Net Promoter Score: How It Works, and Where It Fails

Three businesses, three problems, one identical score. This is why the number alone is not a diagnosis.

Nearly every page ranking for this term is published by a company that sells survey software. So the mechanics get explained accurately and the limits get skipped.

Both halves are here. NPS is a post-purchase check on how a relationship feels, which puts it near the end of the customer journey rather than anywhere in acquisition — and that placement explains most of what it can and cannot do.

What is a Net Promoter Score, exactly?

NPS is the percentage of your customers who answer 9 or 10 to a recommend question, minus the percentage who answer 0 to 6. Passives — the 7s and 8s — are counted in the denominator but contribute nothing to the score.

It comes from Frederick F. Reichheld's "The One Number You Need to Grow" in Harvard Business Review, December 2003, which argued that willingness to recommend correlated with growth better than satisfaction did. Bain & Company's methodology page (fetched 28 September 2026) gives the question as "How likely are you to recommend us to a friend or colleague?" on a zero-to-ten scale, with these bands.

The arithmetic is deliberately blunt. Out of 200 responses, 90 promoters and 30 detractors gives 45% − 15% = +30.

Because both terms are percentages of the same base, the floor is −100 (everybody a detractor) and the ceiling is +100. It is a score, not a percentage, which is why "our NPS is 30%" is always wrong.

One India note: the abbreviation "NPS" is searched enormously here, but most of that volume belongs to the National Pension System. Spell the term out once internally, and do not try to rank for the three letters.

Why two very different businesses can report the same NPS

This is the part most articles leave out, and it is the most useful thing on this page. Because the score is a subtraction, it throws away the shape of the distribution that produced it. Three businesses with genuinely different problems can all report +20.

Business B has four times the detractors of Business C behind an identical headline. Manage against the headline and you treat all three the same, when only one needs its complaints process rebuilt this quarter.

Fisher and Kordupleski make the same point in print: subtracting the two percentages "obscures underlying distributions", and the banding means a company "is basically not distinguishing those of its customers rating them 0 from those rating them 6" while "totally ignoring 'passives'" (Applied Stochastic Models in Business and Industry, vol. 35, 2019, pp. 138–151).

The fix is free. Report the three band percentages alongside the score, always.

What the research says about NPS and growth — and who owns the name

The mechanics are sound. The original claim — that this one number predicts growth better than the alternatives — did not replicate when independent researchers tested it.

Keiningham, Cooil, Andreassen and Aksoy published "A Longitudinal Examination of Net Promoter and Firm Revenue Growth" in the Journal of Marketing (vol. 71, issue 3, July 2007, pp. 39–51), using data from 21 firms and more than 15,500 interviews in the Norwegian Customer Satisfaction Barometer and comparing against the American Customer Satisfaction Index. Their stated result: in the industries Reichheld cites as exemplars, "the research fails to replicate his assertions regarding the 'clear superiority' of Net Promoter compared with other measures in those industries."

Fisher and Kordupleski's 2019 review is blunter — NPS "has not lived up to its claimed benefits" — and adds that in practice it is usually "purely observational, with little or no understanding of demographic factors let alone sampling biases".

That is narrower than "NPS is useless":

• What is contested: that NPS beats satisfaction and retention measures as a predictor of revenue growth.
• What is not contested: that a recommend question, asked consistently on the same base, tracks something real about how customers feel.
• What follows: use it as a trend on your own customers. Do not use it as a forecast, and do not use it as evidence in a board pack that growth is coming.

On the name: Bain's trademarks page (fetched 28 September 2026; the page carries no date of its own) states that Net Promoter®, NPS® and NPS Prism® are registered trademarks of Bain & Company, Inc., NICE Systems, Inc. and Fred Reichheld, and that Net Promoter Score℠ and Net Promoter System℠ are service marks. So the widespread line that "Net Promoter Score is a registered trademark" is not quite right on Bain's own account — the registered marks are Net Promoter and NPS.

Relationship NPS or transactional NPS: which one to run

Relationship NPS asks the whole customer base how the relationship feels, once or twice a year. Transactional NPS fires after a specific event — a delivery, a support call, an onboarding — and tells you about that event. They answer different questions and they are not interchangeable.

Bain's "Three Types of Net Promoter Scores" (fetched 28 September 2026) names relationship NPS, "solicited when there is no initiating trigger"; experience NPS, "triggered by the completion of specific customer actions"; and a double-blind competitive benchmark run by a third party. Bain states plainly that relationship and experience NPS "should not be compared directly", and recommends limiting surveys to once every three months per customer.

For most Indian D2C and services businesses under a few thousand customers, one relationship wave every six months plus transactional NPS on your single worst touchpoint is the whole programme. Quarterly waves on a small base produce movement you cannot read — see the maths below.

When you ask changes what you get

Timing is part of the measurement, and changing it invalidates your trend line. The same customers surveyed at three different moments can produce scores that differ by double digits, with none of the three wrong.

• Immediately after purchase measures the buying experience and the anticipation. The product has not been used yet. Scores here are systematically kinder.
• After first real use measures whether the thing works. This is usually the most decision-useful moment and the hardest to time.
• Immediately after a support ticket closes measures the resolution, not the company. Run it, but keep it in its own bucket.
• On a fixed calendar, months after purchase measures the relationship. Slower, steadier, and the only version worth trending over years.
• Never mid-problem. A survey that lands while a refund is pending measures the refund.

