Net Promoter Score: Calculate NPS and Turn It into Action
An NPS of 34 can describe two very different customer populations. In one illustrative survey of 100 respondents, 52 are Promoters, 30 are Passives, and 18 are Detractors. In another, 67 are Promoters and 33 are Detractors, with no Passives at all. Both produce the same Net Promoter Score. One population leans positive; the other is sharply divided.
That is the central bargain in Net Promoter Score. NPS compresses a set of recommendation ratings into one number that leaders can track. The compression makes the result easy to communicate, but it also removes information. Used well, NPS tells a business that customer advocacy may be shifting and points to the customers, experiences, and segments worth investigating. Used alone, it cannot explain the shift, prove loyalty, or forecast revenue with certainty.
The practical question is therefore larger than “What is our NPS?” A useful program must establish what was measured, calculate the score consistently, show what the aggregate hides, and connect the responses to evidence about customer experience and behavior.
NPS turns a recommendation rating into an aggregate signal
Net Promoter Score begins with a question about recommendation intent, usually phrased as: “How likely are you to recommend us to a friend or colleague?” Respondents answer on a 0-to-10 scale. The Net Promoter System methodology groups a 9 or 10 as a Promoter, a 7 or 8 as a Passive, and a 0 through 6 as a Detractor.
Those labels are classifications applied to survey responses. A person who selects 6 has reported lower recommendation intent under this system; the rating does not, by itself, establish that the person will churn or criticize the company. Likewise, a 10 is stated intent, not a recorded referral. The labels become more useful when the company tests whether its own Promoters, Passives, and Detractors later behave differently.
Calculate the score from all valid responses
The formula is: NPS = percentage of Promoters − percentage of Detractors.
Suppose 200 customers submit valid ratings:
- 110 give a 9 or 10, so Promoters are 55% of respondents.
- 50 give a 7 or 8, so Passives are 25%.
- 40 give a 0 through 6, so Detractors are 20%.
The NPS is 55 − 20 = 35. Report it as NPS 35, not 35%. The inputs are percentages, but the result is a score ranging from −100, when every valid response is a Detractor, to 100, when every valid response is a Promoter.
Passives do not appear as a term in the subtraction, yet they still affect the score. All 200 valid ratings form the denominator for the Promoter and Detractor percentages. Removing the 50 Passives would make the remaining groups 73.3% Promoters and 26.7% Detractors, producing 46.6 instead of 35. That would be a different calculation.
A spreadsheet or survey platform should retain the raw 0-to-10 rating even when the dashboard displays only the three categories. Raw ratings make it possible to audit the grouping, correct a calculation, and see whether movement occurred within a category. A shift from many 0s to many 6s is invisible in NPS even though the underlying responses improved substantially.
Use the count of valid ratings for the denominator and state what made a response valid. A blank rating is not a Passive. A duplicated response, an ineligible account, or a test submission needs a documented treatment that is applied in every reporting period. Otherwise, a data-cleaning change can masquerade as a customer change.
The same score can hide opposite distributions
Return to the two illustrative samples with NPS 34. The first contains 52% Promoters, 30% Passives, and 18% Detractors. The second contains 67% Promoters, no Passives, and 33% Detractors. A leader who sees only 34 cannot tell whether the customer base is broadly favorable or polarized between strong advocates and strong critics.
That lost detail matters operationally. The first distribution suggests a sizeable middle group whose reasons may reveal what prevents a stronger recommendation. The second has almost twice the Detractor share, so recurring failures among those customers could deserve more immediate attention even though the headline score is identical. Neither interpretation proves a cause, but each sets a different investigation.
Always put four figures beside the score: the valid response count, the Promoter percentage, the Passive percentage, and the Detractor percentage. A trend chart becomes more useful when it also shows the survey population and field dates. If the score rises because the Detractor share fell, the business learned something different than if Detractors held steady while Passives became Promoters.
This is also why averaging already calculated NPS values can mislead. Imagine one segment with 20 responses and NPS 80 and another with 1,000 responses and NPS 10. Their simple average is 45, which gives the tiny segment the same influence as the large one. For a combined score, pool the underlying Promoter, Passive, and Detractor counts from populations measured under the same rules, then recalculate. Keep the segment scores visible because the aggregate can conceal the place where the experience actually changed.
A recommendation question can refer to a company, product, service relationship, or recent interaction. Those objects are not interchangeable. A buyer might recommend a software platform overall while rating a recent support case poorly; another might appreciate a helpful account manager but hesitate to recommend the product because implementation remains difficult.
