Net Promoter Score (NPS)· Customer-experience metric
Net Promoter Score (NPS) — Definition, Formula, and What Counts as Good (2026)
Quick answer: NPS asks one question ("How likely are you to recommend us to a friend or colleague?"), scores it 0-10, sorts the answers into promoters (9-10), passives (7-8), and detractors (0-6), and subtracts the detractor percentage from the promoter percentage. The result runs from -100 to +100 and measures loyalty to your company, not satisfaction with any single interaction. That scope is the detail most operators miss: CSAT grades the conversation that just ended, NPS grades the relationship, and swapping one in for the other produces numbers that look precise and mean the wrong thing. For an SMB the score comes with a second warning label — at small sample sizes a handful of respondents can swing it by double digits, so read the movement, not the point value.
What it is
NPS is a survey metric with an unusually rigid recipe. The question wording is fixed: "How likely are you to recommend us to a friend or colleague?" The scale is fixed: 0 (not at all likely) to 10 (extremely likely). And the buckets are fixed:
- Promoters (9-10) — customers enthusiastic enough to put their own reputation behind a recommendation.
- Passives (7-8) — satisfied but unenthusiastic; they count toward neither side.
- Detractors (0-6) — everyone else, including the 6s who thought they were giving you a decent grade.
The formula subtracts one share from the other:
NPS = % promoters − % detractors
Two properties fall out of that subtraction and both surprise first-time users. The score is a spread, not a percentage: an NPS of 40 does not mean 40% of anything, it means the promoter share exceeds the detractor share by forty points. And passives vanish from the arithmetic entirely, so a company where everyone answers 7 scores exactly zero — the same zero as a company splitting evenly between fanatics and furious.
Where the number came from
Fred Reichheld of Bain & Company introduced NPS in a 2003 Harvard Business Review article, "The One Number You Need to Grow," arguing that the recommend question outperformed longer satisfaction surveys at predicting revenue growth. The metric spread fast precisely because it was one question with a memorable output, and Bain built a full Net Promoter System of practices around it. The names are worth respecting in print: Net Promoter, NPS, and the related marks are registered trademarks held by Bain & Company, Satmetrix Systems (now part of NICE), and Reichheld.
The growth-prediction claim did not survive scrutiny unchallenged. Independent academic replications, most prominently in the Journal of Marketing in 2007, failed to reproduce NPS's superiority over other satisfaction metrics, and the measurement literature has treated the "one number" framing skeptically since. Reichheld himself published a recalibration in 2021 (Net Promoter 3.0), pairing the survey with an accounting-based "earned growth rate" partly because self-reported scores had become so easy to game. None of this makes the metric useless. It makes it a loyalty thermometer rather than a growth oracle, which is how careful teams use it.
NPS is not CSAT (and neither is sentiment)
Three measurements get conflated because all three claim to read how customers feel. They differ on what they grade, when they ask, and who answers:
| Metric | What it grades | When it is collected | How |
|---|---|---|---|
| CSAT | One interaction ("Was this conversation helpful?") | Immediately after the conversation | Survey, answered by a minority |
| NPS | The whole relationship ("Would you recommend us?") | On a slow cadence or after milestones | Survey, answered by a minority |
| Sentiment analysis | The live message ("Is this customer frustrated right now?") | Every message, as it arrives | Inferred by a model, asked of no one |
The practical rule: match the survey to the scope of what you want to learn. A customer-service chatbot that just resolved a shipping question should ask a CSAT-style "did this help," because that is the thing the customer can actually evaluate in the moment. Asking the recommend question there mostly measures the last five minutes wearing a relationship costume: a customer fresh out of a frustrating conversation will torch a company they otherwise like, and a smooth bot answer will flatter a company the customer is quietly outgrowing. Our chatbot NPS guide covers the timing rules that keep the two from contaminating each other.
What counts as a good score
The honest answer is that "good" is relative, and anyone quoting a universal number is selling something. Three readings hold up:
Above zero means more advocates than critics. That is the only threshold with intrinsic meaning, because it falls straight out of the subtraction. Everything above it is comparative territory where industries differ enormously — customers grade software, airlines, and insurance on very different curves, so a score that embarrasses one sector leads another.
Your competitors are the benchmark that matters. Bain's own guidance emphasizes relative NPS, your score against rivals in the same market, over absolute targets. An SMB without access to competitive survey data can substitute the next-best comparison: its own score last quarter.
Small samples make the score jumpy. Each respondent in a 50-response survey moves the score by two points on their own; one customer who shifts from detractor to promoter moves it four. A ten-point swing on that base is ordinary noise, not a trend. This is the single biggest reason SMBs misread NPS, and it is why the movement of a 300-response quarterly survey beats the precision theater of a weekly score built on twelve replies.
