ISSUE 009
Does Net Promoter Score Predict Anything?
Somebody will eventually send you a one-question survey asking how likely you are to recommend them, on a scale of nought to ten. It is the most widely used customer metric in the world and it has a substantial academic literature arguing it is no better than the obvious alternative. I went to read both sides. The honest summary is duller than either camp would like, and it has a practical consequence for anyone with fewer than a few hundred customers.
An idea worth thinking about
Eleven points of information, thrown away on purpose
The mechanics first, because they are the argument. You ask one question — how likely are you to recommend us — on a nought to ten scale. Anyone answering nought to six is labelled a detractor, seven or eight a passive, nine or ten a promoter. The score is the percentage of promoters minus the percentage of detractors. Eleven possible answers collapse into three groups, and then two of the three become a single number.
A statistician at the University of Sydney and a retired customer-measurement man from AT&T published a critical review of this in a statistics journal in 2019. Their objections are structural. A customer who rated you nought and a customer who rated you six are recorded identically, which is a strange thing to do to the two most different unhappy people you have. The sevens and eights vanish entirely, and the authors argue that there is no such thing as a passive customer. The case for the metric rests on correlation with growth, and they note, drily, that correlation does not equate to causation. Their summary line is that there appears to be little hard evidence that the score is a lead indicator of business outcomes.
The one that should worry a small business most is mechanical. Your score can rise because detractors leave. Once they have gone, you stop surveying them, so the arithmetic improves while the business gets worse. The authors illustrate it with a bank and with Oldsmobile, and it is the sort of failure that is invisible from inside. They also give an older example from AT&T in the 1980s: combining the Good and Excellent answers produced a 95 percent satisfaction figure, while about 40 percent of the people who said Good turned out not to be very willing to buy again. Forty-year-old data, and still the clearest illustration of what grouping costs you.
Now the disclosure, which matters here. Both authors have written books promoting Customer Value Management, the paper recommends Customer Value Management as the better approach, and the PDF is hosted on one author's own consultancy site. This is a critique of one method by people with another. It also reports no new data and runs no replication — it is a conceptual and structural argument, not a measurement. Read it for the structure, which is sound, and take the recommendation with the same caution you would take any other.
A tool I disappeared into
Find out how wide your score really is
VassarStats is a set of free statistical calculators that Richard Lowry has kept online since about 2001, and it looks exactly as old as that. No account, no cookie banner, no product. The page I want you to open computes the 95 percent confidence interval for a proportion using the Wilson procedure, with and without a continuity correction, following Newcombe's 1998 comparison of the methods. Two boxes: the number of observations of interest, and the total.
Here is the use. Say forty customers answered your survey and twenty-two of them gave you a nine or a ten. Enter 22 and 40. The interval that comes back is the honest range your 55 percent promoter share actually represents, and it is much wider than 55 percent looks. Do the same for your detractors. Then look at the two ranges and ask whether the difference between this quarter and last quarter could survive them.
The limit matters and I want to be exact about it. This calculates an interval for one proportion. A net promoter score is a difference between two proportions drawn from the same respondents, and the interval around that difference is wider than the interval around either piece — that last point is my inference from how such intervals combine, not something the page says. So what you get here is a floor on your uncertainty, not the true figure. If the floor already swallows the change you were about to report, you have your answer and you did not need the exact version.
This is the most useful thing in this entry for a business with a few hundred customers. Most small-business scores are computed from twenty or thirty responses and reported to the nearest whole number as if the second digit meant something. Five minutes on a 25-year-old web page will cure that.
Learn something
A number that cannot tell you what to do about it
Douglas Grisaffe published a careful critique in 2007, four years after the original claim, in an open-access academic journal, and the PDF is free. He is explicit that his objections are conceptual rather than empirical — he points at a separate 2007 paper for the empirical challenge — and the conceptual objections are the ones a business owner can act on.
His central complaint is about actionability, and he builds it out of the original's own sentences. The promise was that replacing a complex satisfaction survey with one question lets companies finally put survey results to use. Grisaffe's response is that it is counter-intuitive for one general question to be easier to act on than several specific ones. To move the number, you have to know which underlying thing is dragging it, and a single score cannot contain that. His line is that how to make the score move up is not knowable based on the score itself.
Then he catches the original contradicting itself. The source he quotes recommends follow-up questions to unearth the reasons behind the feelings, tailored differently to detractors, passives and promoters — which is to say, a set of specific diagnostic questions. Which was the thing the one-question approach was supposed to replace.
His last point is the one I had never thought about. Is likelihood to recommend a cause of loyalty, a measure of loyalty, or a consequence of it? Grisaffe argues the measurement science was never settled, and that it matters: if the answer is a consequence, then managing it directly is managing a symptom. The paper is from 2007 and some of the debate has moved on, but nobody has repealed that question, and it is worth asking of any single metric you are being told to manage.
What we’re seeing
120,990 skiers, and the ordinary satisfaction question predicted better
The best test of whether this metric predicts anything was published open access in early 2026, and it is unusually well built. Researchers got electronic lift-access data from 38 of the top 55 Alpine ski resorts — actual turnstile counts, not self-reported anything — across eight seasons from 2010/11 to 2018/19. They matched it to bi-annual surveys of skiers at those same destinations running from 2011/12 to 2017/18, totalling 120,990 responses. That gave 136 destination-year observations, analysed with panel regression and fixed effects for destination and year.
First finding: it works. The net promoter score significantly predicts skier visits. Anyone who tells you the metric is worthless has to get past this study, and it is the better-designed one.
Second finding: so does asking people how satisfied they are, and better. The authors ran a formal test of whether the two coefficients differ — F of 4.55 on one and 37 degrees of freedom, significant at five percent — and satisfaction has the significantly stronger influence. Their sentence: the net promoter score does not outperform satisfaction in its influence on skier visits. Perceived customer value, the third metric they tested, did not predict visits at all. I am deliberately not comparing the two coefficients by size, because they are measured on different scales and the comparison would be meaningless; the test of difference is the thing.
Third finding, and the strangest. The predictive power falls as the surveyed skiers get more familiar with the resort. Past an average destination experience of 2.3 — which the authors gloss as the average respondent having visited slightly more than five times — the score no longer significantly predicts visits at all. Read that against your own customer base before you run a survey on your regulars. There is also an internal inconsistency worth flagging: the abstract says the two-year lag has the strongest explanatory power while the results section says the one-year lag does. The results section is the more detailed account.
The limits are large and they are the usual ones. Alpine ski destinations are not firms, a lift turnstile count is not revenue, and none of this is a six-person company in Cape Girardeau. What transfers is not the level but the comparison, and the comparison is now fairly clear across the sources in this entry. The critics say there is little hard evidence the score leads business outcomes; this study finds evidence that it does. Both can be true, because the useful question was never whether it correlates with anything. It was whether it beats the question you would otherwise have asked. On the best evidence available, it does not.
So, practically. If you are already running it, keep it, because a consistent series you understand is worth more than a better metric you restart every year — but put a confidence interval on it before you report a change, and stop treating a two-point move as news. If you are not running it, there is no evidence here that adopting it will tell you more than asking people how satisfied they are and why. And whichever you ask, it remains downstream of the measurement this pillar settled on back in entry five: take everyone who bought from you in one month last year, and count how many came back.
One question
If you could ask your customers only one question, what would it be?
One line. I am curious how many of us would pick the recommend question if nobody had ever suggested it.
More in this pillar as it's written.
— Shana
Sources link to the original publisher, never to a summary.