Bain distinguishes competitive benchmark, relationship, and experience NPS. Competitive benchmark research asks a market sample about multiple providers under a comparable design. Relationship NPS asks known customers about the broader relationship without a specific triggering event. Experience NPS follows an action such as a purchase or service call. Identical arithmetic does not make scores from these programs comparable.
The distinction becomes especially consequential in B2B markets. One end user, an executive sponsor, a procurement lead, and an administrator can experience the same vendor differently. The respondent’s role therefore belongs in the interpretation. Calling one contact’s rating “the account’s NPS” silently turns an individual answer into an organizational view that was never measured.
Write the measurement contract before reading the trend
A stable NPS series depends on a stable set of choices. Document them before launch, and keep them with the results:
-
Name the object. State whether the question covers the company, a named product, the overall account relationship, onboarding, support, or another completed experience. The open-text follow-up should refer to the same object.
-
Define eligibility. Specify which customers and contacts can receive the survey, their lifecycle stage, required product exposure, active or churned status, and any exclusions. In a B2B account, decide which roles can speak to the question and whether multiple contacts per account are allowed.
-
Set the trigger and timing. A relationship pulse might use a recurring window, whereas an experience survey follows a defined event. The delay after that event changes what the respondent is likely to remember and rate.
-
Freeze the question and scale. Preserve the wording, the 0-to-10 direction, endpoint labels, language treatment, and display order. A shorter label or reversed mobile layout may look cosmetic but can change how people answer.
-
Define valid data and the calculation. Record how partial, duplicate, test, and late responses are handled; which date assigns a response to a period; how percentages are rounded; and whether the displayed score is rounded only after subtraction.
-
Choose reporting slices in advance. Product, plan, tenure, region, channel, journey stage, and contact role can all be useful. Choose slices tied to plausible decisions, then show their response counts so a dramatic movement in a tiny group is not mistaken for a stable pattern.
-
Assign the follow-through. Decide who can read open text, contact a respondent when appropriate, investigate a recurring issue, and approve a broader change. Collecting criticism without a path to examine it produces a dashboard, not a feedback system.
If the population, trigger, wording, scale, or calculation changes, mark a break in the series rather than presenting the movement as customer improvement or decline.
The contract also prevents an attractive but invalid comparison. A post-support experience score will overrepresent customers who recently needed help. A relationship survey sent to all active customers observes a different population at a different moment. Comparing the two as if one team “beat” the other says more about survey design than performance.
Pair the rating with one concise, open-ended question asking for the main reason behind it. The comment supplies a clue about price, reliability, support, usability, outcomes, or expectations, but it is still one respondent’s account. Preserve the event, product, segment, role, and date alongside the answer; without that context, comments are difficult to interpret and easy to cherry-pick.
Read a change as a measurement result before calling it a business result
When NPS moves from 28 to 34, the arithmetic says the Promoter share minus the Detractor share increased by six points among the valid respondents under the survey’s rules. It does not yet say that the customer base became six points more loyal, that an initiative caused the rise, or that revenue will follow.
First ask whether the surveys observed comparable populations. Customer mix may have changed because a product grew in a region with historically different ratings, a renewal campaign reached more executive sponsors, or dissatisfied customers left before the next survey. A rising score after unhappy customers churn is mathematically possible. Cohort and segment views can expose that composition effect.
Then inspect when the responses arrived. A service incident near the start of one field period and a product launch near the end of another can affect the comparison. Neither event invalidates the survey, but the dates and exposure should remain visible before someone assigns the movement to a quarterly program.
Treat precision, response bias, and benchmarks as separate questions
Response count governs precision, while respondent selection governs representativeness. They are related but different problems. A larger sample can narrow statistical uncertainty around the responding population and still be biased if important customer groups rarely answer.
NPS is a difference between two proportions drawn from three response categories. A 2026 methodological study of NPS confidence intervals showed that interval performance depends on both sample size and the Promoter-Passive-Detractor distribution; common normal approximations can be too narrow in small samples or when a category is near zero. Report a confidence interval when the sampling design supports one, and do not use a universal response-count threshold as a substitute for the precision required by the decision.
For a directional company trend, a wide interval may still be honest and useful. For a decision to reallocate budget between two small segments, the same uncertainty may be too large. Plan the sample around the smallest segment that must support a decision, not merely around the company total. If the analysis repeatedly creates segments with only a handful of responses, collect more data over a stable window or leave the result descriptive.
Response rate answers another question: how many eligible or invited people responded under a specified definition. The American Association for Public Opinion Research cautions that response rate alone does not reliably distinguish accurate from inaccurate survey estimates and recommends looking for evidence of nonresponse bias. A 60% response rate is not automatically representative; a 15% rate is not automatically unusable.
