Guides  /  customer measurement

Customer satisfaction and NPS

Net Promoter Score is quoted to one decimal place, tracked monthly and used to set bonuses. It is also one of the noisiest measures in common use, roughly half again as variable as the plain percentage it is built from. This guide works an NPS out with its confidence interval, shows how large a change you can actually detect, and sets out when CSAT or CES tells you more.

Saad Tariq Written and reviewed by Saad Tariq, Principal Statistician
Updated 4 November 202613 min read
How do you calculate NPS?

Ask how likely someone is to recommend you on a 0 to 10 scale. Scores of 9 and 10 are promoters, 7 and 8 are passives, and 0 to 6 are detractors. NPS is the percentage of promoters minus the percentage of detractors, reported as a whole number between minus 100 and plus 100. The calculation is trivial. The part almost everyone skips is the confidence interval, and without it a tracked NPS is mostly measuring its own sampling error.

Definitions

The three measures, and what each asks

Three metrics dominate customer measurement. They are not interchangeable, and the most common mistake is using one to answer a question another was designed for.

MeasureThe questionScaleAnswers
NPSHow likely are you to recommend us?0–10, collapsed to three groupsOverall relationship, loosely
CSATHow satisfied were you with this?1–5 or 1–7, reported as % satisfiedA specific interaction, just now
CESHow much effort did this take you?1–7Friction in a process

NPS asks about intention, not behaviour, and intention to recommend is a weaker predictor of actually recommending than its popularity suggests. CSAT is tied to a moment and decays fast, so asking two weeks later measures memory. CES is the narrowest and often the most actionable, because reducing effort is something a team can be asked to do.

The grouping is the unusual part

NPS is the only one of the three that throws away the scale it collected on. An 11-point response is compressed into three categories, two of which are then subtracted from one another. That compression is where both its simplicity and its statistical trouble come from.

Worked example one

NPS with its interval

400 customers respond. 180 score 9 or 10, 140 score 7 or 8, and 80 score 0 to 6.

Step 1. The three proportions

promoters = 180 / 400 = 45%
passives = 140 / 400 = 35%
detractors = 80 / 400 = 20%

Step 2. The score

NPS = 45 − 20 = +25

Reported as a number, not a percentage, which is a convention worth keeping because it reminds everyone it is a difference rather than a share.

Step 3. The standard error

NPS is a difference between two proportions drawn from the same sample, so its variance has a specific form:

Var = [ pₚ + pₔ − (pₚ − pₔ)² ] / n
   = [ 0.45 + 0.20 − 0.25² ] / 400
   = 0.5875 / 400 = 0.001469
SE = √0.001469 = 0.0383, so 3.83 points

Step 4. The interval

95% CI = 25 ± 1.96 × 3.83 = +17.5 to +32.5

A span of 15 points. The organisation reporting "our NPS is 25" has a number that could reasonably be 18 or 32.

Sample sizeNPS95% CIWidth
100+25+10.0 to +40.030 points
400+25+17.5 to +32.515 points
4,000+25+22.6 to +27.45 points

Identical score, three very different levels of knowledge. A monthly NPS from 100 responses is not a measurement, it is a wide guess that happens to be expressed as a whole number.

The structural problem

Why NPS is noisier than a percentage

Compare like with like on the same 400 responses.

SE of the promoter percentage alone = √(0.45 × 0.55 / 400) = 2.49 points
SE of NPS = 3.83 points

NPS is 1.54 times noisier than simply reporting the promoter share, on exactly the same data. There are two reasons, and neither is a flaw in how you collected it.

✓It subtracts two quantities that both vary. Every difference carries the uncertainty of both its parts.
✓It discards the passives entirely. 35% of your respondents, 140 people here, contribute nothing to the number. You paid to collect their answers and then deleted them.

This is why a tracked NPS bounces. Teams interpret the bouncing as customer sentiment responding to their work, when a large part of it is arithmetic.

Worked example two

How big a change can you actually detect?

The question that should be asked before any NPS target is set, and almost never is.

Detecting a change needs the uncertainty of two waves, not one, so the standard error of the difference is larger again by a factor of about 1.41.

Responses per waveSmallest NPS change you can reliably detect
40015.2 points
1,0009.6 points
2,5006.1 points
5,0004.3 points
10,0003.0 points

Read the first row carefully. With 400 responses a wave, which is a perfectly respectable survey, nothing under about 15 points is distinguishable from noise. An organisation celebrating a rise from 25 to 31 has observed nothing.

To track NPS monthly at a useful resolution you need thousands of responses every month. Most organisations running monthly NPS have a few hundred, which means the chart on the wall is substantially a picture of sampling variation.

