The method
How to calculate a churn risk score for every customer
A churn risk score turns the behaviour you already record — logins, tickets, failed payments, seats — into one number per account, so you can sort the list and work the top of it. Below is the exact model this site runs: six signals, six weights, one formula, and a worked example you can check on a calculator before you trust it with your own data.
The formula, in one line: score = the sum, across six signals, of severity × weight, where each severity is a number from 0 to 1 and the six weights add up to 100. Round the total and you have a 0–100 score. Higher means more risk.
What a churn risk score is — and what it is not
It is a ranking device. You take a fixed set of signals you can get for every customer, score each one, weight them, and add them up. The output sorts your book of business from “about to leave” to “fine”, so a small team can spend its week on the accounts where a conversation still changes the outcome.
It is not a prediction. There is no trained model behind it, no probability, and no claim that a 70 leaves more often than a 60 at your particular company. It is not comparable across companies either: someone else’s “health score” is built from different columns with different weights, so a 62 there means nothing here. What it is good for is comparison inside your own book, this month, and comparison of the same account against itself last month.
That constraint is deliberate. A model you can read in a table is one you can argue with, fix, and explain to a founder or a board without a data scientist in the room.
The six signals and their weights
Each signal is converted to a severity between 0 (no risk on that dimension) and 1 (full risk), then multiplied by its weight. The weights are judgement, not science: they say that going quiet matters more than being new, and that a failed payment is worth more than an unused seat. They add to exactly 100 so the total lands on a 0–100 scale.
| Signal | Weight | Full severity means | Severity rule |
|---|---|---|---|
| Recency | 30 | Last login 30+ days ago | days since last login ÷ 30, capped at 1 |
| Login frequency | 18 | Zero logins in 30 days | 1 − (logins in the last 30 days ÷ 10), floored at 0 |
| Payment health | 18 | Two or more failed payments in 90 days | failed payments in 90 days ÷ 2, capped at 1 |
| Support load | 14 | Five or more tickets in 90 days | tickets in 90 days ÷ 5, capped at 1 |
| Seat usage | 12 | Under 30% of paid seats in use | 1 − (seats used ÷ seats licensed ÷ 0.7), floored at 0 |
| Tenure | 8 | Signed up 30 days ago or less | 1 − ((days since signup − 30) ÷ 335), floored at 0 |
Why these six, in that order
- Recency carries the most weight (30) because a customer who has stopped opening the product has already left in the only way that shows up before the cancellation. Nothing else on the list is as direct a reading of whether the product is still in their week.
- Login frequency (18) and payment health (18) sit together as the two next-most-actionable facts. Low logins is a slow signal you can still intervene on; failed payments is a fast, unambiguous one that is often a billing-detail problem rather than a decision to leave.
- Support load (14) counts up as well as down: an account with a pile of open tickets is either struggling or has a broken integration, and both end in a renewal conversation you'd rather have first.
- Seat usage (12) catches the account that pays for ten and uses three — the seats they cancel first are the ones nobody logged into.
- Tenure (8) is the smallest weight because being new is a phase, not a fault. It matters at the margin: a 20-day-old account that also hasn't logged in is worse off than a two-year account with the same login count.
When a column is missing
You will rarely have all six columns for all customers. Rather than treating a blank as a failure (which would rank your best-documented accounts as the healthiest), a missing input gets a fixed middle value, so the score is driven by the columns you do have:
| Missing input | Severity used | Points it contributes |
|---|---|---|
| Recency | 0.60 | 18.0 of 30 |
| Login frequency | 0.50 | 9.0 of 18 |
| Payment health | 0.30 | 5.4 of 18 |
| Support load | 0.40 | 5.6 of 14 |
| Seat usage | 0.35 | 4.2 of 12 |
| Tenure | 0.40 | 3.2 of 8 |
That is why the audit tells you when a column was missing instead of quietly guessing: if most of your list has no payment history, the payment signal is contributing the same points to everybody and the ranking is effectively decided by the other five.
A worked example you can check by hand
Take one made-up account — the numbers are invented for the arithmetic, not a benchmark. It pays $450/month, signed up 300 days ago, last logged in 42 days ago, logged in twice in the last 30 days, has 3 support tickets and 1 failed payment in the last 90 days, and uses 12 of its 20 seats.
| Signal | Severity | × weight | Points |
|---|---|---|---|
| Recency | 42 ÷ 30 = 1.40, capped at 1 | 30 | 30.0 |
| Login frequency | 1 − (2 ÷ 10) = 0.80 | 18 | 14.4 |
| Payment health | 1 ÷ 2 = 0.50 | 18 | 9.0 |
| Support load | 3 ÷ 5 = 0.60 | 14 | 8.4 |
| Seat usage | 1 − (0.60 ÷ 0.70) = 0.14 | 12 | 1.7 |
| Tenure | 1 − (270 ÷ 335) = 0.19 | 8 | 1.6 |
| Total → rounded | 65 | ||
65 lands in the High band. The interesting part isn’t the number, it’s the shape of it: recency alone is contributing 30 of the 65 points, so the save play writes itself — a founder email naming what shipped since their last visit, and two specific call slots this week. If instead the 65 had come from support load and seat usage, the right move would be a consolidated fix plan, not a re-engagement email.
The bands, and what to do in each
Work these this week, highest MRR first. Each one has a save play attached in the report — book the call, don't queue an email.
Real damage on at least one signal. Get a human touch in front of the account before the next renewal date.
One signal is drifting. Put them in a monthly review rather than a weekly one, and watch the trend, not the single number.
Healthy. Move them to a quarterly check-in, and ask for the referral or case study while they're happy.
Five mistakes that make a score useless
- 1
Scoring against today instead of the data
If you score a CSV exported last month using today's date, every account looks more neglected than it is and the whole list inflates. The reference date should be the most recent activity in the file, so re-scoring the same export gives the same answer.
- 2
Treating a missing column as a failing one
An account with no payment data isn't an account with failed payments. Missing inputs should land near the middle of their range, which is what the defaults below do — then the score tells you about the account, not about your export.
- 3
Reading the score as a probability
A score of 70 is not a 70% chance of leaving. It is a rank: this account shows more of the behaviour that precedes churn than one scoring 40. Sorting and trend are the signal; the absolute number is a scale you defined yourself.
- 4
Changing the weights every month
A score is only useful if it moves because the customer moved. Reweight the model between cycles and you can't tell a recovering account from a re-tuned one. Pick weights, keep them, revisit them once a quarter at most.
- 5
Folding MRR into the score
Revenue belongs next to the score, not inside it — otherwise a quiet $2,000 account outranks a loud $200 one and you can no longer see which accounts are actually in trouble. Sort by risk, then multiply by MRR to decide what to work first.