Churn Health Scores: Why Most Fail and How to Build One That Actually Predicts Cancellation
73% of CS leaders say their health score doesn't reliably predict churn. Here's what to track instead, and how to build a score that does.
Most SaaS companies past a few hundred customers eventually build some version of a customer health score — a single number that's supposed to tell you who's about to cancel before they do. Most of those scores don't work. Not because the idea is wrong, but because they're built almost entirely from signals that are easy to pull from a dashboard rather than signals that actually move before a customer decides to leave.
Why most health scores fail before they start
Ask most CS or growth teams what feeds their health score and you'll hear some combination of: revenue tier, NPS, support ticket count, and maybe a login flag. Every one of those is a lagging indicator. Revenue tells you what already happened at the account level. NPS is a snapshot from whenever someone last bothered to answer a survey. A support ticket count only moves after something has already gone wrong enough for someone to file a ticket.
None of that is useless — it's just not predictive. A score built from lagging indicators is really a description of the recent past, dressed up as a forecast. That's the actual reason behind the 73% figure above: it's not that the underlying data is bad, it's that most scores are architecturally built to arrive after the decision has already been made rather than while there's still time to change it.
The signals that actually move before someone cancels
Leading indicators are harder to instrument, which is exactly why fewer teams use them systematically. But they're the only category of signal that gives you a window to act.
| Signal | Leading or lagging | What it actually tells you |
|---|---|---|
| Login / session trend | Leading, but noisy | A sustained multi-week decline is meaningful. A single quiet week around a holiday or a busy sprint is not — track the trend, not the raw count. |
| Feature adoption depth | Leading | Whether an account is using the specific feature(s) that correlate with retention for their segment, not just whether anyone logged in that week. |
| Billing behavior shift | Leading | A downgrade from annual to monthly billing, or seats removed ahead of a renewal date, signals intent well before anyone opens a cancel page. |
| Support ticket pattern | Leading, easy to misread | Volume alone is a weak signal. A spike usually means friction; zero tickets on a complex product often means quiet abandonment, not satisfaction. |
| Power user density | Leading | The ratio of active to paid seats. Shadow seats — paid for but unused — are simultaneously a churn risk and an expansion opportunity, depending which way the gap is moving. |
| Sentiment in support & sales calls | Leading, most underused | Qualitative language often shifts before any usage metric does. Most mid-market teams still don't capture this systematically. |
| NPS / survey score | Lagging | Infrequent, low response rate, and by definition it reflects an experience that already happened rather than one that's coming. |
Signal categories drawn from Gainsight's 2025 Pulse customer success research and ChurnZero's 2025 Customer Revenue Leadership Study.
The pattern across that table is consistent: the signals that actually predict cancellation are the ones that require looking inside product usage and account behavior, not the ones you can pull from a support ticketing tool in an afternoon. That's not an accident — it's why most teams default to the easy version and end up with a score that's descriptive rather than predictive.
Login frequency alone will mislead you
Raw login count is the most commonly tracked signal and one of the least reliable on its own. Two accounts with identical login frequency can be in completely different states: one is quietly satisfied and checks in for five minutes a day to do exactly what they need, the other logs in constantly while evaluating a competitor and is a week from leaving. Frequency without depth can't tell them apart.
Depth is the fix — are they touching the feature that correlates with retention for their plan and use case, or are they only opening the app to check one number and leave? A support-heavy product might see healthy accounts generate a steady trickle of tickets and unhealthy ones generate none, because they've stopped engaging enough to hit friction worth reporting. Any single metric read in isolation will produce false positives. That's why a workable score needs several signals combined, not one dashboard number promoted to a decision.
Building a score without a data team
You don't need machine learning to start — you need a small number of leading signals, sane weights, and thresholds you're willing to act on. A simple weighted point system works as a first pass:
- Usage trend (30 points): full points if session frequency is flat or growing over a trailing 30 days; zero if it's down 40%+ over the same window.
- Core feature adoption (25 points): full points if the account used your product's primary value-driving feature in the last 14 days; zero if not.
- Billing signal (20 points): full points on stable or annual billing; zero if they've downgraded frequency or removed seats in the last 60 days.
- Support pattern (15 points): full points for normal ticket cadence; reduced for an unresolved-ticket streak or a sudden spike.
- Seat utilization (10 points): full points if active seats are 70%+ of paid seats; reduced as that ratio drops.
