NPS Calculator
Net Promoter Score is the percentage of promoters (9 or 10) minus the percentage of detractors (0 to 6) among everyone who answered the recommend question. The subtraction takes a second. The hard part is knowing whether the number moved because customers changed or because a different few hundred people answered this time. This calculator gives you the score, the range it plausibly sits in, and a test for whether this period really differs from the last.
Calculate your NPS, and how far to trust it
Enter promoters, passives and detractors, or paste the raw 0-10 answers from your survey export. You get the score, its 95% confidence interval, how many responses a tighter read needs, and whether a change from last period is bigger than noise.
NPS +19.9, 95% interval +11 to +29, 312 responses.
Net Promoter Score
+19.9
95% confidence interval
+11 to +29
Responses
312
Who answered what
45.2% promoters minus 25.3% detractors = +19.9. The 29.5% passives count in the total and nowhere else.
Where your true NPS plausibly sits
Marker = measured score. Band = adjusted-Wald 95% interval, about ±9.0 points (the textbook Wald formula gives ±9.1). Dashed line = zero, where promoters and detractors are equal. One more promoter would move this score by +0.26 points, one more detractor by -0.38; one detractor becoming a promoter moves it by +0.64.
Response rate: 21.5% of 1,450 surveys.
Responses needed for a tighter read
| Margin of error | Responses needed | You have |
|---|---|---|
| ±3 points | 2,842 | 312 |
| ±5 points | 1,023 | 312 |
| ±10 points | 256 | 312 (enough) |
Planned at your observed answer mix. To spot a real 10-point change between two periods with 80% power you need about 1,045 responses in each period; a 5-point change needs about 4,180.
Embed this NPS calculator on your site (free)
<iframe
src="https://chatbotscape.com/embed/tools/nps-calculator/"
width="100%" height="1000" frameborder="0"
title="NPS Calculator by Chatbotscape"
loading="lazy">
</iframe>How to calculate NPS
The recipe is fixed, which is most of its appeal. One question: “How likely are you to recommend us to a friend or colleague?” One scale, 0 to 10. Three buckets: promoters answer 9 or 10, passives 7 or 8, detractors 0 to 6. Then:
NPS = % promoters − % detractorsThe default numbers in the calculator are a synthetic quarter for a small online store, not customer data: 141 promoters, 92 passives and 79 detractors, 312 responses in all. Promoters are 45.2% of the total and detractors 25.3%, so the NPS is +19.9. The 29.5% who answered 7 or 8 are in the denominator and nowhere else.
Two properties of that subtraction trip people up. The score is a spread, not a percentage: +20 does not mean 20% of anyone. And passives drag the score toward zero without ever being subtracted, so the shortcut of dividing by promoters plus detractors overstates it. Our Net Promoter Score glossary entry covers where the buckets came from and the criticism the metric has drawn since 2003.
The margin of error nobody reports
An NPS from a survey is an estimate of the score you would get if every customer answered. How far off it can be depends on two things: how many people answered, and how split they were. A customer base where everyone answers 7 has no spread at all; one divided between 10s and 0s has the most. The per-respondent variance captures both:
variance = p(promoters) + p(detractors) − (p(promoters) − p(detractors))²The textbook 95% margin is 1.96 times the square root of that variance divided by the number of responses. For the worked example it comes to ±9.1 points. The calculator reports a slightly different interval, the adjusted-Wald version, which adds three quarters of a response to the promoter and detractor counts and three to the total before doing the same arithmetic. Brendan Rocks compared interval methods for NPS in The American Statistician in 2016 and found that variations on the adjusted-Wald interval performed better than the plain formula; Jim Lewis and Jeff Sauro publish the specific adjustment used here, one of the variants Rocks tested. On 312 responses the two agree closely: the adjusted interval runs from +10.7 to +28.7, about ±9.0 points.
At small samples they do not agree, and the gap is the point. Take 50 responses split 20, 15 and 15. The NPS is +10, the plain margin is ±23 points, and the adjusted interval runs from -13 to +32. That survey cannot say whether promoters outnumber detractors. One more promoter would lift the score by about 1.8 points, one more detractor would drop it by about 2.2, and one detractor who would have answered 9 instead moves it by 4.
How many responses you need
Turn the margin formula around and it tells you how many responses a target precision costs. At the worked example's answer mix, ±10 points needs 256 responses and ±5 points needs 1,023. Halving the margin costs about four times the responses. The calculator plans at your own observed mix, because a split audience needs more responses than a uniform one. It never plans below a quarter of the worst-case spread, so a lopsided early sample cannot promise a tiny survey, and with no spread to go on it plans at the worst case, half promoters and half detractors.
