How to Check Whether a Customer Health Score Is Actually Good

Here is how I check whether a customer health score is actually good. A good customer health score is one whose bands separate the accounts that churn from the accounts that renew. There is no universal "good" number: you benchmark a health score by comparing churn rates across its bands, and if the lowest band churns several times more often than the highest band (say 21% vs 3%), the score works.

To rebuild past scores for this test, use the free health score calculator, then measure each band with the churn rate calculator. This is Part 5 of the Customer Health Score series. It builds on the formula, the alerts and my retention analytics playbook from Part 4.

Why is an average health score misleading?

Two teams can both report an average health score of 70 and be in very different shape. In one, most accounts sit close to 70. In the other, there is a cluster near 95 and another near 40, and the average hides a lot of risk.

Three things matter more than the average:

• Distribution: how accounts spread across the range.
• Link to real outcomes: churn, contraction, expansion.
• Trend: a stable 68 is not the same as a 68 that was 82 last quarter.

On one team I worked with, the average score went up a few points each quarter and leadership was pleased. Churn did not move. When we looked closer, a new usage signal had been added that almost every account scored well on. The customers had not improved. The formula had drifted.

How do I benchmark a customer health score?

Band churn rate  = accounts in the band that churned ÷ accounts in the band at the start
Separation ratio = churn rate of the lowest band ÷ churn rate of the highest band
Coverage         = churned accounts that were in a risk band ÷ all churned accounts

Band             = a score range (for example Red 0–49)
At the start     = scores frozen on the first day of the period, before outcomes are known

Freeze the scores at the start of the period. If you use scores from the day of churn, the score will look predictive only because it already saw the cancellation.

What does a predictive score look like? A 500-account example

Here is an illustrative mid-market B2B book of 500 accounts, scored at the start of the year, with churn counted at year end.

Band	Accounts	Share	Churned	Band churn rate
Red (0–49)	70	14%	15	21.4%
Yellow (50–74)	225	45%	16	7.1%
Green (75–100)	205	41%	6	2.9%
Total	500	100%	37	7.4%

Separation ratio = 21.4% ÷ 2.9% ≈ 7.4
Coverage (Red only)         = 15 ÷ 37 ≈ 41%
Coverage (Red + Yellow)     = 31 ÷ 37 ≈ 84%

Red holds 14% of accounts but 41% of the churn. A separation ratio around 7 means the score sorts accounts well. If Red and Green churned at similar rates, the ratio would be close to 1 and I would fix the formula before anyone relied on the number. Run your own bands through the churn rate calculator, and adjust weights in the health score calculator if they do not separate.

How should I set health score bands?

Do not pick round cutoffs because they look tidy. Anchor bands to churn and expansion data from the last two to four quarters. The ranges below are illustrative, not industry benchmarks:

Band	Score range	Example churn rate	Recommended action
Red	0–49	15–25%	Executive escalation, save plan
Yellow	50–74	5–10%	CSM check-in, adoption push
Green	75–89	2–4%	Standard cadence, upsell watch
Champion	90–100	Under 2%	Advocacy, case study, referral ask

Teams often skip the Champion tier. Accounts above 90 are more than safe. They are your best source of references and expansion, and treating them like a 76 wastes that.

Moving a cutoff is a trade-off. Raise the Red line from 50 to 55 and coverage goes up, but more healthy accounts get flagged too. This is the precision and recall trade-off described in Google's machine learning course, and the same reasoning applies to a health score.

How often should health score benchmarks be updated?

A benchmark set with 50 customers will not hold at 500.

• Re-check the link to churn every quarter, not once a year.
• Watch for score inflation. If the average rises but churn does not improve, the formula is drifting, not the customers.
• Segment once you have enough volume. A 65 can be excellent for enterprise and weak for SMB.

I track the benchmark itself on the monthly dashboard:

Metric	Target	This month	Last month	Change
Average health score	Stable	68	66	+2
Accounts in Red	<15%	14%	16%	−2 pts
Separation ratio (Red ÷ Green churn)	5+	7.4	6.8	+0.6
Coverage (Red + Yellow)	80%+	84%	79%	+5 pts

What I've seen in practice

• Regions needed their own bands. Working with partners across MENA and Europe, I found that some markets logged in less and preferred a phone call to the platform. A low engagement score there often reflected a working style, not risk. Separate bands per region fixed a run of false alarms.
• Rising averages hid a broken formula. The quarter our average rose fastest was the quarter the score predicted least. Since then, I never report the average without the separation ratio next to it.
• Small Red bands can still be the whole story. In books I have worked on, a small share of accounts carried most of the churn. That is the pattern you want to see, as long as the Red band is small enough to work through each week.
• Too few churned accounts mislead. With fewer than 15 or 20 churned accounts in a period, band rates swing wildly. I pool two to four quarters before drawing conclusions.

This week, export last year's starting scores and outcomes, group accounts into your current bands, and calculate churn for each band. If the bands do not separate clearly, revisit your weights before you change anything else.

Frequently asked questions

What is a good customer health score?

There is no universal number. A score is good when its bands predict outcomes: accounts in the lowest band should churn several times more often than accounts in the highest band. If Red and Green accounts churn at similar rates, the formula needs work.

How should I set health score bands?

Anchor them to your own churn and expansion data from the last two to four quarters rather than picking round numbers. A common structure is Red (0–49), Yellow (50–74), Green (75–89) and Champion (90–100), each with its own recommended action.

How do you test whether a health score predicts churn?

Freeze scores at the start of a period, group accounts into bands, and compare each band's churn rate at the end. Divide the lowest band's churn rate by the highest band's: a ratio well above 1 means the score separates risk.

How often should health score benchmarks be updated?

Re-check the link with churn every quarter, watch for score inflation, and segment benchmarks by customer size, region or industry once you have enough accounts.

Sources

• Classification: accuracy, recall, precision, Google Machine Learning Crash Course
• Classification: ROC and AUC, Google Machine Learning Crash Course
• Essential Considerations For Your First Customer Scorecard, Gainsight