Score below 50 and flat or declining over two consecutive check-ins: treat as high risk and route to proactive outreach. This is a starting framework, not a formula to copy exactly — the weights should reflect which signals actually correlate with churn in your own historical data once you have enough of it to check. Run the resulting at-risk count through your churn calculator to see what saving even a fraction of that cohort is worth in MRR before you decide how much proactive outreach effort it justifies.
What changes when the score is backed by real tooling
The gap isn't just about which signals you track — it's about whether the score actually triggers action. ChurnZero's 2025 study, surveying customer and post-sales leaders across SaaS and technology businesses, found a clear split based on whether teams used a dedicated customer success platform to run their health scoring and playbooks.
Source: ChurnZero, 2025 Customer Revenue Leadership Study (nearly 800 respondents)
Six points of NRR is a meaningful gap, and it isn't really about the software license. Teams with a unified platform aren't smarter about which signals matter — they're faster at turning a score change into an action, because the playbook fires automatically instead of waiting for someone to notice a spreadsheet cell turned red. A health score that lives in a doc nobody revisits weekly isn't a health score. It's a snapshot you took once.
The score will still be wrong — plan for that
Even a well-built score misses accounts. Some churn for reasons no signal captures — a champion leaves the company, a budget gets cut company-wide, a competitor undercuts on price after a procurement review you never saw coming. Treating a health score as a complete solution is how you end up back at that 73% figure: a score that was accurate on day one and never gets checked against actual outcomes drifts out of sync with reality within a couple of quarters.
Two habits keep a score honest. First, backtest it quarterly — pull the accounts that churned last quarter and check what your score said about them 30 and 60 days out. If it didn't flag most of them, your weights are wrong, not your effort. Second, treat the accounts your score missed as a second line of defense rather than a failure to accept. This is exactly where a cancellation flow earns its keep — it catches the subscribers who never showed up as at-risk in your dashboard and still clicked cancel, asks why in the moment, and routes the specific offer that matches their actual reason, the way we break down in why customers cancel.
The two systems feed each other well. Cancel-reason data tells you which of your health score's leading indicators actually held up against real churn and which ones were noise, so the score gets sharper each cycle instead of staying static. Combine that with the proactive outreach playbook we cover in our broader guide to reducing SaaS churn, and you're intervening at two different points in the funnel instead of relying on either one alone to catch everything — which, per the numbers above, it won't.
None of this requires ripping out whatever health scoring you already have. Start by adding one or two leading indicators — feature adoption depth and billing behavior shift are usually the highest-leverage additions — and check the result against your last two quarters of actual cancellations. If the score would have flagged those accounts weeks earlier, you've found real signal. CancelFlow sits downstream of all of this: when a customer your score didn't catch tries to leave anyway, it captures why, offers the right save, and hands you the exact data you need to make the score better next quarter.
Frequently asked questions
What is a customer health score in SaaS?+
A customer health score is a composite number, usually 0–100 or a red/yellow/green rating, meant to summarize how likely an account is to renew versus churn. It's built by combining signals like product usage, support activity, billing behavior, and survey data into a single weighted score, so a team can sort accounts by risk instead of reviewing each one manually.
What's the difference between a leading and a lagging churn indicator?+
A leading indicator changes before a customer decides to cancel — a declining login trend, a downgrade from annual to monthly billing, or a drop in feature adoption depth. A lagging indicator only moves after something has already gone wrong, like an NPS score or a support ticket spike. Most health scores lean on lagging indicators because they are easier to measure, which is exactly why they arrive too late to act on.
How accurate are churn health scores in practice?+
Not very, on average. ChurnZero's 2025 Customer Revenue Leadership Study, based on a survey of nearly 800 customer and post-sales leaders, found that 73% say their current health score doesn't reliably predict churn. The gap is usually a design problem, not a data problem — scores built mostly from lagging, easy-to-measure signals catch risk after it's already too late to intervene.
Do I need machine learning to build a churn health score?+
No. A weighted, rule-based score using a handful of leading indicators — usage trend, feature adoption depth, billing behavior, support ticket pattern — will outperform no score at all and most naive lagging-indicator scores. Machine learning becomes worth the investment once you have a few hundred labeled churn outcomes to train against and want to move past manually-set weights.
Stop losing subscribers today
One script tag. One function call. A live cancellation flow in under 10 minutes.
Start free trial →