Net promoter score
In one sentence
Net promoter score is a relationship metric that subtracts the percentage of detractors, who score zero to six, from the percentage of promoters, who score nine or ten, on a zero to ten recommendation scale, producing a number between minus one hundred and plus one hundred.
Not to be confused with CSAT.
Definition
Net promoter score asks how likely someone is to recommend a company on a zero to ten scale, then subtracts the percentage of detractors from the percentage of promoters. The result sits somewhere between minus one hundred and plus one hundred.
It measures how people feel about your company as a whole. It does not measure whether one conversation went well.
NPS was introduced by Fred Reichheld and Bain and popularized through a 2003 Harvard Business Review article. Its staying power owes more to its simplicity than to its predictive strength, which has been contested in the academic literature ever since.
How the number is built
- Promoters score nine or ten.
- Passives score seven or eight and are excluded entirely.
- Detractors score zero to six.
- NPS equals the promoter percentage minus the detractor percentage.
The structural criticisms
- Discarding the passives throws away information from a large share of respondents.
- Two very different response distributions can produce an identical score, which the single number hides.
- The evidence that NPS predicts better than a simple satisfaction question has been disputed rather than confirmed.
- It is a relationship metric, and it gets used, frequently, as a transactional one.
Why it persists anyway
- A single number that executives understand and can track over time.
- Wide adoption creates a network effect in benchmarking, which survives the comparability problems rather than solving them.
Its relevance to a voice agent, which is limited
- NPS measures the relationship with the company, not the quality of one interaction.
- A single conversation with a voice agent rarely moves it.
- Attributing an NPS change to a voice deployment needs controls most organizations do not have, because dozens of other factors move the relationship.
- Transactional CSAT and customer effort score are both better suited to judging an interaction.
Common misconception
That NPS measures satisfaction with an interaction. It measures relationship-level advocacy. Using it transactionally, to score a single conversation, is a category error, and it produces noisy data that reads as signal.
Why it matters commercially
NPS turns up in board reporting, so anyone selling into an organization has to understand it. It is a poor choice for evaluating a voice agent deployment, and saying so plainly is more useful than producing a number that cannot be attributed to anything.
In voice specifically
A single website conversation is exactly the unit NPS is worst at reading. Speech does not change that. The signals that do move on one conversation are transcript sentiment and whether the visitor got an answer, neither of which is what an advocacy survey is asking about.
Where AsqVox fits
NPS is not an appropriate evaluation metric for a single website conversation. Sentiment on transcripts and the outcome of the conversation are the signals the Orb dashboard surfaces, and those are the ones worth reading against a deployment.
Visual
The number that discards the middle
Two bands on the scale get counted, the middle gets thrown away, and two very different customer bases can land on the same score.
Statistics
Every figure carries its source and year. Vendor numbers are labelled as vendor numbers, and where no reliable figure exists this page says so rather than borrowing one.
Net promoter score was introduced by Fred Reichheld and Bain and popularized through a Harvard Business Review article, and its calculation excludes the passives scoring seven or eight, so a substantial share of responses never contribute to the score.
2003independentHarvard Business Review, Fred Reichheld and Bain, 2003 - The 2003 origin is the one hard fact on this page and belongs in any citation. The passives exclusion is structural, not incidental: it is why the same score can describe two different customer bases.
Cross-industry NPS benchmarks are widely published and are of limited value, because scores vary systematically by industry, geography and survey timing rather than by service quality.
-no reliable figureIf a target NPS is quoted at you as an industry standard, ask which industries, which regions and which survey timing produced it. The comparison usually collapses at that question.
The academic literature contesting whether NPS predicts better than simpler satisfaction measures is substantial, and the honest position is to present it as contested rather than settled in either direction.
-no reliable figureDo not cite NPS as a proven predictor of growth, and do not cite it as debunked. Both overstate what the evidence supports.
There is no published research establishing an NPS movement attributable to a voice agent deployment, which is unsurprising given how hard the attribution is.
-no reliable figureA single conversation rarely moves a relationship metric, and too many other factors move it at the same time to isolate one deployment. Measure the interaction with a transactional metric instead.
Examples
In practice
An organization sets an NPS improvement target for a voice agent deployment. After six months NPS has moved within its normal variance and no attribution is possible, because dozens of other factors affect the relationship. The evaluation is rebuilt around transactional CSAT on contained conversations and repeat contact rate, both of which move measurably and can actually be attributed to the deployment.
The everyday version
Net promoter score asks whether someone would recommend you, out of ten, and turns the answers into one score. It is a reasonable temperature check on how people feel about your company overall. It is the wrong tool for judging whether one conversation with your website went well.
Usage
Who says it
- Executives and board-level reporting, prominently, where a single tracked number is the point.
- CX teams, often with reservations, because they know its limits better than the board does.
Where it turns up
- In board packs, annual reporting and strategic dashboards.
- Occasionally in RFPs as an outcome target for a deployment, which is usually a mistake because the effect on it cannot be isolated.
Common misuse
- Applying it transactionally to individual interactions, which is a category error.
- Comparing scores across industries as if they were on a common scale.
- Setting it as the success target for a deployment whose effect on it cannot be separated from everything else.
Questions people ask
How is net promoter score calculated?
Ask how likely someone is to recommend you on a zero to ten scale. Promoters score nine or ten, passives score seven or eight and are excluded entirely, and detractors score zero to six. NPS is the promoter percentage minus the detractor percentage, so it ranges from minus one hundred to plus one hundred.
What is the difference between NPS and CSAT?
NPS measures relationship-level advocacy, how people feel about your company overall. CSAT measures satisfaction with a specific interaction, asked right after it. Using NPS to score one conversation is the most common way the two get confused, and it produces noisy data because a single interaction rarely moves a relationship metric.
Is NPS a good metric for a voice agent?
No, not for evaluating one. NPS measures the relationship with the company, and a single conversation with a voice agent rarely moves it. Attributing any change to the deployment needs controls most organizations lack. Transactional CSAT and customer effort score are the appropriate metrics for judging an interaction.
Is there a good benchmark NPS score?
Cross-industry NPS benchmarks are widely published and of limited value, because scores vary systematically by industry, geography and survey timing rather than by service quality. Whether NPS even predicts better than a simpler satisfaction question is contested in the academic literature rather than settled.
Last reviewed 4 August 2026. Written and reviewed by Dhruv Dholakia, founder of AsqVox.