Guide
Customer health scores that predict churn — and how to test yours.
Most health scores are built from what was easy to query, tuned by opinion, and never tested against reality. Then everyone acts surprised when green accounts churn. Here's what the scores that actually work do differently — and how to check yours.
Four principles of scores that work.
Build from outcomes, not opinions
Start with your churned accounts and work backward: what was true about them 90, 60, and 30 days before they left? The public designs worth studying — GitLab's and PostHog's handbooks both publish their thinking — share this trait: the score is derived from history, not assembled from available metrics.
Weight people signals, not just product signals
Logins and ticket volumes are easy to query, so they dominate most scores. But churn post-mortems keep naming the same culprits: champion disengagement, sentiment shifts, sponsor changes. If your score has no people signals, it's measuring the product's health, not the account's.
Back-test before anyone trusts it
Run the design against 12+ months of your renewal history. What would it have flagged, when, and what would it have missed? A score that can't predict your past churn has no business predicting your future churn. This step is skipped almost everywhere — it's also the only step that earns trust.
Ship reasons, not just a number
A CSM who sees '42' learns nothing. A CSM who sees '42 — champion quiet 47 days, sentiment down on last 6 tickets, two power users deactivated' knows exactly what Monday looks like. Every score should arrive with its evidence attached.
The 30-second test.
Before redesigning anything, grade what you have. The back-tester needs three columns — account, score, churned or active — and answers the only question that matters: would this score have caught your churn?
- Catch rate and false-alarm rate at any threshold you choose
- Works with 0–100 scores or red/yellow/green
- 100% client-side — your data never leaves the browser
Questions, answered honestly.
How do I test my current score right now?
Use the free Health Score Back-tester: paste a CSV of accounts, scores, and churn outcomes, and it computes your catch rate and false-alarm rate at any threshold — entirely in your browser, nothing uploaded. Most teams get an uncomfortable answer in under a minute.
What catch rate should we aim for?
In a 2026 poll of CS leaders, 82% said their health score caught less than half of churn — so anything above 50% already beats most teams. Above 75% is strong. But the number matters less than the trend: is the score getting better as you learn, or has it been the same untested formula for two years?
Should we build the score in our CS platform or our warehouse?
Wherever the data already lives and the CSMs already look. The platform vs warehouse debate is secondary — what matters is that the score is computed from joined sources (CRM + tickets + calls + product), not from whichever single system was easiest.