The gaming patterns are worth knowing because they are everywhere: survey timing chosen right after happy moments, "anything less than a 9 counts against me" coaching at the point of sale, and surveys sent only to segments already known to be warm. Each inflates the number while draining its meaning — the same trap as optimizing CSAT by escalating every hard conversation, which our CSAT improvement guide warns against.
Where chatbots meet NPS
Bots and NPS intersect in two places, one healthy and one not.
The healthy one: a chatbot is a strong delivery channel for a relationship survey. A 0-10 ask sent through a messaging channel the customer already uses, built with quick replies or a numeric prompt, tends to collect more answers than the same question sent by email (benchmark against your own email survey rather than trusting any published rate), and the bot can branch on the answer in the same conversation: a detractor gets a follow-up question and a fast path to a human, a promoter gets a well-timed review or referral ask. Flow-first platforms (SendPulse, Manychat) treat this as an ordinary flow with tagging; support-desk products (Intercom, Tidio) tend to keep NPS in their survey or reporting layers — where each product actually puts it, and at which tier, is the kind of detail that moves between plans, so verify against current documentation or the individual review before you build.
The unhealthy one: treating NPS as a chatbot performance metric. The bot did not earn the customer's loyalty and cannot lose it alone; it is one touchpoint inside a relationship built by product, pricing, and every other interaction the company has. A bot's own scorecard lives in the metrics stack — containment, resolution, CSAT — and NPS belongs a level up, on the company dashboard next to the numbers a board actually asks about, as our board-reporting guide lays out.
Related terms
- Chatbot CSAT — the transaction-scoped satisfaction survey NPS is most often confused with.
- Sentiment analysis — the inferred, per-message feeling signal; no survey involved.
- Human handoff — where a detractor response should land within days, not quarters.
- Chatbot quick reply — the UI element that makes a 0-10 scale answerable in chat.
- Customer service chatbot — the bot category whose conversations CSAT (not NPS) should grade.
FAQ
What is a good Net Promoter Score?
Above zero means promoters outnumber detractors, which is the only universally meaningful line. Beyond that, good is relative to your industry and your own history: sectors grade on different curves, and Bain's guidance stresses comparing against competitors rather than chasing an absolute target. For an SMB, "higher than last quarter, on a sample big enough to trust" is a more useful definition of good than any league-table number.
Why does a 7 count as zero?
By design. The buckets came from matching self-reported scores against actual repurchase and referral behavior, and 7s and 8s described themselves as satisfied without acting loyal, so the system counts only enthusiasm (9-10) for you and everything from 0 to 6 against you. Customers do not know this, which produces the classic mismatch: someone hands you a 7 believing it is a compliment, and the arithmetic shrugs.
What is the difference between NPS and CSAT?
Scope and timing. CSAT asks about one interaction right after it happens; NPS asks about the whole relationship on a slower cadence. CSAT tells you whether the conversation your bot just handled was any good; NPS tells you whether customers would stake their reputation on recommending the company. Use both, at their own scopes, and be suspicious of any setup that fires the recommend question at the end of a support chat.
What is the difference between relationship and transactional NPS?
Relationship NPS surveys customers on a schedule (quarterly, biannually) regardless of recent activity, and measures the brand overall. Transactional NPS fires after a specific event, such as onboarding or a delivery, and measures that touchpoint. The transactional variant drifts toward CSAT territory; if what you want to grade is a single interaction, CSAT's direct question usually does the job with less ambiguity.
How many responses do I need for a reliable NPS?
More than most SMBs collect. With 50 responses each respondent moves the score two points, so swings of five to ten points are ordinary noise. There is no magic threshold, but treat scores built on fewer than a couple hundred responses as directional at best, watch trends across periods rather than single readings, and resist re-running the survey until enough new answers exist to change the picture.
Sources
- Bain & Company. Measuring Your Net Promoter Score — Net Promoter System. netpromotersystem.com/about/measuring-your-net-promoter-score (verified 12 July 2026).
- Frederick F. Reichheld. The One Number You Need to Grow. Harvard Business Review, December 2003. hbr.org/2003/12/the-one-number-you-need-to-grow (verified 12 July 2026).
- Bain & Company. Trademarks and Licensing — Net Promoter System. netpromotersystem.com/resources/trademarks-and-licensing (verified 12 July 2026).
- Chatbotscape Glossary. Chatbot CSAT. /glossary/chatbot-csat (verified 12 July 2026).
- Chatbotscape evaluation methodology. /methodology (continuously updated).