How Do You Find the Real Reasons Behind Customer Churn — Beyond NPS?
The most-mentioned reason customers give is usually NOT the biggest lever to keep them. Here's how causal AI finds the churn drivers that actually move retention.
The short answer
Your NPS score tells you that customers are leaving. It doesn't tell you why — at least not in a way you can act on. A number that drifts from 42 to 38 is an alarm, not a diagnosis. The "why" lives in the open-text verbatims customers write next to the score, but reading them at scale is the problem, and counting the most-frequent complaint is the wrong fix. The reason customers mention most is often the one you can least change. To find the churn drivers that actually move retention, you have to read every verbatim and model it causally — not tally it.
Why NPS alone can't tell you why customers leave
NPS is a monitoring instrument. It compresses a rich, emotional decision — to stay or to go — into a single 0-to-10 rating, and then into one headline number. That's useful for spotting a trend, but it strips out exactly the information you need to act: the reasons. Two brands with the identical score can be losing customers for completely different reasons, and NPS can't tell them apart.
The reasons are sitting right there in the open-text box, where customers explain their rating in their own words. But most organisations never read these verbatims at scale — there are thousands of them — and the few who do usually fall into a trap: they count how often each theme is mentioned and treat the most-mentioned one as the biggest problem. Frequency is not importance. This is the same failure we describe in why market research is broken: measuring what's loud and stated, not what actually causes behaviour.
The most-mentioned reason was not the biggest loyalty lever
SUPRA read the open-text NPS responses instead of just the scores. NLP auto-coded the free-text answers into themes; a causal AI model (Universal Structure Modeling) then estimated which themes actually drive loyalty — capturing indirect effects through sentiment that a linear model averages away.
The Hidden Lever: the single most-mentioned reason — cited by about one in two respondents — was not the biggest loyalty lever, because the brand was already maxed out there with no room to improve. The real levers were lower-frequency themes with headroom to move. Acting on the loudest complaint would have wasted effort where nothing could change; acting on the causal drivers moved retention. The client now auto-codes incoming new-customer NPS in nine languages.
Counting complaints vs. modelling churn causally
| Dimension | NPS + complaint counting | Causal churn-driver analysis (SUPRA) |
|---|---|---|
| Question answered | What do customers complain about most? | Which themes actually cause loyalty, and by how much? |
| Method | Score + manual/keyword tally of verbatims | NLP auto-coding + causal AI (USM) — indirect effects |
| Use | Alerting that something is wrong | Deciding which lever to pull to keep customers |
| Blind spot | Treats the loudest theme as the biggest lever | Finds hidden levers where you still have headroom |
"The reason customers mention most is often the one you can change least. Frequency is not importance — and the hidden lever is where the retention is."
What churn-driver analysis looks like in practice
You don't need a new survey. The open-text field on your existing NPS already holds the "why." SUPRA layers analysis on the data you already collect and returns:
- Auto-coded themes at scale — NLP reads every verbatim and sorts it into consistent categories, so thousands of responses become structured data instead of an unread archive.
- A ranked set of causal loyalty drivers — the themes that actually move retention, separated from the ones customers merely mention often, with indirect sentiment effects captured.
- Hidden levers with headroom — the lower-frequency drivers where you can still improve, rather than the maxed-out complaint everyone already knows about.
The method is the same causal AI and implicit-measurement foundation behind our brand development work, and the reasoning is developed in Dr. Buckler's books Why Brands Grow and The Top 5%.
Frequently asked questions
Why are customers leaving, and how do I keep them?
Customers rarely leave for the reason they name most loudly. NPS tells you satisfaction is slipping, not which underlying driver, if fixed, would keep people. The "why" lives in the open-text verbatims — SUPRA auto-codes them into themes with NLP, then uses causal AI to estimate which themes actually drive loyalty. The levers that keep customers are frequently not the complaints you hear most often.
How do I find the real reasons behind customer churn?
Don't rely on the NPS number or on tallying the most-mentioned complaints. The "why" is in the verbatims, but reading them at scale — and separating what customers say from what causes them to leave — is the hard part. SUPRA uses NLP to auto-code open-text NPS into themes, then causal AI (USM) to find which drive loyalty. In a Sonos study, ~3,100 US responses were coded into ~50 categories; the causal model reached R²=0.70 — about 40% more explanatory power than linear.
What makes customers stay loyal — beyond NPS?
Loyalty is driven by a handful of causal levers hidden in the verbatims, not by the headline score. In the Sonos study, the most-mentioned reason — cited by about one in two respondents — was not the biggest lever, because the brand was already maxed out there. The real levers were lower-frequency themes with headroom, and indirect sentiment effects added ~41% of the explanation. Frequency is not importance.
Can AI analyze NPS verbatims at scale?
Yes. NLP can auto-code open-text NPS into consistent themes, replacing slow, subjective manual coding, and a causal AI model can then estimate which themes actually drive loyalty. That turns thousands of verbatims from an unread archive into a ranked list of retention levers. After the Sonos study, the client now auto-codes new-customer NPS in nine languages automatically.
Find out what your NPS verbatims aren't telling you
Bring your existing NPS open-text data. We'll show you whether a causal read reveals churn drivers — and hidden levers — your score alone is missing.
Get my AI Diagnostic →