Net promoter score (NPS)

How the score is built

Every respondent lands in one of three bands, and the middle band is the one that surprises people — it is counted in the denominator but contributes nothing to the score:

• Promoters, 9 to 10. Enthusiastic enough to put their own reputation behind a recommendation. This is the group the metric is really trying to count.
• Passives, 7 to 8. Satisfied and entirely unmoved. They dilute your percentages by sitting in the base, which is why a product everyone mildly likes scores far lower than founders expect.
• Detractors, 0 to 6. Note how wide this band is: a 6 out of 10 is a fairly ordinary answer in everyday language and a full detractor here. The scale is deliberately harsh, and comparing it to a percentage-satisfied figure will mislead you.

The arithmetic is percentage of promoters minus percentage of detractors. Twenty replies made up of ten promoters, five passives and five detractors give 50% minus 25%, or 25. The score is conventionally written as a bare number rather than a percentage, because it is a difference between two percentages and not a proportion of anything.

What the number is good at, and what it cannot tell you

Its real strength is comparability over time inside one company. Asked the same way, of the same kind of customer, at the same point in their relationship with you, the trend is a genuine signal — and because it is one number, it survives being reported to people who will never read a research summary. That is a smaller virtue than it sounds and a real one.

What it cannot do is explain itself. A score that fell six points tells you nothing about whether a feature broke, a competitor got better, support got slower, or you simply started surveying a different segment. It is also easy to move without improving anything, by choosing when and whom to ask: survey people right after a successful action and the score rises, survey everyone including the accounts that quietly stopped logging in and it falls. Because the sampling choice is invisible in the final number, two companies quoting the same figure may have measured genuinely different things — the same reason a cross-company churn rate comparison rarely means much.

Early on, the number is noise and the comments are the asset

There is a specific arithmetic reason NPS misleads small companies, and it is worth doing once. Each respondent is worth 100 divided by your sample size in percentage points, and moving a single person from detractor to promoter shifts both percentages at once — so one changed mind moves the score by 200 divided by your sample size. With 20 responses that is 10 points; with 10 responses, 20 points. A founder watching their score swing from 30 to 10 is usually watching two people, not a trend, and reacting to it means reorganizing a roadmap around sampling noise.

First-party note from building AgentCeres — the AI Growth Officer at agentceres.com: we run a customer-feedback specialist whose job is to read reviews, support conversations and survey responses and rank the themes that keep recurring, and everything that has actually changed a decision came out of the free-text half rather than a score. That matches the arithmetic above. The practical version for anyone under a few hundred customers is to keep asking the rating question if you want the trend line later, treat the number as decoration until the sample can carry it, and put your attention on the box underneath it. A useful follow-up is not "why?" but "what would have made that a 10?", which asks for a specific change instead of a justification. See how to collect customer feedback for the wider loop this sits inside.

NPS and the metrics it gets confused with

CSAT — customer satisfaction — asks how satisfied you were with one specific interaction, usually right after it, and is reported as the percentage giving a positive rating. It answers "did that go well?" where NPS attempts "how do you feel about us overall?". CES, customer effort score, asks how much work something took, and predicts repeat behaviour better than either when the thing you are diagnosing is friction. None of the three replaces the others, and running all three on a small user base mostly produces survey fatigue.

The more useful contrast is against behaviour. Stated intent to recommend is a survey answer; referrals that actually arrive are an outcome, and the two diverge routinely because recommending something is effortful and rating it is not. The same applies underneath: what people do shows up in activation rate and retention long before it shows up in a satisfaction score, which is why an NPS reading that contradicts your usage data is usually the one that is wrong. Where a survey does earn its keep beyond diagnosis is social proof — a promoter who left a specific, quotable sentence is a testimonial you already have permission to ask about.