Building a Customer Health Score That Actually Predicts Churn

Why Your Current Health Score Isn't Predicting Churn

Your customer success team probably tracks something called a "health score." Maybe it's a simple formula: login frequency plus feature adoption plus support ticket volume. You assign each customer a number from 1-100, flag the low ones as at-risk, and hope your team reaches out in time.

Here's the problem: this approach feels safe and measurable. But it rarely predicts who actually churns.

We looked at data from hundreds of SaaS companies. The customers with the lowest engagement scores weren't always the ones who left. Sometimes dormant accounts renewed. Meanwhile, customers with healthy scores churned without warning.

The issue is that most health score systems measure activity, not intent. Activity tells you what customers are doing. Intent tells you why they might leave.

Building a real churn scoring system requires you to think differently about what data matters.

The Three Components of a Real Customer Health Score

A customer health score that actually predicts churn combines three distinct layers of information. Most teams only track one.

1. Usage Pattern Changes (The Signal)

Don't measure absolute usage levels. Measure changes in usage patterns.

A customer who uses your product three times a week consistently? That's stable. The same customer who suddenly drops to once a week? That's a signal.

This distinction matters because what's "normal" varies wildly. A marketing automation platform might have monthly batch users who log in once a month. That's healthy for their workflow. You can't flag them as at-risk based on frequency alone.

Instead, track velocity. Look at:

• Week-over-week change in login frequency
• Month-over-month active feature usage (not just logins)
• Time spent in key workflows compared to their baseline
• Whether they're using new features or staying in legacy workflows

A 40% drop in logins from a customer's own baseline is more predictive than any absolute number.

2. Business Context (The Reality Check)

Your customer's industry, company size, and use case determine what healthy looks like.

An enterprise customer in financial services might interact with your product differently than a startup in e-commerce. Both could be perfectly healthy. You need to segment your churn scoring by customer cohort.

Collect and weight:

• Industry vertical (some verticals have higher baseline churn)
• Company size and contract value (larger deals churn differently)
• Use case and feature adoption (feature X might be crucial for healthcare, irrelevant for SaaS tools)
• Tenure (month 3 customers behave differently than month 36 customers)
• Time until renewal (the 60 days before renewal is different from month 6 of a 12-month contract)

This context transforms your churn scoring from one-size-fits-all to realistic.

3. Engagement Triggers (The Warning Signs)

Certain behaviors are direct warnings that a customer might leave.

We're not talking about declining usage here - we're talking about specific actions or inactions that correlate with churn decisions:

• Key stakeholder stops logging in (especially if they're the main user)
• Support tickets spike or stop entirely (silence can be worse than complaints)
• Decreased adoption of new features you've released
• Expansion purchase attempts fail or get cancelled
• They request billing information or contract terms
• Reduced interaction with your customer success manager (if you have one)

Each of these is more predictive than raw engagement numbers.

Building Your Churn Scoring Model

Now that you understand what matters, here's how to actually build this.

Step 1: Decide On Your Scoring Range

Pick a scale that your team will actually use. 0-100 is standard because humans intuitively understand percentiles. We recommend this breakdown:

• 80-100: Healthy. Stable or growing usage.
• 60-79: At-risk. Usage declining or engagement triggers present.
• 40-59: High-risk. Multiple warning signs. Immediate action needed.
• 0-39: Critical. Customer is likely churning. Emergency outreach only.

Your team should have different action triggers for each band.

Step 2: Weight Your Signals

Not all signals matter equally. You need to weight them based on what actually predicts churn at your company.

Here's a realistic example from an actual SaaS company with $2M ARR and mostly mid-market customers:

• Usage pattern change (40% of score) - This was the strongest predictor
• Support sentiment (25%) - What customers say matters
• Feature adoption (20%) - Using new features = staying for new value
• Expansion behavior (15%) - Customers who grow with you rarely churn

Your weights will be different. You need to analyze your own data.

Step 3: Start Simple, Then Iterate

Don't try to build the perfect model. Build a 1.0 version with just your top three signals. Use it for two months. Track which customers it correctly identified as at-risk.

Your goal is to find customers with a 40%+ probability of churning within 90 days. If your model gets 70% accuracy in month 2, great. Iterate from there.

Most teams either build nothing or build something so complex that no one uses it. Don't be either of those teams.

The Data You Actually Need to Collect

Here's what your system needs to track to make customer health score work:

• Daily or weekly login data (not aggregated monthly data)
• Feature-level usage, not just account-level (which features, how often)
• When customers use your product (time of day, day of week - behavioral patterns matter)
• Support interactions and sentiment (tickets opened, closed, response times)
• Billing events (usage-based charges up or down, expansion attempts)
• Email engagement (do they open your emails, click links)
• NPS or CSAT if you collect it
• CSM notes and interaction logs (if you have a success team)

You probably have most of this data in your systems right now. The issue is connecting it.

How to Actually Use This Score

A customer health score is only useful if it changes behavior.

Your team needs specific workflows:

For customers scoring 60-79: Your CSM reaches out with a genuine check-in. Ask what's changed. Sometimes the usage drop is intentional (they adopted a workflow that required less engagement). Sometimes they're exploring competitors. You find out.

For customers scoring 40-59: This is a business review meeting. You want to understand their goals, see if they're getting value, and often uncover a problem you can fix. Many of these customers can be recovered.

For customers scoring below 40: This is a save meeting. Be direct. Ask if they're planning to renew. Listen to why they might not. Fix what you can.

The score is a tool to decide who you talk to and what you talk about. It's not a replacement for actual conversation.

Common Mistakes That Tank Churn Scoring

As you build this out, avoid these pitfalls:

Using only login frequency. You'll flag hibernating accounts that are actually fine, and miss customers who log in every week but hate your product.

Not accounting for seasonal patterns. If you sell to schools, summer usage drops aren't warning signs. You need to adjust for your customer's business calendar.

Making the model too complex. If your score requires 47 data inputs, no one will trust it. Keep it simple enough that a human can understand why a customer got that score.

Ignoring the contract timeline. Every customer should be more important as their renewal date approaches. A low score in month 11 of a 12-month contract is an emergency. A low score in month 2 might be normal onboarding behavior.

Not updating based on outcomes. Every 60-90 days, look back at the customers your model flagged. How many actually churned? What did you miss? Adjust your weights.

Benchmarking Your Health Score

If you want to compare your approach to industry standards, here's what we see from the Churn Analyzer blog research:

Companies with a formal customer health score system have 20-30% lower churn than those without. But only if the score actually predicts churn - which means it needs to be tested and refined.

The median time from health score warning to churn is 40 days. That means if your score flags someone, you have roughly five weeks to take action.

Teams that use a health score primarily for reactive saves tend to see 35-40% recovery rates on high-risk customers. Teams that use it proactively (reaching out to 60-79 band customers before they're critical) see 55-65% retention lift.

Next Steps

Start this week.

First, audit what data you have. What can you pull from your product, billing system, and email platform today? You probably have 80% of what you need.

Second, pick your three most important signals. Don't overthink it. Usage change, support sentiment, and feature adoption work for most SaaS companies.

Third, calculate a rough score for your current customers. See who lands in the at-risk band. Manually review a few of them. Does it feel right? Adjust if needed.

Building an effective churn scoring system doesn't require a data science team. It requires intentionality about what predicts churn at your company, and the discipline to test and iterate.

If you want to skip the infrastructure work, start a free trial of Churn Analyzer to see how automated customer health scoring can identify at-risk customers before they leave. But whether you build this manually or use a tool, the framework here is what separates effective customer health scores from vanity metrics.