The practical consequence for targets

If a bonus depends on moving NPS by five points and the survey can only detect fifteen, the bonus is a lottery. Work out the detectable change before the target is agreed, and set the target above it. This is a conversation to have with the board once, early, rather than every quarter afterwards.

Worked example three

Breaking NPS down by segment

The first thing anyone does with an NPS is split it. By customer size, by tenure, by region, by product. This is where the noise problem stops being theoretical, because every split divides the sample.

The same 400 responses, cut five ways.

SegmentnNPS95% CIWidth
Enterprise50+44+24.7 to +63.339 points
Mid-market110+29+15.0 to +43.228 points
SMB160+25+13.3 to +36.723 points
New, under 6 months47+9−15.0 to +32.047 points
Legacy33+6−20.4 to +32.553 points
Overall400+25+17.5 to +32.515 points

The story the table appears to tell

Enterprise customers love us at +44. Legacy customers are unhappy at +6. That is a 38 point gap and someone is about to build a retention programme around it.

The story the intervals tell

Enterprise runs from +24.7 to +63.3. Legacy runs from −20.4 to +32.5. Those ranges overlap. On this data you cannot say with confidence that Enterprise customers are more positive than Legacy ones at all.

The apparent 38 point gap is built on 50 responses against 33. Neither number is a measurement; both are wide guesses that happen to differ.

Why the smallest segments always look most extreme

Look at the widths. Legacy, with 33 responses, has an interval 53 points wide, more than three times the overall figure. Small samples produce extreme values by chance, so when you rank segments the smallest ones reliably appear at the top and the bottom. That is a property of the arithmetic, not of your customers.

It is the same mechanism that puts tiny departments at both ends of an engagement league table, and the same reason mixed effects models shrink small-group estimates toward the average.

✓Set a minimum segment size before you cut the data, and suppress anything below it. A hundred responses is a reasonable floor for NPS given how noisy it is.
✓Show every segment with its interval. A chart of overlapping bars tells the truth; a ranked table does not.
✓If you are comparing all segments against each other, you are running multiple comparisons. Five segments is ten pairs.
✓Report the mean of the raw 0 to 10 by segment as well. It uses every respondent and its intervals are narrower, so differences that are real show up sooner.
A rule that saves a lot of argument

Decide the segmentation and the minimum cell size before the data arrives, and write both into the analysis plan. Choosing how to cut after seeing the numbers is how a business ends up investing in a segment difference that was never there.

Information loss

What NPS throws away

Three specific distortions follow from compressing eleven points into three groups.

Customer movementReal changeEffect on NPS
6 → 8Meaningful improvementNone. Detractor becomes passive; the score rises only through the detractor share.
7 → 8Small improvementNone at all. Both are passives.
8 → 9Small improvementSame impact as a customer moving 0 → 9.
0 → 6Enormous improvementNone. Still a detractor.

A programme that moves a lot of customers from 5 to 8 has done real work and will show almost nothing. A programme that nudges a handful from 8 to 9 will show a clear rise. The metric rewards movement across two arbitrary boundaries and ignores everything else.

The mean of the 0 to 10 scores uses all the information and is far more sensitive to genuine change. It is less memorable, which is the only serious argument against it.

Choosing

When CSAT or CES tells you more

You want to knowUseWhy
Whether a specific interaction went wellCSATTied to the event, asked immediately, easy to act on
Where a process is painfulCESPoints at friction, and effort is something a team can reduce
Overall relationship, tracked slowlyNPS, annuallyIts noise is tolerable over long intervals with large samples
Whether sentiment actually movedMean of the 0–10Uses the whole scale, far more sensitive than the collapsed score

A reasonable arrangement for most organisations: CSAT or CES at the point of interaction for operational feedback, and NPS once or twice a year with a large enough sample to mean something. Monthly NPS on small samples combines the worst of both.

Whichever you use, the same self-selection problem applies. Customers who respond to a survey are not a random sample of customers, and the ones who feel strongly answer first.

Doing it

In R, SPSS and Stata

R

No package needed. The variance formula is three lines and the transparency is worth more than a dependency.

nps_ci <- function(prom, pass, det) {
  n <- prom + pass + det
  pp <- prom/n; pd <- det/n
  nps <- (pp - pd) * 100
  se <- sqrt((pp + pd - (pp - pd)^2)/n) * 100
  c(nps = nps, lower = nps - 1.96*se, upper = nps + 1.96*se)
}
nps_ci(180, 140, 80)

For the mean of the raw 0 to 10 instead, t.test(score) gives the interval directly, and it will be narrower than the NPS one on the same data.

SPSS

Create the grouping, then tabulate.

