A customer health score is the most important metric in customer success — and the most commonly built incorrectly. When done well, a health score gives CS teams an early warning system that identifies at-risk accounts weeks or months before the customer expresses dissatisfaction. When done poorly, it gives CS teams a false sense of security about accounts that are quietly heading toward churn.
89% of CS teams that use a formal health score report improved ability to predict churn — but only 31% say their health score is actually accurate Source: Gainsight, State of Customer Success, 2024
The gap between having a health score and having an accurate one is significant. This article covers the methodology for building a health score that actually predicts churn — including the signals that matter, the ones that don't, and the weighting framework that makes the score actionable.
Why Most Health Scores Fail
The most common failure mode in health score design is including too many signals without validating which ones actually predict churn. A health score that averages 15 different metrics — login frequency, feature adoption, support tickets, NPS, contract value, days since last contact, and so on — is not more accurate than one that uses 5 well-validated signals. It's less accurate, because the noise from irrelevant signals dilutes the signal from relevant ones. The second failure mode is using lagging indicators instead of leading ones. NPS scores, for example, are a lagging indicator: they tell you how the customer feels after the fact. Login frequency is a leading indicator: it tells you what the customer is doing before they tell you how they feel. A health score built primarily on lagging indicators will always be too late.
The Signals That Actually Predict Churn
Research from Gainsight's analysis of over 10,000 B2B SaaS accounts identifies the signals most strongly correlated with churn. The top predictors, in order of predictive power:
- Product usage frequency: Accounts that drop below their baseline usage frequency for two consecutive weeks are 4.2× more likely to churn than those maintaining baseline.
- Feature adoption breadth: Accounts using fewer than 30% of the features they purchased are 3.1× more likely to churn — they haven't achieved the value they bought.
- Champion engagement: When the primary champion stops attending check-in calls or responding to emails, churn probability increases by 67%.
- Support ticket sentiment: Accounts submitting tickets with frustration-indicating language ('still not working', 'this is unacceptable') are 2.8× more likely to churn within 90 days.
- Time since last executive contact: Accounts where CS hasn't spoken to an executive-level contact in over 90 days are 2.3× more likely to churn at renewal.
4.2× higher churn probability for accounts that drop below baseline usage frequency for two consecutive weeks Source: Gainsight, Churn Prediction Research, 2024
The Signals That Don't Predict Churn (As Much as You Think)
Several commonly used health score inputs have weaker predictive power than most CS teams assume. NPS scores, for example, have a correlation with churn that is real but weaker than usage data — and they're collected infrequently enough to miss the early warning window. Contract value is almost entirely uncorrelated with churn probability (large accounts churn too). And the number of support tickets submitted is a weak predictor — what matters is the type and sentiment of the tickets, not the volume.
Building the Score: A Practical Framework
Step 1: Define Your Churn Cohort
Before building a health score, analyze your historical churn data. Look at the 20–30 accounts that churned in the last 12 months and identify the signals that were present 60, 90, and 120 days before churn. This analysis is the foundation of a validated health score — it tells you which signals actually predicted churn in your specific customer base, rather than which signals are theoretically correlated with churn in general.
Step 2: Select 4–6 Leading Indicators
Based on your churn cohort analysis, select the 4–6 signals that had the strongest predictive power. Prioritize leading indicators (usage data, engagement data) over lagging ones (NPS, CSAT). Ensure each signal is measurable, consistent, and available in near real-time.
Step 3: Weight by Predictive Power
Assign weights to each signal based on its predictive power in your churn cohort analysis. A signal that was present in 80% of churned accounts should have a higher weight than one that was present in 40%. The weights should sum to 100 and should be revisited quarterly as you accumulate more churn data.
Step 4: Define the Thresholds
For each signal, define the threshold that triggers a health score change. What level of login frequency is healthy vs. at-risk? What feature adoption percentage is healthy vs. concerning? These thresholds should be based on your churn cohort data — the point at which the signal became predictive of churn in historical accounts.
Step 5: Validate and Iterate
A health score is a hypothesis, not a fact. Validate it by tracking whether accounts that score as 'at-risk' actually churn at higher rates than those that score as 'healthy'. If the score is not predictive, adjust the signals and weights. Most health scores require 2–3 iterations before they become reliably accurate.
Making the Score Actionable
A health score that sits in a dashboard and is reviewed once a month is not a health score — it's a report. A health score that triggers automatic alerts, CSM interventions, and escalation workflows is a health score. The most effective CS teams build their health score directly into their workflow: when an account drops below a threshold, the CSM gets an alert, a task is created, and the intervention playbook is triggered automatically.
34% lower churn rate for CS teams that use automated health score alerts vs. those that review health scores manually Source: ChurnZero, Customer Success Automation Report, 2024
The health score's ultimate value is not in the number — it's in the conversations it triggers. A CSM who reaches out to an at-risk account because their health score dropped is having a proactive conversation. A CSM who reaches out because the customer called to cancel is having a reactive one. The difference in outcome between those two conversations is the entire ROI of a well-built health score.