Churn Signal

Cancelling looks like something that happens the day someone clicks the button, but the decision is usually weeks old by then. The person who used it moved to another tool, or the champion changed jobs and nobody opened the product again. Or they hit one obstacle and never came back. Cancelling is the paperwork on a decision already made weeks earlier.

Those weeks were invisible to the team building the product. The dashboard held two events, signup and cancellation, and everything between them sat in the logs. So the monthly meeting opened with "32 accounts cancelled" and closed with "we'll bring that down next month." The 32 were already gone, and nobody could name the accounts leaving right now.

Large accounts have a person assigned, so a couple of calls are enough to sense trouble. The problem is the thousands of teams on free and low-priced plans. There is nobody to assign, so there is no atmosphere to read, and emailing all of them every week annoys the ones who are perfectly happy.

The churn signal replaces that human sense with behavioral records. Turn "they've been quiet lately" into countable conditions — login gaps, whether a given feature gets used — and a daily sweep of thousands of teams can hand you the sixty worth looking at.

Find what departed customers did in common before they left, then keep only the behaviors that actually track with leaving. Four steps.

1. Gather the customers who left. Take every account that cancelled in the last six months and look at the 90 days before each cancellation. A shorter window leaves too few cases; a longer one mixes in records from a product that has since changed.
2. Pull out candidate behaviors. Widening login gaps, a long-used feature going untouched, no new teammates being invited, the billing contact changing, support tickets spiking or going silent. Anything recorded automatically can be a candidate; anything a person has to notice and type in cannot.
3. Measure each candidate's hit rate. Compare the churn rate among people who did the behavior against the overall churn rate. If the overall rate is 5 percent and this group runs at 7 percent, the behavior is worthless as a signal; at 30 percent it's worth using. Skip this step and you end up with a list of behaviors that merely sound plausible.
4. Set a threshold. Not "usage dropped" but "weekly active users below half of four weeks ago" — something a machine can decide. A loose line produces more accounts than anyone can act on; a tight one catches only people who are already gone.

One behavior alone is rarely enough, because healthy teams go quiet for two weeks in August too. So a flag goes up only when two or three signals overlap: usage halved, no new document in 30 days, and nobody invited last month.

Which behaviors count as signals differs by product. In a collaboration tool used daily, a four-day gap is a signal; in a billing tool used once a quarter, two months of silence is normal. Borrow another company's list and the emails go to the wrong people.

Settled signals feed a daily evaluation that branches three ways.

1. Flagging an at-risk user. Accounts where signals overlap get collected into a daily list. Signals attach to single behaviors; this list is a conclusion drawn at the account level, and only here does "who should we contact" produce names.
2. Triggering automation. A reactivation campaign — the set of emails that call a quiet user back — takes this list as its input. The content can branch by signal: teams that stopped inviting people get help adding teammates, teams stuck at one feature get a walkthrough of that feature.
3. Fixing the product. If signals light up disproportionately after one particular screen, that screen is the problem. Repairing the blockage costs less than chasing everyone who hit it, and it stops the list from refilling with the same accounts next month.

It also connects to the health score, a single number per customer that combines usage and responsiveness. The items that pull that number down are mostly these signals. A signal answers true or false while a score answers by degree, so the score sets the order you call people in and the signals tell you what to say.

"This account has three churn signals lit. Someone should reach out this week."

The behavioral record crossed the threshold, so it's a person's turn to step in.

"That's too weak to use as a churn signal. It barely differs from the baseline churn rate."

People who do the behavior and people who don't leave at similar rates, so it separates nothing.

"Signals doubled because it's vacation season. I don't think we should let the emails go out."

The threshold was crossed, but the silence has another explanation, so pause the automation.

At Pagedeck, Ravi owns this list. Every Monday he opens last week's flags and starts by counting how many accounts lit up and then went active again with no intervention at all. A high share of those means the conditions are too loose, and he resets the thresholds until the weekly list lands near sixty accounts, which is what the team can work through.

Once a quarter he rebuilds the signal list itself. Last quarter the mobile app shipped and the web-only conditions fell apart — many teams that had gone quiet on the web were using the app daily — and he shut off the emails going to them within a week.

The most common error is reading a signal as a cause. It appears before people leave, so removing it looks like it would remove the churn. That produces "usage drops before they leave, so let's raise usage," and daily notifications push logins up by force. But the drop in usage is the result of the reason to use it disappearing, not the cause of the departure. Pushing the mercury down by hand doesn't lower the fever. Teams whose logins were raised by notifications cancel the next month with the signal switched off.

The other one is assuming more signals are better, since fewer people slip through. Expand to twenty conditions and hundreds of accounts land on the list every week, which means nobody reads it. The list is only useful when it stays within the number of accounts someone can actually contact that week.