RECODE score (9 THRU 10 = 1) (7 THRU 8 = 2) (0 THRU 6 = 3) INTO npsgroup.
VARIABLE LABELS npsgroup 'NPS category'.
VALUE LABELS npsgroup 1 'Promoter' 2 'Passive' 3 'Detractor'.
FREQUENCIES VARIABLES = npsgroup.

SPSS will not produce an NPS interval. Take the promoter and detractor percentages from the frequency table and apply the formula above, or use the bootstrap module on a computed NPS variable.

Stata

recode score (9/10 = 1 "Promoter") (7/8 = 2 "Passive") (0/6 = 3 "Detractor"), gen(npsgrp)
tab npsgrp
gen nps_i = 100 * (score >= 9) - 100 * (score <= 6)
ci means nps_i

That last trick is the neat one. Scoring each respondent as +100, 0 or −100 makes NPS an ordinary mean, so ci means returns the correct interval directly and every standard command works on it.

Check the scale direction before anything else

Some platforms collect 1 to 10 rather than 0 to 10, and some reverse the scale so 0 is best. Applying the standard 9–10 and 0–6 cut-points to a reversed or shifted scale produces a plausible-looking number that is completely wrong. Tabulate the raw responses before recoding.

Avoid these

Six mistakes that cost you credibility

MistakeWhy it matters
Reporting NPS without an intervalAt 400 responses an NPS of 25 spans 18 to 33. The bare number implies a precision it does not have.
Tracking it monthly on small samplesWith 400 a wave, nothing under 15 points is detectable. The monthly chart is largely sampling noise.
Setting targets below the detectable changeA five point target on a survey that resolves to fifteen makes the bonus a lottery.
Comparing against published benchmarksDifferent wording, different industries, different response rates, and no interval on their figure either.
Treating it as a measure of behaviourIt records stated intention to recommend, which is a weaker predictor of recommending than its reputation suggests.
Ignoring who did not respondUnhappy customers often leave rather than answer. A rising NPS can mean detractors churned.
Is this quarter's movement a real change?

NPS has a wider interval than almost anyone reports, so a three-point move on a few hundred responses is usually noise presented as a trend. Send the responses and you get the score with its interval, and a straight answer on whether it moved.

See survey and insight analysis
Saad TariqSaad TariqPrincipal Statistician

Writing it up

How to report it

✓The score with a confidence interval and the number of responses behind it.
✓The three group percentages, so a reader can see whether a move came from gaining promoters or losing detractors. Those have different causes.
✓The response rate with its denominator, and who was excluded from the invitation.
✓The smallest change this sample can detect, stated once, early.
✓The mean of the raw scale alongside it, which will move earlier and more honestly than the collapsed score.

The sentence worth drafting carefully: "NPS was +25 (95% CI +18 to +33) from 400 responses, a 31% response rate. At this sample size changes smaller than about 15 points cannot be distinguished from sampling variation."

Common questions

Frequently asked questions

How do you calculate NPS?

Subtract the percentage of detractors, who score 0 to 6, from the percentage of promoters, who score 9 or 10. Passives scoring 7 or 8 are excluded. The result runs from minus 100 to plus 100 and is reported as a whole number rather than a percentage.

What is the confidence interval for NPS?

Its variance is [p promoters + p detractors minus the squared difference between them] divided by n. With 180 promoters, 140 passives and 80 detractors from 400 responses, NPS is +25 with a standard error of 3.83 and a 95% interval of +17.5 to +32.5.

Why is NPS noisier than a normal percentage?

Because it subtracts two proportions that each vary, and because it discards the passives entirely. On the same 400 responses the promoter percentage has a standard error of 2.49 points while NPS has 3.83, so NPS is about 1.5 times noisier on identical data.

How many responses do I need to track NPS?

More than most organisations collect. At 400 responses per wave the smallest reliably detectable change is around 15 points. Detecting a 5 point change needs roughly 5,000 per wave, and a 3 point change needs about 10,000. Monthly NPS on a few hundred responses is largely measuring its own sampling error.

What is a good NPS?

There is no universal answer, and published benchmarks are weaker than they look because question wording, industry, channel and response rate all differ and none of them comes with an interval. Your own trend, measured consistently with a large enough sample, is more informative than any comparison.

Should I use NPS, CSAT or CES?

CSAT for a specific interaction, asked immediately. CES where you want to find friction in a process, because effort is something a team can be asked to reduce. NPS for an overall relationship tracked slowly with a large sample. For detecting genuine movement, the mean of the raw 0 to 10 beats all three.

Can a rising NPS be bad news?

Yes. If unhappy customers churn rather than respond, they leave the sample and the score rises while the business worsens. Always read NPS alongside retention and the response rate, not on its own.