Key Takeaways
- Product adoption metrics measured as levels don’t work: in 6.5 years of data on 10,279 subscribers, active and churned accounts were statistically indistinguishable on usage volume, yet models on the same data predicted churn a year out with 85%-plus accuracy from the shape of usage over time.
- The strongest churn predictor in a B2B SaaS study of 8,878 accounts wasn’t a product signal at all: 37–55 days without a positive interaction outranked login inactivity and core-feature disuse.
- Silent accounts churn more than complaining ones. TARP’s research found dissatisfied customers who complained and got nothing were still more loyal (46%) than those who said nothing (37%).
- Activation is the one adoption metric with real evidence: 69% of products in the top quartile for seven-day activation were also top quartile for three-month retention across 2,600-plus companies.
- No customer success vendor publishes how often its health score marks a churning account green. The widely quoted “73% of CS leaders say health scores don’t predict churn” figure doesn’t exist in the study it’s attributed to.
- Your support and knowledge data already holds the two dependency signals that beat logins: whether the account finds answers without asking, and whether it keeps asking the same question.
The renewal notice arrives on a Wednesday. The account is at 14 months. Its health score has been green for the last 90 days: logins steady, three of five modules touched, NPS a 7 from the champion in April. Nobody on the CS team flagged it. Nobody could have, from that dashboard. The score didn’t break on Wednesday. It was never measuring the thing that decided the renewal.
Every leader who has watched a green account churn asks the same question afterwards: which product adoption metrics actually predict renewal? Most answers list twenty metrics and rank none. The evidence, which exists and is more interesting than the lists, says something else. The metrics on your dashboard aren’t wrong. They’re measured the wrong way, and they leave out the two signals that matter most, which live outside the product.
No vendor publishes how often its health score is wrong
Start with the number you’ll be tempted to quote. A widely circulated claim says 73% of CS leaders admit their health score doesn’t reliably predict churn, sourced to ChurnZero’s 2025 study of about 800 leaders. The study exists: 793 respondents, sixth annual edition. The finding doesn’t. The only 73% in the document is the share of teams using AI for call summarisation. The sample size matched, so the fabrication travelled.
Look for the real version and you find silence. Gainsight’s post on why green customers churn contains four qualitative causes and no percentages; one of the four is “incorrect health scores”. ChurnZero’s guide to measuring health-score effectiveness poses the question of low-risk accounts that churned as an exercise for the reader. The vendors that sell the score don’t publish its false-negative rate. The only source of that number is your own churn list.
So the first move is procedural. Take every account that churned or contracted in the last four quarters. Pull its health colour at 90, 60 and 30 days before the renewal date. Count the greens. That confusion matrix, drawn by hand in an afternoon, is worth more than any benchmark, because it’s the only one that describes your score on your customers.
Product adoption metrics measure levels. Renewal is decided by trajectories
The most rigorous study on this question used 6.5 years of real subscription data from an academic publisher: 10,279 users, 5,141 of whom stayed and 4,552 of whom lapsed. Roberts, Deza, Ihshaish and Zhu compared the two groups on usage volume and found “both populations are not significantly different, from a statistical point of view”. Active users downloaded a bit more per day. Not enough to separate them. Their conclusion: predictions based on statistical properties “will only work in extreme cases”.
Then they fed the same data to time-series models and hit 90% accuracy 200 days ahead, and above 85% a full year ahead. Same accounts, same usage. The signal wasn’t in how much a customer used the product. It was in the shape of usage over time: the slope, the variance, the pattern of gaps.
That result dissolves the “which metric” framing. Logins, DAU/MAU, feature counts, minutes in app: every one of them is reported as a level, a number for this month. A level tells you where an account is. Renewal depends on where it’s going, which is the whole case for managing renewals where the decision is actually made. The same login count, differenced over time, becomes predictive. The metric was never the problem; the aggregation was.
One more finding from the paper will save someone a data-lake budget. Accuracy peaked when usage was resampled to 17-day windows, and every model got worse on daily, high-frequency data. The authors say further customer data “can be extraneous”. If you’re being sold event-level telemetry as the road to renewal prediction, the best evidence available says coarser is better.
The product adoption metrics that beat logins weren’t in the product
A 2025 study of 8,878 small-business customers of a Brazilian HR-software company ranked 34 features by how strongly they predicted churn. The winner, with a coefficient of 1.51, was an absence of positive interactions with the company for 37–55 days. Login inactivity of 8–44 days scored 1.42. Not using the core feature for 69–823 days scored 1.23. The authors call the silence signal “an even stronger indicator of churn than inactivity in logins”.
That matches a much older result from a different tradition. TARP’s research for the US Office of Consumer Affairs, later confirmed across 500-plus surveys, measured loyalty by complaint behaviour. Among customers with a small problem, 37% of those who stayed silent remained loyal, against 46% of those who complained and were left unsatisfied, and 70% of those whose complaint was resolved. For problems worth over $100 the gap doubled: 9% loyal among the silent, 19% among unsatisfied complainers. A customer who complains and gets nothing is still more likely to stay than one who never spoke.
For a support and enablement leader that has a sharp edge. The ticket queue is an engagement signal, not only a cost. An account that files three tickets in a quarter is telling you it still wants the product to work. An account that filed none, whose champion stopped replying, and whose users stopped searching the help centre is the one to call. Deflection targets that push engaged customers into silence can be retention-negative, which is a reason to measure self-service by resolution rather than by avoided contacts.
