A 265,000-Company Study Found Zero Correlation Between NPS and Churn. Here's What Actually Predicts It.
A decade of data across 265,000 subscription businesses found no statistical link between NPS and churn risk. Here's what to track instead.
Most customer success teams still treat NPS as an early-warning system — a detractor score pops up, someone gets flagged for outreach, a save play gets triggered. It feels like it should work. A customer who's unhappy enough to mark you a 3 out of 10 seems like a customer who's halfway out the door. The data doesn't back that up, and the study behind it is bigger than anything else in this space.
The study: 265,000 companies, a decade of data, no correlation
In 2023, Customer Cross — a Ghent, Belgium-based customer success firm that works with subscription businesses across the US and Europe — ran the largest study we've seen on this specific question. They pulled a decade of data from roughly 265,000 subscription and recurring-revenue businesses: SaaS, telecom, data services, anything billing on a renewal cycle. The question was simple: does NPS actually correlate with which customers churn and which ones stay?
It doesn't. Not in either direction. High-NPS customers weren't reliably more likely to renew. Low-NPS customers weren't reliably more likely to cancel. The report puts it plainly — there's no statistical evidence that a customer's satisfaction score, whether NPS or CSAT, predicts how long they'll stick around. That's a harder finding to dismiss than the usual "NPS is flawed" take you see in a LinkedIn post, because it's not a theoretical critique — it's a quarter-million accounts of actual renewal behavior checked against the scores those same accounts gave.
Why the score breaks down at the individual level
The mechanism matters more than the headline, because it tells you whether the problem is fixable with a better survey tool or whether it's structural. Customer Cross points to three things, and all three hold up against what anyone running a CS org has probably already noticed:
- Response rates are too low to be representative. Most NPS programs see single-digit to low-teens response rates. The customers who bother to answer aren't a random sample of your base — they skew toward people with either a strong compliment or a strong complaint already in mind.
- Responses cluster at the extremes. A 9 or a 10, or a 2 or a 3. The middle of the scale, where most of the actual signal about gradual disengagement would live, gets very little traffic.
- Your most engaged customers are often your toughest graders. A power user who's deeply invested in your product and plans to renew for years will still mark you a 6 because they have three specific complaints — and a customer who's quietly checked out and barely opens the product anymore might mark you an 8 just to be polite. The score measures how someone feels answering a survey, not how committed they are to staying.
None of that is a reason to stop running NPS. It's a reason to stop wiring it into your churn alerts as if a dropping score means a specific account is about to leave.
What NPS is actually good at: expansion, not retention
Here's where it gets more useful than a simple "NPS doesn't work" takeaway. Gainsight's review of the research, drawing on ProfitWell data, found NPS does correlate with something — just not churn. It tracks expansion revenue reasonably well.
Source: ProfitWell data, reported via Gainsight. Bottom quartile earns ~5% less expansion revenue than median, ~15% less than top quartile.
That gap is real money, and it's worth budgeting for — a customer who loves your product enough to give you a 9 is a better upsell target than one who grudgingly gives you a 6, even if both of them are equally likely to renew next year. The mistake is using the same number to answer two different questions. NPS is a reasonable proxy for "who's primed to buy more." It is not a reasonable proxy for "who's about to leave."
Where NPS does earn a seat at the churn table
There's one documented exception worth knowing, because it keeps the picture honest rather than just swinging to "NPS is worthless." Gainsight's analysis found that companies landing in the top 25% of NPS within their specific vertical see a real 5–10% bump in renewal rate. The score isn't meaningless at the extreme top end of a benchmarked comparison — it's meaningless as a general-purpose early-warning trigger for an individual account.
| What's measured | Correlation found | Source |
|---|---|---|
| Churn / renewal risk, general population | None — no statistical link | Customer Cross, 2023 (265,000 companies, 10 years of data) |
| Renewal rate, top 25% NPS within vertical | +5–10% renewal bump | Gainsight review of NPS research |
| Monthly expansion revenue, bottom vs. median quartile | ~5% less than median | ProfitWell data via Gainsight |
| Monthly expansion revenue, bottom vs. top quartile | ~15% less than top quartile | ProfitWell data via Gainsight |
Read that table as a map of what to automate and what to leave to judgment. Expansion motion and top-tier benchmarking are fine places to let an NPS threshold trigger a playbook. An individual account's churn risk is not, and building your save-play triggers on that assumption means you're spending CS time chasing low scorers who were never going to leave, while accounts that quietly stopped logging in — and gave you an 8 last quarter because they were being polite — slide to cancellation with nobody watching.
