Sales & PartnershipsFree
Churn Analysis Tested Against Renewals
A warning sign that also appears in the accounts that renewed is not a warning sign, and the review says which is which.
River returns a document and a sheet. The document is the timeline with two dates marked: the day the customer did something they could not walk back, and the day you needed to know in order to reach your slowest save play. The sheet holds every warning sign present on the account, scored against the accounts that renewed in the same window, with the days of runway each one leaves and the play that runway can still reach.
Presence proves nothing. In an illustrative review, six of eight churned accounts had lost their champion, which sounds decisive until you count the twenty-nine of sixty-three that renewed and had also lost theirs. That signal raises the chance of leaving from 11.3 percent, the overall churn rate, to 17.1 percent, and it flags most of your book. Every candidate signal here runs past the same set of renewal decisions, and only the ones genuinely rare among accounts that stayed survive, with the raw counts beside every rate.
Written for the customer success lead running the review, the account manager who has to explain the loss, and the founder who wants the review to change something. It ends with rules rather than a lesson, written so they can be run against today's accounts and return names. The forward-looking read on a live account is the renewal risk brief, the expansion side is the expansion brief, and the meeting this belongs in is the QBR pack.
Why a list of warning signs is the wrong output
The move is borrowed and old. When investigators traced a 2003 hepatitis A outbreak to one Pennsylvania restaurant, asking only the sick what they ate would have shown which dishes were popular. So they also asked diners who stayed well. Salsa: 94 percent of the sick had eaten it, against 39 percent of the well. The key feature of analytic epidemiology is a comparison group, and a churn review that reads only the accounts that left has none.
Rarity is only half of it. What a warning sign is worth also depends on how common churn is. Cancer screening shows the same pattern: the lower the prevalence, the lower the predictive value, meaning the rarer the disease, the less a positive result means. When one account in nine leaves, most familiar warning signs are followed by churn less than a quarter of the time. The second test is runway: a sign only triggers a save if it leaves more days than your slowest play needs.
Take an illustrative account: Brackenhall Retail Group, 184,000 dollars, notice on 3 October. The account was lost on 6 June, when somebody exported 41 times their own twelve-month median. Their three sharpest signals, at 62 to 80 percent, all arrived inside 31 days, and the exec play takes 34. The usage rule did fire on 8 May with 29 days left, which was time enough for three of the four plays and not the one the team chose.
How it works
Add the lost account
Dated events from the account that left: usage, tickets, exports, people changes and the notice.
Add the renewals
Every renewal decision from the same period, kept and lost, with the signals each one showed.
Score against the kept
How often each signal shows up in accounts that stayed, compared with your overall churn rate.
Match play to runway
How early each signal arrives, set against how long each of your save plays takes.
What you get
- Every signal scored against the accounts that renewed in the same period, raw counts shown
- The loss point dated from the first act the customer could not walk back
- A detection deadline counted backwards from how long your slowest save play takes to arrange
- Median runway per signal, so a reliable signal that arrives too late is labelled that way
- Where each signal's data lived, and whether anybody ever queried it before the notice
- The surviving rules run across today's accounts, returning names and how many would really leave
- Quiet signals tested too, such as no feature requests in [your request register](/templates/voice-of-customer-to-product-pack)
Common questions
We have eight churned accounts. Is that enough to score anything?
It is enough to rank and not enough to prove. Every rate here is printed beside its count, so 5 of 8 never appears as 62 percent on its own, and any signal resting on fewer than five accounts is marked as tentative. Ranking eight losses against sixty-three renewals still beats reading the eight alone.
Champion departure is on every churn checklist there is.
In the illustrative review it was on the lost account too, 232 days before the notice. It was also on 29 of the 63 accounts that renewed, which puts it at 17.1 percent against an overall churn rate of 11.3. As a trigger it flags 51 of 61 live accounts. It is a real signal and nowhere near strong enough to spend an escalation on.
Our strongest signals arrive too late to act on. So what use are they?
They correct the forecast, which is a different job from saving the account. In the illustrative example, two or more irreversible acts runs at 83 percent and flags three accounts holding 412,000 dollars. Two tripped it inside the last 34 days and get one attempt. The third tripped it 71 days ago, so the honest move is planning the replacement revenue.
We have no play book with lead times attached.
Then the review assumes rough lead times for common plays, labels them as assumptions, and shows how the deadline moves if they are wrong. Better still, estimate from the last three saves you actually ran: days between deciding and the customer feeling it. Without lead times every signal looks actionable, which is how a review ends in a resolution nobody can run.
Is this not just blaming a CSM with the benefit of hindsight?
The opposite, because the deadline comes from your play book rather than from the outcome. In the illustrative example it lands on 3 May, and the verdict is that one signal fired on 8 May with 29 days of runway, and the team reached for the play that needs 34. That is a mismatch in the process, not a missed catch.
Some of this was never recorded anywhere. Then what?
Every signal lands in one of three buckets: on a dashboard somebody read, held in a system nobody checked, or not recorded at all. On the illustrative account all four of the highest-scoring signals sat in the second bucket, so the fix was a regular report on data already held rather than a new tool or a new health score.
How is this different from scoring renewal risk?
Risk scoring reads one live account against a model. This reads one dead account against the accounts that renewed alongside it to work out which inputs that model should have. Run this after a loss and it produces the rules; the renewal risk brief applies them forward. The lost-deal equivalent is the deal post mortem.
Sources
Built by
Chandler Supple
Co-founder & CEO, River
Chandler has spent the last six years working in AI, building tools for millions of users and Fortune 100 companies. In 2025, he co-founded River to build a new kind of operating system for AI agents.
Churn Analysis Tested Against Renewals
Describe what you need, add any files you have, and your workspace opens with the work already underway.