NPS, or net promoter score, is a single number between minus one hundred and one hundred, produced from one question about how likely somebody is to recommend you.
It is the most widely used customer measurement in software and it is the one most often asked to do work it cannot do.
This page is what it is, how the arithmetic works, and where the number runs out. We do not calculate it, sell it or store it, which is the reason this page can be blunt about it.
How is NPS calculated?
Ask one question, sort the answers into three groups, subtract one percentage from another.
The question is how likely somebody is to recommend the product to somebody else, answered on a scale of zero to ten.
| Answer | Group | Counts as |
|---|---|---|
| 9 or 10 | Promoters | Positive |
| 7 or 8 | Passives | Nothing at all |
| 0 to 6 | Detractors | Negative |
The score is the percentage of promoters minus the percentage of detractors. Passives are counted in the denominator and then discarded, which is the first thing worth noticing about the arithmetic.
A hundred people, forty scoring nine or ten, forty scoring seven or eight, twenty scoring six or below, gives forty per cent minus twenty per cent, which is a score of twenty.
Two consequences follow from that shape and neither is obvious.
Two very different populations produce the same score. Forty promoters and twenty detractors scores twenty. So does eighty promoters and sixty detractors. The second company has a much louder problem and the number hides it completely.
A seven is thrown away. Someone who answered seven is a person who took the time to answer, is broadly content, and is counted as if they had never replied. Almost half of a typical response set can end up in that bucket.
What is a good NPS score?
There is no threshold that means anything on its own, and any page telling you otherwise is filling space.
The number moves with the question's wording, the moment you asked, the channel you asked through, the country the person is in and whether the people who answered are representative of the people who did not. Two companies running the survey differently are not comparable, and the same company changing when it asks is not comparable with itself.
What the number is genuinely good for is one thing: watching your own score move, with everything about how you collect it held still. A score that fell eleven points this quarter is information. A score of thirty two is a number.
What are the four things it cannot tell you?
Four, and each of them is a decision somebody tries to make with it anyway.
What to build. This is the big one. A score says people are unhappy. It does not say what about, and the free text box under the question produces sentences that have to be read one at a time and cannot be added up. A score cannot be sorted into a queue of work.
Who is unhappy. Unless you attach identity to the response, a score is an average over a population you cannot separate. The customers paying you the most and the customers paying you nothing are in the same figure, weighted equally.
Whether the people who answered are the ones who matter. Everybody who fills in a survey chose to fill in a survey. That is a filter, and it selects for people with time and strong feelings. The shape of that problem is worth reading before trusting any voluntary response set.
Whether anything you did worked. A score taken after a change tells you the score. Attributing the movement to the change requires that nothing else moved, which is never true.
What answers those questions instead?
A list of specific requests with a count on each, which is a different instrument rather than a better one.
A score compresses many opinions into one figure and loses the detail. A ranked list of requests keeps the detail and loses the summary. You can read the top of a list and know what to build next week. You cannot read a score and know that.
The two are complements. If you want one sentence on the difference: a score tells you that something is wrong, and a list tells you what.
On a board, the equivalent of the NPS number is the count on a card, and the equivalent of the free text box is the comment thread underneath it, which stays attached to the thing it is about. Every vote has a person behind it, so the count can be split by who cast it rather than averaged over everybody.
And where a score weights every respondent equally by construction, a vote can carry what the customer pays, which is the difference between knowing how many people are unhappy and knowing how much of your revenue is. Thirty two boards that keep that list, priced on their own pages and grouped by what each one bills on, is where to compare the instrument itself.
Do we measure NPS?
No. There is no survey, no score, no scale of zero to ten and none is planned.
What this product does is collect requests, count the people behind each one, optionally weight each vote by what that person pays, and report back to the people who asked when something ships. That is the whole of it.
If a score is what you need, buy a survey tool, and it will do that job better than any voting board. If what you need is to know which thing to build first, a score will not get you there whatever it costs.
How to collect customer feedback, and what each method quietly loses, covers surveys alongside the other five ways, with the same honesty about each.