The 8 Signals That Belong in a Customer Health Score in 2026

Almost every published health score model includes support as a signal, and almost every one of them counts tickets. Contact volume as a proxy for health is close to backwards: engaged power users file more tickets than quiet accounts, and the account that churns without warning is typically the one that stopped writing in months ago. The signal is not how much a customer contacted you. It is what they said.

Eight signals belong in a customer health score: usage trend, adoption breadth across the buying team, what support contacts were about, sentiment trajectory, unresolved asks, champion movement, commercial signals, and competitor mentions. Usage-based models cover the first two well. Signals three through eight require reading text, which is why most health scores are blind to them.

Usage tells you they logged in. It does not tell you they are staying.

The strongest finding in the health-scoring literature is that trend beats level: an account whose usage dropped 40% this month is usually a bigger risk than a steady low-usage account. That is correct and it is also the ceiling of what telemetry can do. Usage tells you behavior. It cannot tell you intent, and renewal decisions are made on intent.

The accounts that churn with a healthy score are almost always accounts where the words went bad before the numbers did. A frustrated admin keeps logging in right up until the day the contract ends.

What a health score has to do

1. Model the account, not the user. One disengaged user is noise. Three inactive admins is a signal. Every component should roll up at the account level with a stated rule for how user-level activity aggregates.
2. Read support content, not support volume. This is the difference between a score that predicts and one that correlates. An adaptive taxonomy categorizes what an account is contacting about, so a score can distinguish an account filing tickets about advanced configuration from one filing tickets about the same broken export for the fourth month.
3. Weight by what the account is worth and where it is. A customer context graph ties feedback to segment, revenue, lifecycle stage, and renewal date, which is what turns a score into a work queue rather than a leaderboard.

The 8 signals that belong in a customer health score

1. Usage trend against the account's own baseline

Percentage change over the trailing period against that account's normal, not against a global benchmark. Segment-relative percentile works better than absolute thresholds, because a low-touch product and a daily-use product have different healthy patterns.

2. Adoption breadth across the buying team

How many distinct users, how many distinct teams, and whether the admins are among them. Single-user dependency is one of the most reliable churn predictors in B2B, because the account leaves when that person does. Track seat utilization against contracted seats as the commercial version of the same signal.

3. What support contacts were about

Not the count. The categories. Contacts about advanced use are an expansion signal. Contacts about a recurring unresolved defect are a churn signal. Contacts about basics from an account 18 months in are an enablement failure. These three produce identical ticket counts and opposite health readings, which is why volume-based support components add noise rather than predictive power.

4. Sentiment trajectory across every channel

Direction over time, not a point-in-time score. NPS gives you a number twice a year from whoever responded. Sentiment pulled from tickets, calls, reviews, and community gives you a slope from everyone who said anything. The slope is the signal. See going beyond CSAT scores to understand customer sentiment.

5. Unresolved asks

The requests this account has made that you have not shipped, weighted by how long they have been open and how often they have been repeated. An account that asked for the same thing in three consecutive quarterly calls has a documented reason to leave, and it is usually absent from the health model entirely. See feedback signals that indicate customer churn risk.

6. Champion movement

Whether the people who ran the evaluation are still present and still engaged. A champion going quiet, a new decision-maker appearing in threads, or the original buyer's email bouncing are all leading indicators that arrive well before usage moves.

7. Commercial signals

Invoice status, seat utilization, contract term remaining, and the history of the last renewal. These are lagging relative to the others, but they belong in the score as hard overrides rather than weighted components: an account in billing dispute should be capped regardless of how good its usage looks.

8. Competitor mentions from the account itself

Anyone at the account naming an alternative in a ticket, a call, or a public review. This is the highest-specificity signal on the list and the rarest in published models, because it requires reading text from channels the CS platform does not ingest.

The score that predicts is the one that reads

Two companies build health scores. Both include usage trend, adoption breadth, NPS, support tickets, and billing. One counts support tickets. The other reads them. Both scores will correlate with churn in a backtest, because usage decline eventually shows up in every account that leaves.

The difference is lead time. Usage decline is the last thing to happen before a churn, not the first. By the time an account's logins fall, the decision was made in a meeting you were not in, usually weeks earlier, and the reasons were stated out loud in a support thread or a call that nobody scored.

That is the whole argument for text-based components. They do not make the score more accurate in a backtest. They make it earlier, and earliness is the only property of a health score that has any value, because a score that identifies churn you cannot prevent is a report rather than an early-warning system. See churn risk detection from support data and correlating CSAT with churn risk.

How to build it

Normalize each signal to a 0 to 100 subscore before weighting, or one large raw number swamps everything else. Add hard overrides for critical events rather than letting them average out. Exempt accounts under 90 days from the usage-trend rule, since new accounts have no baseline.

Start with signals one, two, three, and seven. That combination is buildable and it already beats most published models, because signal three is the one usually missing. Add four and five when you can categorize feedback across channels. Add six and eight last.

Backtest against your own churn history before rollout, refresh the weights quarterly, and check false positives as carefully as false negatives. A score that flags everything is the same as no score.

The decision rule: if a signal cannot change what a CSM does this week, it does not belong in the score.

FAQ

What signals should a customer health score include?

Usage trend against the account's own baseline, adoption breadth across the buying team, the categories of what support contacts were about, sentiment trajectory across channels, unresolved requests, champion movement, commercial signals as hard overrides, and competitor mentions from the account itself.

Why is support ticket volume a bad health signal?

Because engaged power users file more tickets than disengaged accounts, so volume alone points in the wrong direction as often as the right one. What predicts is the content: contacts about advanced configuration mean something opposite to contacts about the same unresolved defect for the fourth month, and both produce the same count.

How often should a customer health score be recalculated?

Daily or weekly for the score itself, with the weights reviewed quarterly against actual churn outcomes. Scores refreshed monthly lose the lead time that justifies having one, and weights left unreviewed drift away from what currently predicts as the product and customer base change.

How does Enterpret contribute to a customer health score?

Enterpret categorizes every piece of feedback an account produces across tickets, calls, reviews, surveys, and community with an adaptive taxonomy, which supplies the content-based support signal, the sentiment trajectory, the unresolved-ask history, and competitor mentions. The customer context graph ties all of it to the account, its segment, its revenue, and its renewal date, so the output is an ordered list of who to call rather than a set of scores.

Should health scores be a single number or multiple?

Report one composite for triage and keep the component subscores visible, because the composite tells a CSM which accounts to look at and only the components tell them why. A single opaque number produces a work queue nobody trusts, since the first question on any flagged account is what drove the change.

If your health score reads usage but not language, see how Enterpret handles voice of customer software.

‍