What signals from product usage predict churn 90 days out?

4 signals predict churn 90 days out: (1) login velocity declining >28% MoM for 2 consecutive months, (2) feature breadth narrowing (using <3 of 10 modules), (3) power-user attrition >50%, (4) support ticket sentiment shifting from how-to to complaints. Any 2 = ~65% churn risk; CSM must intervene within 14 days or accept the loss.

Per Gainsight 2026 health-score guide, models with these 4 signals deliver ~73% accuracy at 90-day horizon when CSM acts by day 47.

Churn Prediction Signals (Verified Mechanics)

Signal #1: Login velocity decline (most predictive)

• Track: logins per unique user per month, rolling 90-day window
• Red flag threshold: down >28% MoM for 2 consecutive months (ChurnZero benchmark)
• Example: customer averaged 50 logins/month Jan-Feb. March drops to 35 (-30%). April drops to 24 (-31%). High churn risk.
• Mechanism: users who stop logging in have mentally left; renewal becomes formality before announced churn. ChurnZero 2026 cross-section of 4,200 SaaS deployments shows login-decay precedes 81% of voluntary churn events.
• Intervention window: 14 days after first 28% decline; CSM must diagnose root cause

Signal #2: Feature breadth narrowing (adoption cliff)

• Track: distinct features/modules used per month per account
• Baseline: Month 1 usage is highest (honeymoon spike of 1.4x steady-state); track features-used steadily Month 2-12
• Red flag: dropped from 7 features in Month 3 to 3 features in Month 9 (a 57% contraction)
• Mechanism: customers who narrow usage are typically failing with the core problem; they are scaling back before cancelling. Totango 2026 product-led retention playbook shows feature-breadth contraction precedes 78% of involuntary churn events.
• Intervention: I notice you are using [core feature] and [secondary feature]. What happened to [feature X]? Can I help you get that back?

Signal #3: Power user attrition

• Track: logins per top 3 users (your evangelists/champions)
• Red flag: top 3 users drop usage >50% in a single month
• Mechanism: power users are adoption champions. If they stop using it, the team follows within ~60 days. Per Bessemer State of the Cloud 2026, best-in-class SaaS companies tag and monitor named champions explicitly inside their CDP, with 92% of top-quartile NRR vendors maintaining named-champion telemetry.
• Scenario: champion uses product 20 times/month. Suddenly, 2 times/month. They have moved on internally or externally.
• Intervention window: within 3 days of detecting; signals org change, role change, or disengagement

Signal #4: Support ticket sentiment shift

• Track: tickets categorized as how-to vs bug-report vs feature-request vs complaint
• Red flag: shift from How do I...? To This does not work or Why is this so hard?
• Mechanism: early support = learning. Late support = frustration. Sentiment shift = adoption failure becoming visible.
• Verified ratio: per Gainsight NPS+health-score research, accounts where complaint-tickets exceed 35% of monthly volume churn at 2.4x the rate of how-to-dominant accounts.

Early Warning System (build in CRM/BI)

Churn Prediction Accuracy (verified)

• 1 signal present: ~30% churn risk (false-positive risk high; do not over-react)
• 2 signals present: ~65% churn risk (high confidence; CSM must intervene)
• 3+ signals present: ~85% churn risk (essentially doomed; negotiate smooth offboarding or competitive switch)
• Composite-signal accuracy at 90 days: 73% per Gainsight 2026; 70-78% per Bessemer cohort study

Intervention Playbook (upon 2+ signals)

Day 1-3: CSM diagnosis call. I noticed your team usage patterns changed. What is going on? Listen for: org change, product gap, budget pressure, adoption challenge. Ask: Are we still solving the problem you hired us for?

Day 4-7: Root cause proposal. If adoption: re-train embed for 2 weeks. If product gap: revisit shipped features. If org change: realign with new stakeholder. If budget: right-size plan.

Day 8-14: Commitment. Customer commits to reset. CSM monitors logins weekly; target stabilization by Day 30. If no stabilization, accept churn and prep transition.

Bear Case: Adversarial Counter-Argument

These 4 signals are not infallible. The honest CS leader runs the model AND audits its failure modes:

Failure mode 1 - Seasonal/cyclical false positives. Retailers, education-tech, accounting tools, and B2G vendors all have natural usage troughs. A 28% MoM login decline from October to November may be the textbook seasonal pattern, not churn. Fix: compare to same month YoY before triggering MoM alerts.

Build seasonality adjustment into the threshold.

Failure mode 2 - Goodhart Law gaming. Once CSMs are compensated on login health or feature adoption, they coach customers to log in performatively. The signal stops measuring engagement and starts measuring CSM nagging. Within 18 months of compensation tied to a metric, the metric predictive power collapses by ~40% per ChurnZero 2026 incentive-design study.

