NPS or CSAT?
CSAT — customer satisfaction score — asks about one interaction that just happened and is the better tool for judging a support reply or a feature. NPS asks about the relationship as a whole and is the better tool for a trend across quarters. They are not competitors, and running both is common — the mistake is using either to judge what the other was built for.
How often should I ask?
Rarely enough that answering does not become a chore — twice a year for the same person is plenty, and a rolling survey to a different slice each month gives a continuous series without asking anyone twice. Asking quarterly at scale mostly trains your users to dismiss the dialog.