Compare respondents with nonrespondents using attributes already available and appropriate to the analysis: plan, tenure, account size, product use, region, support history, renewal status, or contact role. Large differences do not tell you exactly how nonrespondents would have scored, but they reveal where the result may lean. Report invitations, valid responses, the response-rate definition, and the material differences you found.
A “good” NPS adds a third question: good relative to what? An internal trend is defensible only when the measurement contract and customer mix are sufficiently stable. A competitive benchmark is useful only when it covers a comparable market, object, population, timing, mode, wording, and calculation. Bain’s benchmark guidance emphasizes comparison with competitors and distinguishes overall results from journey-level views; a self-selected customer survey and a double-blind market study do not create equivalent scores simply because both use NPS.
There is no context-free cutoff at which an NPS becomes good. A positive score means Promoters outnumber Detractors among respondents. That may be encouraging, but it says nothing about competitive position, business economics, or the experience of nonrespondents. Prefer a relevant competitor set or a stable internal baseline to an industry number whose sampling method is unknown.
Use NPS to decide where to investigate
NPS becomes actionable when it changes the next question. A low score for recently onboarded administrators can direct attention to implementation records and enablement gaps. A Detractor cluster among executive sponsors approaching renewal can trigger account-specific review. A company-wide rise accompanied by worsening scores in a fast-growing product line warns that the aggregate is hiding a future problem.
The score itself does not rank fixes. Response volume, customer impact, commercial exposure, recurrence, and the team’s ability to change the underlying experience all matter. A frequently mentioned inconvenience may be less consequential than a rare failure that blocks deployment for a strategic customer.
Choose NPS, CSAT, or CES by the decision you need to make
NPS is often grouped with Customer Satisfaction Score and Customer Effort Score, but the three questions observe different things. They can disagree without any measurement being wrong.
| Metric | What the respondent judges | Useful starting decision | What the score does not establish |
|---|---|---|---|
| NPS | Likelihood of recommending a company, product, or service | Whether relationship-level advocacy is strengthening for a defined population | The cause of the rating, an actual referral, or future revenue |
| CSAT | Satisfaction with a named product, service, event, or interaction | Whether a specific experience met expectations | The ease of the process or health of the whole relationship |
| CES | Effort or ease in completing a named task or interaction | Where a workflow creates friction | Satisfaction with the outcome or willingness to recommend the company |
Under a common five-point implementation, CSAT is the percentage of valid responses rated 4 or 5. Other programs may report an average, so the label alone does not guarantee comparable math. CES implementations also vary in wording, scale, direction, and aggregation; one common method averages the response values.
A customer can recommend the vendor, feel satisfied that a support issue was resolved, and still report that resolution required too much effort. High NPS, high CSAT, and poor CES would form a coherent picture: the broader relationship remains strong, the outcome met expectations, and the process contains friction. Averaging the three into one customer-health number would erase that diagnosis.
Use NPS when the decision concerns advocacy across a relationship, product, or market. Use CSAT when a team owns a defined outcome and needs to know whether it satisfied customers. Use CES when the team can remove steps, repetition, waiting, transfers, or ambiguity from a task. If the real question is why customers churned, none of the three is sufficient without behavioral and account evidence.
Follow the score into causes, actions, and observed outcomes
A disciplined NPS cycle can remain compact:
-
Locate the movement. Decompose the change into Promoter, Passive, and Detractor shares, then inspect preselected segments and cohorts. Look for both large groups and small groups with material business exposure.
-
Read the reasons in context. Review open-text themes with the respondent’s product, role, journey stage, and triggering experience. Keep a path back to the original comment so a theme label does not replace what the customer actually said.
-
Test the clue against operations. If respondents mention slow resolution, examine wait time, transfers, reopen rates, and case histories. If they mention poor value, inspect product use, promised outcomes, pricing changes, and renewal conversations. Agreement strengthens a hypothesis; disagreement tells you where to look again.
-
Choose a change an owner can make. Define the affected population, the experience to change, the expected operational result, and the customer outcome that should follow. “Raise NPS” is not a workable intervention. Reducing repeated identity checks during support is.
-
Watch the leading and lagging results. Confirm that the operational change occurred, then observe relevant survey responses and later behavior such as adoption, renewal, expansion, contraction, referrals, or churn. Keep the time sequence visible and avoid claiming causation from a simultaneous movement alone.
Close the loop with individual customers when the survey terms, consent, context, and issue make contact appropriate. A Detractor comment describing an unresolved outage needs a different response from an anonymous criticism of pricing. The first may support prompt service recovery; the second belongs in a pattern analysis. Promoter comments also deserve attention because they identify value the company must avoid damaging during a redesign.