Detecting a change is more expensive than measuring a level, because both periods carry their own error. To have an 80% chance of spotting a real 10-point move at p < 0.05, the worked example needs about 1,045 responses in each period; a 5-point move needs about 4,180. At the worst-case mix a 10-point move needs 1,570. If your survey collects fewer than that per period, that is not a reason to skip NPS, but it is a reason to compare across longer windows, and to stop reporting single-digit moves as news.
Did your NPS really change?
Here is the trap in numbers. Last quarter the same store collected 268 responses: 104 promoters, 85 passives and 79 detractors, for an NPS of +9.3. This quarter it is +19.9. A 10.5-point rise is the kind of number that ends up on a slide with an arrow.
The two-sample test says otherwise. The 95% interval for the change runs from -2.8 to +23.7 points, z = 1.54, and p = 0.123. If nothing had changed, a gap this large would still turn up about one time in eight, so the honest report is that the score may have improved and the data cannot yet show it. Notice what does hold up: last quarter's interval (-0.5 to +19.0) included zero, and this quarter's does not. The store can now say promoters outnumber detractors. It cannot yet say the quarter changed anything.
The test is the one Lewis and Sauro describe for comparing two Net Promoter Scores: adjusted proportions in each sample, the standard error of the difference as the square root of the sum of each sample's variance over its size, and a two-tailed z. The same test works for two segments in the same period, such as customers who used your chatbot against customers who did not, as long as the question and scale are identical.
What counts as a good NPS
Zero is the only line with meaning on its own: above it, more customers would recommend you than would warn people off. Past that, a score is good or bad relative to something. The usual advice is to compare against competitors in the same market, because sectors grade on very different curves. A small business without competitor data has a better comparison to hand: its own score last period, on samples big enough to tell the difference.
That is why the verdict in the calculator is about zero and about change, and why this page has no table of “average NPS by industry.” Tables like that circulate widely; we have not found one that states its sampling method clearly enough to check, and a benchmark of unknown method is a worse comparison than your own history.
Where chatbots fit, and where they do not
A chatbot is a good delivery channel for a relationship survey. A 0 to 10 question sent through a messaging channel the customer already uses, with quick replies for the scale, is easy to answer, and the bot can branch on the answer: a detractor gets a follow-up question and a route to a person, a promoter a review or referral ask. Our guide to running NPS through a chatbot covers timing, sampling and closing the loop.
A chatbot is a poor thing to grade with NPS. The recommend question asks about the whole relationship; fired at the end of a support chat it mostly measures the last five minutes. For the conversation itself use CSAT, and our CSAT calculator gives it the same treatment this page gives NPS: an interval and a significance test. If you do compare NPS between customers who used the bot and customers who did not, remember the groups chose themselves, so a significant gap shows a difference between the groups, not that the bot caused it.
Why we built it this way
Many NPS calculators stop at the subtraction; several now add a margin of error, and a few a comparison test. We wanted the three questions an owner actually asks after a survey closes, answered from their own numbers: what is the score, could it really be zero or negative, and is it different from last time. The paste mode exists because survey tools export a column of 0 to 10 answers, and counting buckets by hand is where errors creep in. The planner exists because the usual answer to “how many responses do I need?” is a single number, when the answer depends on whether you are measuring a level or a change.
The engine is checked before every build against hand-computed values and against the worked example Lewis and Sauro publish (their two-sample example gives z = 1.81 and p = 0.07; ours reproduces both). Nothing on this page comes from a customer deployment or from survey data we collected.
Limitations you should state alongside the number
The interval covers sampling error and nothing else. It assumes the people who answered are a random draw from the customers you want to describe. They often are not: strongly satisfied and strongly dissatisfied customers tend to answer more readily, surveys sent only to recent buyers flatter the score, and staff who coach “anything below a 9 hurts me” inflate it. When fewer than 15% of surveyed customers answer, the calculator flags it, because at that point who chose to answer can matter more than how many did.
The test and planner use the normal approximation, which is reasonable at the sample sizes where NPS is worth reading and rough below about 30 responses per sample; the calculator says so when it applies. Testing many segments or many periods at once will turn up a “significant” difference by chance sooner or later, so decide which comparison matters before looking. And NPS is a loyalty thermometer, not a forecast: the claim that it predicts growth better than other survey measures was challenged by Keiningham and colleagues in the Journal of Marketing in 2007. Nothing here is advice about a specific business decision.
Related Chatbotscape tools and resources
Some review links on this site go to platforms we may earn a commission from. That relationship does not affect scores, rankings, or what this calculator computes. See our affiliate disclosure for how it works.