What to wire into your churn alerts instead
The fix isn't a better survey question. It's accepting that churn risk lives in behavior, not sentiment, and building your signal set around that. We've covered the mechanics of this in more depth in our guide to building a health score that actually predicts cancellation — usage trend, feature adoption depth, and billing behavior shift all move before a customer decides to leave, where NPS only reflects a mood at whatever moment they happened to answer a survey.
The second half of the fix is capturing the moment itself instead of guessing at it in advance. A well-built cancellation survey shown in the flow — not a follow-up email after the fact — gets answered by 60–80% of cancelling subscribers, versus under 5% for a post-cancellation ask. That's not a replacement for predictive signal, it's the ground truth you check your predictions against. If your health score flagged an account as safe and it still showed up on the cancel page citing a specific, nameable reason, that's more useful data than another NPS data point — it tells you exactly which of your leading indicators missed, and why. We break down the actual distribution of those reasons in why customers cancel, and it's worth checking your own cancel data against it rather than assuming your NPS detractors and your actual churned accounts are the same list, because the research says they usually aren't.
Keep running NPS. Just stop asking it to do this job
None of this is an argument to kill your NPS program. It's useful for expansion targeting, for marketing testimonials, and as one input into a broader product roadmap conversation. What it isn't is a churn-prediction tool, and treating a detractor score as a cancellation trigger means you're allocating retention effort against noise instead of signal. Pull your last two quarters of actual cancellations and check them against the NPS scores those accounts gave beforehand — if the overlap is thin, which the research says it will be, that's your confirmation to move the budget toward usage and billing signals instead. Run the at-risk segment you do identify through our churn calculator to see what it's actually worth in retained MRR before deciding how much to invest in catching it.
This is also exactly the gap a cancellation flow is built to close. A health score built on the right leading indicators will still miss accounts — nothing catches everyone before they decide to leave. When one of those accounts reaches the cancel button anyway, CancelFlow captures the real reason in the moment, routes the offer that actually matches it, and hands that data back so your next quarter's model gets sharper instead of relying on a satisfaction survey that was never measuring the right thing.
Frequently asked questions
Does a high NPS score mean a customer is less likely to churn?+
Not reliably, no. Customer Cross, a Ghent-based customer success firm, studied roughly 265,000 subscription businesses across the US and Europe using a decade of data and found no statistical correlation between NPS and churn risk — high scorers weren't meaningfully less likely to cancel, and low scorers weren't meaningfully more likely to. The one exception: companies that land in the top 25% of NPS within their specific vertical see a real 5–10% bump in renewal rate, according to Gainsight's review of the research. Outside that top tier, the score doesn't separate stayers from leavers.
If NPS doesn't predict churn, what should SaaS companies track instead?+
Leading indicators that move before a cancellation decision, not after a survey gets answered: usage trend over a trailing 30 days, depth of adoption on the feature that actually drives retention for that account's use case, and billing behavior like a downgrade from annual to monthly. Better still, capture cancel reasons directly at the moment someone clicks cancel — response rates on an in-flow cancel survey run 60–80%, versus under 5% for a post-cancellation email, and the data is tied to an actual decision instead of a mood on a random Tuesday.
Is NPS a waste of time for SaaS companies to track?+
No — it's just solving a different problem than most teams think it is. ProfitWell's data, reported through Gainsight, found companies in the bottom NPS quartile generate about 5% less monthly expansion revenue than the median quartile and roughly 15% less than the top quartile. NPS tracks upsell and cross-sell potential reasonably well. It was never built to flag which specific account is about to cancel, and the data says it doesn't.
Why do NPS surveys fail to predict individual churn risk?+
Three structural reasons, per the Customer Cross research: response rates are low enough that the respondents aren't representative of your base, the responses that do come in skew bimodal — mostly very positive or very negative, with little in the middle — and your most engaged customers are often your most demanding, so they'll hand you a mediocre score with no intention of leaving. A single snapshot survey can't tell those customers apart from someone who's actually on the way out.
Stop losing subscribers today
One script tag. One function call. A live cancellation flow in under 10 minutes.
Start free trial →