Most SaaS churn is predictable weeks before the cancellation email lands. The problem is that teams either have no health score or one built on vanity signals that flag the wrong accounts. This guide shows you how to build a customer health score that actually predicts churn, which signals to include, how to weight them, and how to turn a number into action.
Why generic health scores fail
A health score is a compressed prediction of whether an account will renew. It fails when it measures activity that does not relate to value. Total logins is the classic trap: a user forced to log in daily to fix problems looks ‘healthy’ while hating the product. Meanwhile an efficient power user who logs in weekly looks ‘at risk.’
The fix is to anchor every signal to a hypothesis about value. Ask: does more of this signal genuinely mean the customer is getting more outcome from us? If you cannot answer yes with a reason, drop the signal.
Choose signals across three categories
Product engagement
Measure use of the features that map to your core value, not raw activity. If your product’s value is publishing reports, track reports published and shared, not page views. Depth of adoption (how many key features used) predicts retention better than frequency alone.
Account and relationship
Include signals like number of active seats versus purchased seats, presence of an executive sponsor, and support ticket sentiment. A single-threaded account with one champion is fragile: if that person leaves, the renewal is at risk regardless of usage.
Commercial and behavioral
Track late payments, downgrade requests, and contraction in seats. These are strong, direct signals. A downgrade request is not a warning of future risk; it is risk already in progress.
Turning signals into a score
Start simple. Assign each signal a score band and a weight, then combine them. Resist the urge to build a machine learning model on day one; you rarely have enough clean churn history, and a transparent weighted model is easier for a CSM to trust and act on.
| Signal | Healthy | At risk | Weight |
| Core feature adoption | 3+ key features used weekly | 1 or fewer | High |
| Seat activation | Over 70% of seats active | Under 30% | High |
| Executive sponsor | Identified and engaged | None or left | Medium |
| Support sentiment | Neutral to positive | Repeated escalations | Medium |
| Billing | On time | Late or downgrade asked | High |
Combine into three tiers: green, yellow, red. The tier matters more than the exact number, because it drives a clear response.
A real scenario: the quiet enterprise account
A B2B tool had an enterprise customer whose overall login count looked fine, so no alarm fired. But two signals had quietly moved: seat activation dropped from 80% to 25% over a quarter, and their only champion changed jobs. A weighted score caught what login count hid. The CSM re-onboarded new users and secured a second stakeholder. The account renewed. Under a login-only view, it would have looked safe until the cancellation.
Common mistakes and how to fix them
- Confusing activity with value. Fix: tie every signal to an outcome the customer is paying for.
- One score for all customers. A 5-seat startup and a 500-seat enterprise churn for different reasons. Fix: segment scores by customer type or size.
- Set and forget. A score that never gets validated drifts. Fix: quarterly, compare scores against who actually churned and adjust weights.
- No action attached. A red score that no one owns is a dashboard, not a system. Fix: define a required play for each tier and assign an owner.
- Over-engineering early. Fix: ship a transparent weighted model first; add sophistication only when you have proven history.
Action checklist
- List the outcomes customers buy your product to achieve.
- Pick 4 to 6 signals that map directly to those outcomes.
- Define healthy and at-risk bands for each, with weights.
- Roll signals into green, yellow, red tiers.
- Attach a specific play and owner to each tier.
- Review predictions against real churn every quarter and recalibrate.
Conclusion and next step
A useful health score is honest about value, segmented, and tied to action. Do not wait for a perfect model. Pick your top signals this week, publish a simple weighted score, and give each red account an owner. Refine it once you can measure it against real outcomes.
FAQ
How many signals should a health score use?
Fewer than most teams expect. Four to six well-chosen signals are usually enough. Too many dilutes the score and makes it hard to explain why an account is red.
Should I use machine learning for churn prediction?
Only once you have enough clean churn history and a transparent baseline to compare against. Start with a weighted rules model that CSMs can understand and trust.
How often should the score update?
Frequently enough to catch trends, typically daily or weekly for the score itself, with a quarterly review of the weights and thresholds.
What is the single most predictive signal?
It varies by product, but seat or license activation and direct commercial signals like downgrade requests tend to be strong across many B2B SaaS businesses.
References
- Established customer success practice on health scoring and net revenue retention, as discussed widely in the B2B SaaS field.