Net Promoter Score℠ · NPS® confidence and change
How much does your NPS result tell you?
See the response mix behind the score, estimate uncertainty, and compare two independent results. Add the change that matters to your organization to put the difference in context.
Start with the people who answered.
Use unweighted counts of valid 0–10 responses. The opening data are an example. Paired respondents, overlapping groups, weighted surveys, and repeated transactions need a different analysis.
Opening counts are illustrative. Replace them with your data. No responses are uploaded or saved by this calculator.
Loading calculator. Fixed example shown.
Composition and uncertainty
What does this score contain?
Hypothetical random-sampling benchmark. Your sampling model has not been validated; these intervals are not a population margin of error.
A · Result A
400 valid responses
- Detractors
- 20.0% (80)
- Passives
- 30.0% (120)
- Promoters
- 50.0% (200)
95% hypothetical model interval
+22.14 to +37.41
AW(3,T) adjusted-Wald method; interval centered at +29.78. The circle marks the raw score.
Interpretation and limits
Intervals assume independent, identically distributed unweighted -1/0/+1 responses from a large population. They exclude selection, coverage, nonresponse, and measurement bias.
This is a hypothetical random-sampling benchmark. These counts do not validate a model for your sample or establish population representativeness; do not report it as a population margin of error.
Methods and report text
Read the interval for the change itself.
Two results can have overlapping individual intervals while their difference is distinguishable from zero. The comparison here calculates uncertainty for B minus A directly. A positive value means the observed NPS was higher in B.
A business threshold is a separate decision. An observed increase of six points does not establish a gain greater than five points if the change interval still includes smaller gains or a decline. The calculator compares the full interval with your entered threshold in both directions.
Does an interval containing zero mean there was no change?
No. It means the analysis does not resolve the direction under the stated model. Look at the range of changes still compatible with the data. An interval that includes both a meaningful decline and a meaningful increase leaves the decision uncertain. This tool does not perform an equivalence test.
Can I compare the same customers before and after?
Matched responses need the variability of each person's change. Two sets of category totals lose that information. Keep respondent identifiers or a Detractor/Passive/Promoter transition table and use a paired analysis. Partially overlapping samples also need the relationship between responses. This release shows descriptive scores but withholds their change interval.
What if the sample or questionnaire changed?
The calculation cannot separate customer change from a change in recruitment, target population, survey channel, wording, question order, or weighting. Record those changes in the report and evaluate comparability before interpreting a trend. No interval corrects selection or nonresponse bias.
Can this tool compare a benchmark or plan the next wave?
It analyzes observed counts. A benchmark estimated from another survey has its own uncertainty and needs its counts and sampling details. A published score alone is insufficient. NPS-specific sample planning, fixed-reference tests, matched records, and multi-wave analysis are not included.
The detectable-difference planner supports independent percentages and means; its percentage mode is not an NPS calculation. Discuss next-wave NPS planning using the expected response mix and the decision threshold.
Sources reviewed September 29, 2026
Formulas, assumptions, and sources
Ratings 0–6 are Detractors (D), 7–8 are Passives (Q), and 9–10 are Promoters (P). Valid responses total n = D + Q + P. Missing, refused, or invalid responses are excluded. The observed score is 100 × (P − D) / n.
Single score: the triangular adjusted-Wald interval, AW(3,T), adds 0.75 to D, 1.5 to Q, and 0.75 to P. The interval uses those adjusted shares and n* = n + 3, while the headline retains the raw observed score. This follows the 2026 paper's implementation; AW(3,U) is another supported choice in its simulations, not an option in this release.
Adjusted interval = 100(pP* − pD*) ± z × √[10,000 × {pP* + pD* − (pP* − pD*)²} / (n* − 1)]
The normal critical value z matches the selected two-sided confidence level. Mathematical endpoints are never truncated; charts stop at the physical scale. At least two valid responses are required for a single-score interval. With D=1, Q=2, P=17, the raw score is +80 and the 95% adjusted interval is approximately +44.4 to +94.8.
Independent change: code D, Q, and P as −1, 0, and +1. Use the raw counts and a Welch t interval for the difference of their means, multiplied by 100. This is our implementation choice using the NPS variance and NIST's unequal-variance procedure; it is not a two-sample adjusted-Wald method.
v = 10,000 × {pP + pD − (pP − pD)²} / (n − 1)
SE(change) = √(vA + vB)
df = (vA + vB)² / {vA²/(nA − 1) + vB²/(nB − 1)}
Change interval = NPS(B) − NPS(A) ± t × SE(change)
For this release, the change interval requires at least 30 valid responses and at least 10 outside the largest category in each result. These conservative implementation checks limit sparse-data use; they do not guarantee nominal coverage. No continuity, finite-population, or multiple-testing correction is applied. Counts are limited to 1,000,000 valid responses per result.
In local simulations of two routine comparison scenarios, 95% Welch intervals covered the true change about 95% of the time. In a deliberately concentrated scenario, most samples failed the safeguards; among the 2,638 eligible comparisons out of 100,000, coverage was only 86.1%. Near the cutoff, a reported interval can still be too narrow. Obtain a sparse-data analysis before using a concentrated-sample result for a decision.
Intervals assume independent observations from a common response distribution within each sample. For opt-in, voluntary, invitation-only, or unknown sampling, outputs are labeled hypothetical model intervals. The calculator cannot validate the sampling model or establish population representativeness. A 95% procedure aims for 95% coverage over repeated samples under its assumptions; it does not assign a 95% probability to the fixed population value.
- Bain: standard NPS categories
- Turk, Cinderich & McNeill (2026): NPS confidence intervals
- Rocks (2016): Interval Estimation for the Net Promoter Score
- NIST: independent means and Welch–Satterthwaite inference
- AAPOR: survey methodology and precision disclosure
NPS® is a registered trademark and Net Promoter Score℠ is a service mark of Bain & Company, Inc., NICE Systems, Inc., and Fred Reichheld. Trademark information.