Activation is the one product adoption metric with evidence behind it
Amplitude’s benchmark report covers behavioural data from more than 2,600 products over a year. Its headline: 69% of products in the top quartile for seven-day activation were also in the top quartile for three-month retention. Read that carefully. It’s a co-occurrence between top groups, not a correlation coefficient, though blogs routinely restate it as one. It still makes activation the best-evidenced adoption metric there is, and it’s the reason finding the real activation event matters more than anything downstream.
The same report has two findings the adoption dashboards ignore. First, there was no relationship between products in the top quartile for adding users and those in the top quartile for retention. Seat growth predicts nothing. Second, engagement is savagely concentrated: 10% of products account for 79% of all user engagement, and for half of products more than 98% of new users are inactive two weeks in. Any cross-company DAU/MAU benchmark is comparing products in different universes. For B2B tech specifically, three-month retention is 2.5% at the median and 15.6% at the 90th percentile. If your DAU/MAU target came from a consumer app, it’s not a target.
Dependency: the adoption signals your support and knowledge data already hold
Here is what the studies have in common. Trajectory beats level. Silence beats logins. Activation beats everything measured later. Each is a way of asking one question: does the customer’s own work run through your product, and would they notice if it stopped? That’s dependency, and it’s what a renewal decision is actually about.
Two proxies for dependency show up in the retention data. Price point is one: ChartMogul’s analysis of 2,100-plus companies’ billing data found median net revenue retention of 38.2% for products under $10 a month against 78.5% above $500. SaaS Capital’s survey of 1,500-plus private companies puts median gross retention at about 93% above $25,000 ACV and 90% below. The bigger the contract, the more of the customer’s operation is wired to it. Integrations are the other, but be careful with the numbers. The most quoted one, that integration users are 58% less likely to churn, is what survey respondents reported believing, not a measurement. The ProfitWell integration benchmarks it usually travels with now redirect to a page that no longer contains them.
The proxies you control sit in the surfaces your team runs. They’re better than the ones above because they’re per account and they move weekly.
| Signal | Measured as | Where it lives | What it says about dependency |
|---|
| Self-service resolution by account | Trajectory | Help centre and in-app assistant | Users are finding answers and continuing work without you: embedded |
| Repeat-question rate by account | Trajectory | Search logs, assistant transcripts, tickets | The same question three times means nothing was adopted the first time |
| Days since last positive interaction | Level with a threshold | Inbox, CSM notes, community | The strongest single churn predictor in the B2B study above |
| Onboarding milestone completion vs. plan | Trajectory | Onboarding project | Activation, measured on the customer’s timeline rather than yours |
| Number of distinct people asking questions | Trajectory | Help centre, assistant, tickets | One champion is a single point of failure; five askers are a dependency |
None of these appear in a product-analytics tool, because none of them happen in the product. They happen where the customer meets your team. Which is why the leader who owns support, self-service and onboarding is holding better renewal data than the leader who owns the usage dashboard, and usually doesn’t know it. It’s also why the handful of customer success metrics that change a decision are mostly the ones a blended team can see and a product-analytics tool can’t.
Product adoption metrics scorecard: wrong in a way you can see
The fix isn’t a fifteenth input. It’s a score built so that when it’s wrong, you can tell which input lied. Four rules, all drawn from the evidence above.
- Five inputs, each a trajectory. Activation against plan, usage slope over 30–60 days, days since last positive interaction, self-service resolution by account, repeat-question rate. Drop the login count as a level; keep its slope.
- Weight silence above usage. The one B2B study that ranked them put no-positive-interaction above login inactivity. Until your own confusion matrix says otherwise, borrow that ordering.
- Refit every renewal cycle. The Brazilian study reported its model honestly under rolling evaluation: F1 fell from 0.62 to between 0.06 and 0.15, and the team retrains whenever recall drops below 60%. A score fitted on last year’s churn is describing a company that no longer exists. Gainsight’s own churn interviews say the same in words: metrics set at the start of the journey “may not remain valid indicators”.
- Publish your own false-negative rate. Greens that churned, by quarter, on the same slide as the score. It’s the number no vendor will give you, and the one your CFO will trust.
Rules one and three need the inputs in one place. That’s where most scores quietly fail. Usage sits in the analytics tool, tickets in the help desk, onboarding in a project tracker, help-centre searches in a fourth system. The score reads a monthly export of each. It can’t compute a slope from a snapshot. In MatrixFlows the account, its requests, its onboarding milestones and its knowledge activity are one linked record, so the trajectory is there to read. This account’s users searched 40 times last month and found answers 34 times; this one asked the same firmware question in three channels. Search-gap analytics show which accounts are asking things nobody has answered. The health score built on a foundation like that reads dependency directly instead of inferring it from logins, and it’s the difference between the leading indicators and the ones that confirm a churn after the fact.
The green account on Wednesday wasn’t a data problem. It was a definition problem. Adoption isn’t what the customer does inside your product. It’s what the customer can no longer do without it.
Create a Free Workspace → and run last year’s churned accounts through the five signals before the next renewal quarter.
.svg)
.svg)
.svg)
