Customer Health Score: How to Build One That Predicts Churn

What a Customer Health Score Is

A customer health score is a single number, typically on a 0–100 scale, assigned to each customer account and updated on a regular cadence (usually weekly). The number summarizes how likely that account is to renew, expand, or cancel based on observable behavioral and relationship signals.

The health score is a prediction tool, not a diagnostic one. It answers the question: "Which accounts should my customer success team contact this week?" It does not answer the question: "Why are customers churning at the rate they are?" Those are different questions requiring different tools.

This distinction matters because the two concepts are sometimes conflated. A customer health score is a forward-looking, per-account indicator. The Churn Health Score that RetentionCheck produces is a backward-looking, portfolio-level diagnostic based on AI analysis of cancellation feedback. Both are useful. They answer fundamentally different questions about your retention situation.

For a calculator to estimate how much churn is currently costing you in revenue, see the churn rate calculator.

The 5 Signals That Matter

Most customer health score frameworks list 8–15 signals. In practice, five categories account for the majority of predictive power. Adding more signals beyond these five produces marginal accuracy improvements at significant implementation cost. For most SaaS companies under $20M ARR, a model built on these five signals is more than sufficient.

1. Login Frequency

Login frequency is the most universally available and consistently predictive churn signal. A customer who was logging in daily and has gone 10 days without a login is exhibiting the behavioral signature of pre-churn disengagement. ProfitWell's analysis of 23,000 B2B SaaS cancellations found that 68% of churned customers had a measurable login frequency decline in the 30–60 days before cancellation.

Measure login frequency as a rolling 14-day average compared to the customer's own historical baseline, not against a fixed threshold. A customer who logs in twice a week and drops to twice a month has declined sharply. A customer who always logged in once a week and continues at that rate is healthy. Absolute frequency matters less than trend relative to baseline.

2. Feature Adoption Breadth

Customers who use more of your product's core features churn at lower rates than customers who use a narrow subset. This is true across product categories but is especially pronounced in B2B tools: accounts using 4 or more core features churn at roughly half the rate of accounts using 1–2 features, according to data from Gainsight's 2025 Customer Success Benchmark Report.

Define "core features" as the 5–8 capabilities that deliver your product's primary value proposition. Exclude administrative features (billing management, user settings) and onboarding steps. Track what percentage of core features each account has used in the past 30 days. Weight recent usage more heavily than historical adoption.

3. Support Ticket Volume and Resolution

Support tickets are a double-edged signal. A customer with zero support tickets is not necessarily healthy. Low engagement can indicate the customer is not using the product. A customer who opened and resolved a support ticket typically shows higher retention than a customer with no support history at all, because ticket resolution confirms active engagement and problem-solving.

The churn-predictive signal is unresolved support tickets, not ticket volume. Accounts with one or more open tickets older than 7 days churn at 2.3× the rate of accounts with fully resolved support histories. Weight this signal heavily in your health score model. A single lingering support issue can dominate the customer's perception of your product regardless of everything else going well.

4. Billing Issues

Billing issues are both a churn signal and a churn cause. A customer whose payment method has failed, whose invoice has been disputed, or whose usage has breached a plan limit is under friction that often triggers a cancellation evaluation. Payment-related signals typically provide 14–45 days of lead time before cancellation, making them one of the most actionable categories.

Segment billing signals by type. Failed payment attempts are involuntary churn precursors and should trigger immediate automated dunning, separate from your health score response protocol. Plan limit breaches can signal either growth (positive) or frustration (negative) depending on whether they correlate with feature adoption or with customer complaints. Disputed invoices are almost always negative signals and should trigger CS outreach within 48 hours.

5. NPS and CSAT Scores

Survey-based signals (NPS, CSAT, in-product thumbs ratings) are the most direct expression of customer sentiment but the least frequently updated. They typically provide 45–90 days of lead time before cancellation, making them the longest-lead but least precise signal in the model.

NPS detractor scores (0–6) predict churn at roughly 4× the rate of promoter scores (9–10) in B2B SaaS. CSAT scores below 3 out of 5 on core feature interactions correlate with 2–3× higher churn rates in the 90 days following the survey response. Include the most recent survey score in your health score model, but weight it below the behavioral signals, which update continuously and reflect current engagement rather than a point-in-time opinion.