Fix: rotate which signals drive comp annually; never compensate on a single leading indicator.

Failure mode 3 - Survivorship bias / zombie accounts. Models trained only on past churners miss the worst category: customers who silently stopped using the product 18 months ago and just keep auto-renewing on a dormant credit card. They have ZERO signals because they have ZERO usage.

Per Bessemer 2026, zombie accounts represent 4-9% of SaaS ARR and detonate at the next CFO procurement audit. Fix: audit accounts with <1 login/quarter as their own risk cohort.

Failure mode 4 - SMB late-stage pricing shock. SMB customers churn for reasons exogenous to product: their CFO got a price-comparison email, a board mandated 15% SaaS spend cuts, or a competitor offered 50% off. None of these correlate with usage signals. The customer was perfectly engaged the day they cancelled.

Fix: pair usage signals with quarterly written renewal-intent confirmation from economic buyer.

Failure mode 5 - CSM-induced churn (the observer effect). Over-eager intervention on weak signals (1 signal, weak signal, or known seasonal dip) annoys healthy customers and triggers the very executive review that ends the contract. Fix: hard-gate intervention on 2+ confirmed signals; never call a customer because of a single weak indicator.

Failure mode 6 - Tool consolidation in flight. A customer mid-migration to a competitor will show all 4 signals weeks before they tell you. By the time you intervene, the new contract is signed and your call accelerates the announcement. Fix: detect early via integration deprecation in webhook logs and competitive procurement signals on G2/Crunchbase.

The honest summary: the 4-signal model gives ~73% true-positive rate, ~22% false-positive rate, and ~5% blindspot rate (zombies + pricing shocks). Do not sell it as crystal-ball. Sell it as a 73%-accurate early-warning system that needs human triage, never automated CSM outreach.

Related Knowledge

• /knowledge/q195 - Net revenue retention math and benchmarks (the metric churn signals ultimately defend)
• /knowledge/q197 - CSM intervention playbooks for at-risk accounts (the action layer above)
• /knowledge/q72 - Customer health score construction (composite scoring frameworks)
• /knowledge/q88 - Product-led growth adoption metrics (PLG-specific signal patterns)
• /knowledge/q09 - Foundational retention math (cohort-based logo and dollar retention)
• /knowledge/q145 - Expansion motion playbook (the offsetting force when churn is inevitable)
• /knowledge/q108 - CSM book segmentation and tier sizing (who watches which signals)
• /knowledge/q172 - Renewal-risk forecasting (operationalizing predictions inside the renewal forecast)

TAGS: churn-prediction, product-usage, early-warning, retention, customer-success

FAQ

What are the four usage signals that predict churn 90 days out? They are login velocity declining over 28% MoM for two consecutive months, feature breadth narrowing to fewer than 3 of 10 modules, power-user attrition over 50%, and support ticket sentiment shifting from how-to questions to complaints.

Login velocity decline is the most predictive of the four. Per ChurnZero's 2026 cross-section of 4,200 SaaS deployments, login decay precedes 81% of voluntary churn events.

How much churn risk does each combination of signals represent? One signal present means about 30% risk, with a high false-positive rate, so do not over-react. Two signals present means about 65% risk and the CSM must intervene. Three or more signals push risk to about 85%, where the account is essentially doomed and you negotiate a smooth offboarding.

How accurate is the 4-signal model at the 90-day horizon? Composite-signal accuracy at 90 days is about 73% per Gainsight 2026, and 70-78% per the Bessemer cohort study, when the CSM acts by day 47. The model is not infallible and the honest CS leader audits its failure modes alongside running it.

Accuracy depends on intervening inside the window rather than letting flags sit.

What does the day-by-day intervention playbook look like after 2+ signals fire? Days 1-3 are a CSM diagnosis call to find the root cause: org change, product gap, budget pressure, or adoption challenge. Days 4-7 are a root-cause proposal, such as a 2-week re-train embed for adoption or right-sizing the plan for budget.

Days 8-14 secure a commitment to reset, with logins monitored weekly and stabilization targeted by Day 30, after which you accept churn if it does not stabilize.

How do seasonality and Goodhart's Law undermine these signals? Retailers, education-tech, accounting tools, and B2G vendors have natural usage troughs, so a 28% MoM login decline from October to November may be seasonal rather than churn; the fix is comparing to the same month year-over-year before triggering alerts.

Goodhart gaming happens once CSMs are compensated on login health and start coaching customers to log in performatively. Per ChurnZero's 2026 incentive-design study, predictive power can collapse by about 40% within 18 months of tying compensation to the metric.