- Chatbot CSAT calculator: the conversation-level survey, with a confidence interval and a bot-vs-human test
- Containment rate calculator: what share of conversations the bot actually finishes
- Chatbot ROI calculator: the cost side, kept separate from loyalty
- Customer lifetime value calculator: what keeping a customer is worth in gross margin
- Net Promoter Score: definition, history and what counts as good
- Customer effort score: the third survey metric, scoped to a single service interaction
- Sentiment analysis: the inferred signal, with no survey at all
- Running NPS surveys through a chatbot: timing, setup and the detractor follow-up
- Improving chatbot CSAT without gaming it
- The chatbot metrics guide: where NPS sits relative to the bot's own scorecard
- The chatbot KPIs a board cares about: reporting a score as a range with its sample size
FAQ
How do you calculate NPS?
Ask "How likely are you to recommend us to a friend or colleague?" on a 0 to 10 scale. Count 9s and 10s as promoters, 7s and 8s as passives, and 0 to 6 as detractors. NPS is the percentage of promoters minus the percentage of detractors, so it runs from -100 to +100. With 141 promoters, 92 passives and 79 detractors out of 312 responses, that is 45.2% minus 25.3%, or +19.9.
Do passives count in NPS?
They count in the total and nowhere else. Passives are never subtracted, but they are in the denominator, so more passives pull the score toward zero. A common mistake is to divide by promoters plus detractors only, which drops the passives and inflates the score.
What is the margin of error for NPS?
It depends on the number of responses and on how split they are. The per-respondent variance of NPS is the promoter share plus the detractor share minus the square of their difference, and the 95% margin is 1.96 times the square root of that variance over n. On the worked example's 312 responses it is about ±9.1 points. With 50 responses split 20, 15 and 15 it is about ±23 points, wide enough that an NPS of +10 cannot be told apart from zero.
How many responses do I need for a reliable NPS?
For a margin of ±10 points at a typical answer mix, a few hundred; the worked example needs 256. For ±5 points, about 1,023. To tell a real 10-point change between two periods from noise with 80% power you need about 1,045 responses in each period, and roughly four times that for a 5-point change.
Is a change in NPS from last quarter statistically significant?
Run a two-sample test rather than eyeballing the gap. In the worked example NPS rose from +9.3 to +19.9, a 10.5-point jump, on 268 and 312 responses. The test gives p = 0.123, so the jump is not statistically significant: if nothing had changed, a gap this large would still turn up about 12% of the time by chance. Turn on the comparison in the calculator to test your own two periods.
What is a good NPS score?
Above zero means promoters outnumber detractors, and that is the only threshold with meaning on its own. Beyond it, good is relative: industries score on very different curves, so the usual advice is to compare against competitors in the same market. For a small business, a score higher than last period on samples big enough to tell the difference is the useful definition. We do not publish an industry benchmark table because we have not found one whose method we could check.
Should a chatbot ask the NPS question after every conversation?
No. NPS grades the whole relationship; a post-chat survey grades the conversation that just ended, which is what CSAT is for. A chatbot is a good channel for delivering a relationship survey on a schedule, and a poor place to ask the recommend question at the end of a support chat.
Does the calculator send my data anywhere?
No. All math runs in your browser, including the pasted-scores mode. Nothing you type is uploaded, stored, or logged, and refreshing the page clears it.
Can I embed this calculator on my site?
Yes, free. Copy the iframe snippet from the embed section on this page. The embed strips Chatbotscape navigation and keeps the calculator plus attribution.
Sources and basis
- Definition.Question wording, the 0 to 10 scale, the promoter, passive and detractor buckets and the formula: Bain & Company, Measuring Your Net Promoter Score, read 10 November 2026. Net Promoter, NPS and Net Promoter System are registered trademarks, and Net Promoter Score is a service mark, of Bain & Company, Inc., NICE Systems, Inc. and Fred Reichheld (per Bain's trademark notice, read 10 November 2026).
- Variance and intervals.Brendan Rocks, “Interval Estimation for the ‘Net Promoter Score’,” The American Statistician 70(4), 2016, pp. 365-372; preprint on arXiv, read 10 November 2026.
- Adjustment and two-sample test. Jim Lewis and Jeff Sauro, How to Statistically Compare Two Net Promoter Scores, MeasuringU, 2 February 2021, read 10 November 2026. The calculator uses their pseudo-counts (three quarters of a response added to promoters and to detractors, three to the total) for both the single-sample interval and the comparison.
- Growth-prediction critique.Timothy L. Keiningham, Bruce Cooil, Tor Wallin Andreassen and Lerzan Aksoy, “A Longitudinal Examination of Net Promoter and Firm Revenue Growth,” Journal of Marketing 71(3), 2007.
- Planning formulas. Responses for a margin e: 1.96² × variance / e². Responses per period for a change d at 80% power: (1.96 + 0.84)² × 2 × variance / d². Standard normal approximations; the engine is held to them by a script that runs before every build.
- Worked example. Synthetic counts chosen to show a plausible quarter-over-quarter move. They are not survey data from any business, and no figure on this page is a benchmark.