Health Score

With twenty accounts nobody needed the number. Say a company's name and the person who owned it could tell you how things were going there. Somewhere past forty accounts that memory stopped working.

The real problem was picking who to contact today. Without a rule, the accounts that reached out most recently got the attention. But churn doesn't come from the noisy accounts. Complaining means a customer still plans to use the product; a company that has already decided to look elsewhere says nothing at all.

Pagedeck learned this the expensive way. One of Sofia's accounts went six months without filing a single support request, and she had it filed under doing fine. Two weeks before renewal an email arrived saying they wouldn't continue. The logs showed active users had been sliding for four months. The evidence was sitting where anyone could read it and nobody did.

A health score turns that memory and instinct into a record. It folds the activity scattered across each account into one value, which puts forty accounts in a single ordered list. The goal isn't an accurate diagnosis of every customer. It's an answer to which one gets looked at first this morning.

Pick what goes in, assign a weight to each input, write a rule that turns each input into a 0-to-100 score, then multiply and add.

The inputs are usage signals — behavior like login frequency or whether a particular feature gets used, which hints at how an account is doing. Which behaviors actually signal anything differs by product, so they get found rather than invented. Behaviors that departing customers shared in the year before they left are the candidates, and behaviors the surviving customers never did filter that list down.

Pagedeck uses four.

1. Weekly active share, 40 points — the fraction of contracted seats that logged in at least once in the last seven days.
2. Core feature use, 30 points — whether two or more people have edited the same document. A team that starts doing this has a hard time moving to another product.
3. Support state over 30 days, 20 points — whether unresolved tickets are piling up and whether any of them carry complaints.
4. Champion turnover, 10 points — when the person who signed the contract leaves, the next renewal becomes a fresh sale.

Take a 50-seat account where 20 people logged in last week. The first input scores 40, and at a weight of 0.4 that contributes 16. Co-editing is in use, so the second scores 100 and contributes 30. Three complaint tickets put the third at 50 for 10 points, and an unchanged champion puts the fourth at 100 for another 10. The account lands at 66.

Scores recalculate daily, and the current value gets read next to the last four weeks of movement. A lone value can't tell an account that hit 66 yesterday apart from one that has sat at 66 for three months.

The line between safe and not doesn't come from outside. It comes from your own churn record. Borrowing 70 because that's what the industry uses means borrowing someone else's ruler without knowing which signals at which weights produced their 70.

The way to find it is to fit the score against last year's outcomes. Pagedeck found that 70 percent of the customers who left that year had dropped below 45 three months before leaving, and set 45 as the line. Anything under it gets filed as a red account automatically. Whether the line does any work shows up in renewal rates: if accounts under 45 renew at 52 percent and accounts above 70 renew at 94 percent, the score predicts something. If the two rates come out close together, the score predicts nothing and the weights need another look.

Direction beats the absolute value. An account that fell from 88 to 61 in three weeks is more urgent than one that has sat at 55 for six months. The second may simply be a light user; the first means something happened in the last three weeks.

Weights get recalibrated about every six months. The first set of weights encodes the assumptions of the day it was built, and when the product changes, so does the question of which behavior counts as a signal. When an account at 82 leaves without a word, that's a moment to revisit the weights rather than blame the customer.

"This account went from 88 to 61 in two weeks. Let's find out what happened before we do anything."

The size of the drop matters more than the value, and judgment waits until the cause is known.

"The score says 72 but I'd call it at risk. Their decision maker changed last month."

An argument for overriding the classification by hand, based on something the score can't see.

"Six accounts are under 45 this quarter. I'll work them in order of contract value."

There aren't hours for all of them, so the list the score produced gets narrowed once more by a different rule.

Pagedeck's customer success team reads the sorted list together every Monday morning. Sofia, the customer success manager (CSM) on most of those accounts, starts with the ones that moved more than 10 points last week and drops the ones she can explain. An accounting-team customer that only works in bursts at quarter end is supposed to dip in the first weeks of a quarter, so that one passes. What's left after the explainable drops come out is usually three or four accounts out of forty, and those get the week's time.

The most common misreading is treating a high score as proof that an account is safe. The number looks like a measurement of health, so the assumption is easy to make. It's calculated purely from behavior inside the product, which means everything happening outside it is missing. An acquisition, or a budget cut for next year, leaves no trace in any usage log. That's how an account at 82 ends in a one-line cancellation. Read the score as a safety certificate and nobody calls the high-scoring customers, which closes the one channel through which outside news would have arrived.

Using it to evaluate the people who own the accounts causes damage of the same size. Once the average score across someone's accounts counts as their performance, the fastest way to raise it stops being helping customers and becomes overriding classifications by hand. That manual override exists so the score can absorb facts it can't see; attach an evaluation to it and the same door gets used to dress up the number. Six months later the team average is up and churn hasn't moved.