Scoring Methodology: Building a Weighted Composite

A weighted composite health score aggregates each signal into a 0–100 number using predefined weights that reflect each signal's predictive importance. The mechanics are straightforward.

First, normalize each signal to a 0–100 scale within its category. Login frequency might score 100 for a customer at their historical baseline and 0 for a customer who has been inactive for 30+ days. Feature adoption might score 100 for an account using 6 of 7 core features and 0 for one using only 1.

Second, assign weights based on predictive importance. A starting point for B2B SaaS:

Third, compute the composite: multiply each signal score by its weight and sum the results. An account scoring 90 on login frequency, 70 on feature adoption, 100 on support, 100 on billing, and 60 on NPS would compute as: (90 × 0.30) + (70 × 0.25) + (100 × 0.20) + (100 × 0.15) + (60 × 0.10) = 27 + 17.5 + 20 + 15 + 6 = 85.5.

Fourth, establish intervention tiers based on the composite score:

• Green (70–100): Healthy. Monitor weekly. Look for expansion opportunities.
• Yellow (40–69): At-risk. Proactive CS outreach within 5 business days. Reference the specific signals driving the decline.
• Red (0–39): High-risk. Immediate CS contact or executive escalation for high-value accounts. Do not wait for the next weekly review cycle.

How to Validate and Calibrate Your Model

A health score model is only as good as its correlation with actual outcomes. After building your initial model, validate it against your historical churn data. Take the last 6–12 months of accounts that churned and compute what their health score would have been 30, 60, and 90 days before cancellation.

If 70%+ of churned accounts would have shown a Yellow or Red score 30 days before cancellation, your model has meaningful predictive power. If the distribution looks flat — roughly equal numbers of churned accounts across Green, Yellow, and Red — your weights need recalibration. Adjust weights toward the signals that diverge most sharply between churned and retained cohorts in your historical data.

Recalibrate annually at minimum. Product changes, new customer segments, and pricing adjustments all shift the relative predictive power of each signal. A weight configuration that was accurate for a self-serve SMB customer base may perform poorly after you add an enterprise tier with different usage patterns.

Customer Health Score vs. Churn Health Score: Key Differences

These two concepts are commonly confused because they share vocabulary. They are different tools that answer different questions.

A mature retention program uses both. The customer health score tells your CS team who to call this week. The Churn Health Score tells your product team what to build next quarter. Neither tool is a substitute for the other.

For the diagnostic approach — understanding the systemic reasons customers are churning across your entire customer base — paste your cancellation feedback into RetentionCheck. The AI analysis identifies dominant churn themes, weights them by severity, and produces a portfolio-level Churn Health Score.

Implementation Without a Data Science Team

Building a customer health score does not require machine learning or a dedicated data science function. A spreadsheet-based model that pulls from your product database, CRM, and support tool can produce meaningful scores for companies at any stage.

The practical minimum to build a useful health score:

• Product usage data: login timestamps and feature interaction events (available from Mixpanel, Amplitude, Segment, or direct database query)
• Support ticket status: open/closed/resolved timestamps (available from Intercom, Zendesk, or Help Scout APIs)
• Billing status: payment method status, failed charge history (available from Stripe or your billing provider)
• Survey data: most recent NPS or CSAT score per account (available from your survey tool's export)

Join these data sources on account ID in a spreadsheet or simple data warehouse table. Compute normalized scores for each signal category and apply your weights. Export a sorted list of accounts by composite score each Monday. Accounts in Red receive CS contact that week. This workflow takes 2–4 hours to set up and 30 minutes per week to maintain.

For a framework on identifying which customers are at risk based on cohort analysis, see cohort retention analysis. For the financial model that quantifies the revenue impact of improving retention, see net revenue retention.

When to Use a Machine Learning Model Instead

Manual weighted health scores outperform reactive retention for most companies. Machine learning models provide incrementally better accuracy but require significantly more infrastructure and labeled training data.

Switch to an ML-based churn prediction model when: you have 500+ monthly churned customers (enough labeled examples to train reliably), you have a data engineering function that can maintain model pipelines, and your manual health score false-positive rate (Yellow accounts that do not churn) exceeds 40%, causing CS team alert fatigue. Below these thresholds, the complexity cost of ML outweighs the precision benefit.

For a full treatment of predictive approaches including logistic regression and gradient boosting, see how to predict customer churn.