The rule: pick one moment per survey type and freeze it. Move the timing and you restart the trend line — say so in the report, because a change in when you asked looks exactly like a change in how customers feel.

Non-response is the other half. People who answer an unsolicited survey are not a random sample: the indifferent middle is least likely to reply, which is precisely the group NPS already discards. That is the sampling bias Fisher and Kordupleski flag, and more sample does not fix it.

The "why?" box is worth more than the score

The free-text follow-up — "what is the main reason for your score?" — is the only part of an NPS survey that tells you what to do on Monday. The number says something changed. The text says what changed.

Critics and vendors agree here: a single ordinal score carries no diagnostic information. Everything actionable comes from the sentence after it.

1. Ask one open question, unprompted. No checklist of reasons — a list teaches people what to say.
2. Read detractor comments first, in full. Twenty comments read properly beat two hundred auto-tagged by sentiment.
3. Tag to a fixable cause, not a feeling. "Delivery took 9 days" is a cause. "Bad experience" is not.
4. Count causes, not scores. Rank the tags by frequency and by how many detractors each one explains.
5. Close the loop on the worst ones. Reply to the person. A detractor who gets a reply is the cheapest retention work available.
6. Put the top three causes in the report above the score. The score is the headline nobody can act on.

The check we run before an NPS number goes in a client report (our own practice, with no measured figure attached to it): we ask for the three band percentages, the response count for that wave, the exact send timing, and whether the question wording or the trigger changed since the last wave. If any of the four is missing, the score goes in as a range with the caveat printed beside it, or it does not go in. We added that habit after decks where a "10-point improvement" turned out to be a survey that had started firing at a different moment. We publish no figure for how often that happens — we have not counted it.

How much sample an NPS number needs before it means anything

Small samples make NPS look far more volatile than reality. Because the score is a difference of two proportions, its margin of error is wider than a single percentage from the same sample — and almost nobody puts a confidence interval on it.

Worked calculation — our own arithmetic, formula shown, on hypothetical Business A above (40 promoters, 40 passives, 20 detractors per 100):

NPS   = p_promoter − p_detractor
Var   = [ p_promoter + p_detractor − (p_promoter − p_detractor)² ] / n
SE    = √Var
95% margin of error = 1.96 × SE  (expressed in NPS points)

With p_promoter = 0.40 and p_detractor = 0.20, the bracket is 0.40 + 0.20 − 0.20² = 0.56. So:

At 200 responses, a move from +20 to +28 is inside the noise. Most Indian SMB waves we are shown have fewer than 200 responses, so the quarter-on-quarter changes being presented are usually not changes.

Two consequences. Report the interval, or at minimum the response count, every time. And on a base that cannot produce a few hundred responses per wave, run NPS annually as a slow trend and spend the attention on the free text. The same discipline applies to every figure in a report — our performance marketing team runs media and measurement together for this reason, and it is cheaper to argue about sample size before a decision than after one.

Why an Indian benchmark and a US benchmark are not the same number

Comparing your Indian NPS against a published global or US benchmark is not like-for-like, and the honest move is to stop doing it. People in different markets use rating scales differently, so part of any cross-country gap is scale use rather than sentiment.

The evidence is in survey methodology, not NPS literature. Johnson, Kulesa, Cho and Shavitt's "The Relation Between Culture and Response Styles: Evidence From 19 Countries" (Journal of Cross-Cultural Psychology, vol. 36, issue 2, March 2005) — India is one of the 19 — found that "power distance and masculinity were found to be positively and independently associated with extreme response style", and that individualism, uncertainty avoidance, power distance and masculinity were each negatively associated with acquiescent responding.

Two things that study does not say, and we will not stretch it: it is about rating scales in general, not the NPS question, and it gives no Indian NPS figure. We are not claiming Indian scores run high or low — only that a cross-country comparison confounds sentiment with response style, so the gap cannot be interpreted. Bain points the same way: it says a double-blind third-party benchmark is how you learn where you stand, not a published table.

We publish no benchmark table here. Not for lack of numbers — the internet is full of them — but because we could not find a set where every figure had a named, dated source we could fetch and stand behind. Most circulating "industry average NPS" figures are one vendor's panel, undated, resampled by everyone else. So:

• Your own trend, same question, same timing, same base, is the only number to manage against.
• If you need a competitive read, commission a double-blind study or accept that you do not have one.
• Segment your own score before you envy anyone else's — first-time versus repeat, metro versus non-metro, prepaid versus COD. Method in our customer segmentation guide.

What NPS cannot tell you, and what to pair it with

NPS is one lagging, self-reported signal about sentiment. Not a growth forecast, not a revenue number, and no substitute for behaviour. The honest boundary, and the metric that covers each gap:

Deciding which of those belong on your actual dashboard is a separate exercise, and our marketing KPIs guide covers the selection logic. The wider measurement stack — attribution, tracking, reporting — sits in our marketing analytics guide.

Where to go from here

Run NPS if you will act on the comments. Skip it if you only want a number for a slide.

1. Pick one type and one moment — relationship or transactional — and write the timing down so the next wave matches.
2. Ask the score plus one open "why". Two fields. Nothing else.
3. Report three band percentages, the response count and the score, in that order.
4. Tag the detractor comments to fixable causes and fix the top one before the next wave.
5. Trend against yourself only, treating any move smaller than your margin of error as flat, and put a behavioural metric beside it so sentiment never travels alone.

Sentiment work is slow, the same way organic is: expect three to six months before a fixed cause shows up as a readable movement, and longer on a small base.