There's a debate that runs through almost every customer success team eventually: do you watch what your accounts are doing inside the product, or do you watch what they're saying to your support team? Most CS programs pick one — usually product usage, because it's more legible and easier to instrument — and quietly neglect the other.
That's a mistake. But understanding why requires pulling apart what each signal actually tells you and where each one breaks down on its own.
What Usage Data Actually Measures
Usage telemetry captures behavioral engagement: how frequently an account logs in, how deeply they've adopted features, how their session cadence has changed over the past 30 or 90 days. A well-instrumented usage signal surfaces things like:
- Login cadence decay — logins dropping from daily to weekly to sporadic
- Feature adoption regression — previously-used features going dormant
- DAU/WAU ratio collapse — active user count shrinking faster than seat count
- Core workflow abandonment — the key job-to-be-done actions going quiet
These are genuinely predictive signals. Across mid-market B2B SaaS cohorts, accounts showing significant usage decay in the 60–90 days before renewal are meaningfully more likely to churn than accounts sustaining steady engagement. The directional correlation is well-established in the CS literature, and platforms like Amplitude, Pendo, and Gainsight have built substantial product surface area around surfacing it.
The problem is what usage doesn't capture: the emotional state of the account.
What Support Sentiment Actually Measures
Support ticket volume and sentiment tell a different story. They capture frustration accumulation — the slow buildup of unresolved friction that precedes a non-renewal decision, often at a time when usage is still nominally active.
Consider what happens when a mid-market account is three months from renewal and has a billing dispute, a broken integration, and two feature requests that haven't shipped. Their power users are still logging in because they have to — the tool is embedded in their workflows. Usage looks fine. But their VP is writing escalated Zendesk tickets, and the primary admin has stopped responding to CSM outreach. The sentiment signal is screaming. The usage signal is murmuring.
This is the pattern we see most consistently in accounts that churn "unexpectedly." Unexpected is usually a failure of signal breadth, not a failure of signal accuracy. The usage data was technically correct. It just wasn't reading the right variable.
The Case Study: Two Signals, Different Timing
Here's a concrete scenario that illustrates the divergence. A 40-seat account at a workflow automation SaaS — call them Brackfield — had been on a Growth plan for 18 months. Their Segment event stream showed steady-to-declining usage: not alarming, but down 15% on session frequency versus the prior 90-day cohort. A usage-only health score would have flagged them as "watch" — amber, but not critical.
Meanwhile, their Zendesk ticket volume had doubled in the same window. Sentiment scoring on ticket text showed escalating frustration language — "still waiting," "this is the third time," "I need to escalate this." Their CSAT on resolved tickets had dropped from 4.2 to 2.9. Two open threads were aging past 14 days without resolution.
A usage-only signal: amber, revisit next quarter. A sentiment-only signal: red, intervention needed now. A combined signal: red, with specific context about what the CSM should address in the outreach (the open tickets, not the login frequency).
This is not a cherry-picked edge case. The pattern — usage at yellow, sentiment at red — is one of the most common configurations in accounts that churn within 90 days. The inverse pattern also exists: sentiment flat, usage crashing. Accounts where the champion has gone quiet but nobody is complaining because they've emotionally checked out already. Both patterns require both signals to surface reliably.
When Each Signal Leads
Neither signal is categorically more predictive. The leading indicator depends on the churn type:
- Champion departure churn — the main internal advocate leaves or changes roles. Usage drops first; sentiment stays flat or even improves temporarily as new users explore the product cautiously. Usage leads here.
- Frustration churn — recurring unresolved support issues erode trust until the account decides not to renew. Sentiment leads; usage may stay normal until very close to the decision point.
- Budget reallocation churn — the account isn't dissatisfied, they're just reprioritizing spend. Both signals stay relatively flat until the last 30–60 days, when usage decays sharply and sentiment turns pragmatically neutral. Neither leads clearly; cohort benchmarking and renewal date proximity matter more.
- Competitive displacement — a rival product is winning internal evaluation. Usage drops as evaluation of the competitor takes time away from the incumbent tool; support tickets may shift in nature (questions about export capabilities, data portability, API access). Both signals are active, but in specific ways.
A scoring model that weights only one signal will systematically miss at least two of these four churn types. The CS teams that catch the most at-risk accounts before renewal have both streams wired into their health scoring methodology.
The Dual-Signal Challenge: Instrumentation Is Harder
We're not saying usage-only health scores are badly built — they're often the pragmatic result of what data is accessible. Product analytics platforms are typically well-integrated into the data stack from early in a company's life. Support platforms are sometimes treated as a separate operational silo, with CX owning Zendesk separately from the CS team that owns health scores.
This organizational split is the real reason most health scores are usage-only. It's not that CS leaders don't understand sentiment's value — it's that wiring Zendesk or Intercom into a health scoring model requires cross-team data access that has historically been hard to configure without engineering involvement.
The instrumentation gap is the gap that dual-signal scoring platforms are built to close. When both streams flow into a single account-level model, the combined risk score surfaces the patterns that either signal alone would miss.
A Note on Weighting
When both signals are present, how should they be weighted? There isn't a universal answer — it depends on your product type and your support model. For self-serve-heavy products where support ticket volume is naturally low, usage telemetry should carry more weight. For high-touch enterprise or mid-market products where support relationship quality is a primary satisfaction driver, sentiment should carry more. For most mid-market B2B SaaS companies, something close to equal weighting with cohort-adjusted baselines performs better than a static 70/30 split in either direction.
The Vendarix approach normalizes both signal streams against your account's own historical baseline and your cohort's average, before combining them. An account that typically opens 2 tickets a month and opens 8 is a very different situation from an account that typically opens 10 tickets and opens 12. Absolute volume matters less than deviation from pattern.
The Practical Question for CS Teams
If you're currently working with a usage-only health score, the first step isn't to tear it down — it's to audit what it's missing. Pull your last 10 churned accounts and check their Zendesk/Intercom history for the 90 days before non-renewal. What was ticket volume doing? What was the tone of the last five support interactions? If you find that most of your "surprise" churn had visible support sentiment signals that never surfaced in health scoring, you have the business case for expanding your model.
If you find that your churned accounts were truly quiet on both signals until very late — that's a different problem, and it points toward either renewal date scoring or competitive displacement signal work rather than sentiment integration.
The question isn't which signal matters more. It's whether your current model is reading both.