For recurring themes, use both frequency and consequence. Ten complaints about a confusing label and two reports of data loss are not comparable merely by count. Link themes to affected workflows, account exposure, severity, and supporting operational data. The resulting priority may challenge the loudest narrative, which is exactly why the additional evidence matters.
Keep customer-level response data separate from broad performance reporting. Frontline teams may need identifiable feedback to recover a service failure; executives usually need trends, distributions, segment context, and material themes. Access rules should follow the purpose and the promises made to respondents.
NPS can accompany growth without proving it
Recommendation intent has a plausible relationship with referrals, retention, and spending, but a plausible chain is not a universal law. Customer behavior also reflects price, switching costs, contracts, procurement rules, product availability, competitive alternatives, and market conditions. In B2B, a user may gladly recommend a tool while having no authority over renewal.
The empirical record is mixed. A peer-reviewed study in the Journal of the Academy of Marketing Science found that earlier direct studies did not confirm the claim that NPS was superior to other customer-mindset metrics for predicting sales growth. Its own analysis covered 193,220 evaluations across seven US sportswear brands over five years and found predictive value under specific modeling and sampling conditions, particularly for changes in a brand-health form of NPS. Those conditions limit how far the result can travel.
More recent evidence preserves the same caution. A 2026 longitudinal study of 38 ski resorts, using more than 120,000 survey responses and 136 destination-year observations, found that NPS predicted skier visits but was not superior to customer satisfaction; the respondents’ destination experience also affected its predictive ability. This is evidence from one industry and design, not a benchmark for every subscription, manufacturing, or professional-services business.
The business response is to validate NPS locally. Link survey cohorts to later outcomes over a defensible time window. Compare the behavior of categories while controlling, as far as the data permits, for tenure, product, market, account size, and other factors that affect both the rating and the outcome. Check whether changes in NPS add useful information beyond existing measures. If they do, the score has earned a role as a leading signal in that setting. If they do not, keep the customer comments and use a measure closer to the decision.
Do not set bonuses directly from an untested headline score. Strong incentives invite survey selection, coaching, suppression of unhappy customers, and pressure on respondents. A metric used for learning can lose that value when people are rewarded for moving the number by any available route.
Start with the decision the score must improve
Before launching another NPS survey, write one sentence naming the decision the result will inform. Then specify the population, object, trigger, and follow-through that make that decision possible. If those details are unclear, the organization is not ready to interpret a trend, however polished the dashboard looks.
The score earns its place when it helps a team find a consequential customer pattern, test the reason, change an experience, and observe what follows. That is more demanding than calculating one number. It is also where the value begins.
Frequently asked questions
Can an individual customer have an NPS?
An individual supplies a 0-to-10 rating and may be classified as a Promoter, Passive, or Detractor; NPS is the aggregate calculated from a defined set of valid responses. In B2B reporting, distinguish the respondent from the account as well: one contact’s rating should not be presented as an account-wide score unless the research design explicitly defines and supports that aggregation.
How many NPS responses are enough?
There is no universal minimum. Set the sample from the precision required for the company total and each segment that will drive a decision, using an expected Promoter-Passive-Detractor distribution. The 2026 confidence-interval study notes that approximate normality depends on the category counts and gives at least five expected Promoters and five expected Detractors as a basic condition, with ten in each being better; that is a condition for an approximation, not a guarantee that the sample is adequate for a business decision.
What is a good NPS survey response rate?
No single percentage establishes survey quality. Report the eligible population, invitations, valid responses, field dates, channel, and exact rate definition, then compare respondents with nonrespondents on available customer attributes. AAPOR’s response-rate guidance says rates do not reliably separate accurate from inaccurate surveys, so evidence of nonresponse bias, missing data, and consistency with other observations matters more than clearing an arbitrary threshold.
How often should a company run an NPS survey?
Match frequency to the type of NPS and the speed of the decision. Bain’s guidance on relationship and experience NPS says relationship NPS is often requested once or twice a year, while experience NPS follows selected customer actions, and recommends suppression rules so a customer is solicited no more than once every three months. A B2B company with long implementations may need a slower relationship cadence, while a high-volume service can collect experience feedback continuously without surveying the same person after every interaction.
What does a negative NPS mean?
A negative NPS means the Detractor percentage exceeds the Promoter percentage among valid respondents: for example, 25% Promoters and 40% Detractors produce NPS −15. It does not mean that 15% of customers are unhappy, and it does not reveal why the balance is negative. Read the full distribution, respondent mix, comments, and operational evidence before choosing a response.