Usage Signal

For a long time, a customer's state after signing was judged by the account manager's read: the tone of the last call, how fast emails came back. Those aren't bad inputs, but they share one weakness. They contain only what the customer chose to say.

Words and behavior diverge often. Someone with a warm relationship with their account manager frequently isn't opening the product at all, and that person won't complain right up to the end — they have already found a replacement, so there's no reason to raise it.

The gap surfaced at Pagedeck. An account Sofia managed gave notice two weeks before renewal. The last call had been two months earlier and had gone fine. Opening the usage records showed that weekly active users had fallen from 9 to 2 three months earlier, and document creation had been zero for over a month. That information had been sitting somewhere nobody looked. Using the customer's behavior instead of the customer's statements as the input to judgment is where this concept comes from.

Signals get found, not invented. There are two ways to find them. Behaviors that customers who actually left had in common in the three months before leaving are candidates, and behaviors that retained customers also did filter those candidates out. Anything both groups did is not a signal.

What gets found tends to cluster into four kinds.

1. Breadth of access. How many people use it. An account with 22 weekly users out of 30 seats and one with 3 are in different states while paying the same amount.
2. Core feature use. Whether the action the product was bought for is happening. At Pagedeck that's creating a document, sharing it, and getting it reviewed.
3. Expansion behavior. Whether new members get invited, whether a folder for another department appears. Movement here means room to grow the contract.
4. Relationship signals. Ticket frequency, response time, whether the customer's own contact changed. Not usage records, but read alongside them.

Reading one alone is guaranteed to mislead. Watch only active users and an accounting team that has always worked in bursts at quarter end turns red every quarter. So signals get read as a set, and against how that customer has always used the product.

Change matters more than the absolute value. "Two weekly users" is far weaker than "nine down to two." An account that was always at two being at two is not news at all. That is also why a signal list is worth nothing in its first month; the baseline it gets compared against has to accumulate first.

Which signal is strongest differs by product. In a document product several people share, five different people touching the same document is a much stronger signal than one person logging in daily. For a tool used alone, the reverse.

The precondition is the record. Event tracking, which logs what users clicked on which screen, has to be in place first, and there is no way to manufacture last quarter's log after the fact.

The first use is producing a score. Several signals with weights attached and collapsed into one value is a health score. Pagedeck assigns 40 points to weekly active share, 30 to core feature use, 20 to ticket state over the last 30 days, and 10 to whether the customer's contact changed. Converting to a number is about ordering: laid out as raw signals, there's no answer to whether the account with falling logins or the account with rising tickets gets attention first.

Once a score exists it drives marking a red account as likely to churn. After the marking, though, the signals behind the score get unfolded again. Forty-five points establishes order but not action, and an account at 45 because logins fell needs something different from one at 45 because tickets piled up.

The prep material for a quarterly business review (QBR) comes from here too. Walking in with what rose and what fell last quarter turns the conversation from checking in to checking facts.

It runs in the other direction as well. Accounts with active expansion behavior become candidates for growing the contract, and a feature whose use drops across many accounts at once may be a product problem rather than a customer problem, which sends it to the product team.

One limit stays. Signals show the state and not the reason. That a customer was acquired, or that next year's budget was halved, appears in no usage record. So this input doesn't replace conversations — it aims them. The practical form of that is a standing agenda: bring the two signals that moved most, ask what changed, and let the answer decide whether the account is actually at risk.

"This account went from 9 weekly users to 2. We need a call."

Picking which account to look at first based on the size of a change.

"The accounting team only uses it at quarter end, so lower the weight on active users for them."

Asking that a customer's normal pattern be reflected.

"This one isn't visible in the usage signals. The budget came up in the review."

Naming a fact behavior records can't hold.

After the churn incident, Pagedeck rebuilt its signal list. Comparing the final three months of records across 17 departed accounts, what showed up in common wasn't a drop in active users — it was a drop in the number of people editing documents together. Accounts turned into something one person read, and then they left.

So weekly editors went to the top of the list, with an alert to the account manager when it falls four weeks running. Of the 11 accounts reached through that alert next quarter, 8 renewed.

The most common mistake is reading login counts as the signal. It's the easiest number to obtain. But a visit where someone did nothing and a visit where they created and shared a document both count as one. Manage by logins and you can raise the number by sending more notification email, and a number raised that way has no relationship to renewal. Left uncorrected, the manager leaves green accounts alone and gets surprised at renewal time.

Assuming more signals means better judgment is the second mistake. More inputs feels like a sharper picture. In practice a list of twenty is a list nobody reads. A signal earns its place only if it can change what the manager does next. Pagedeck's list holds four, with a rule that adding one means removing one.