How do you know which users will churn?

Churn prediction starts with direct observation, not dashboards. The founders most likely to catch churners early are the ones still close enough to individual users to notice when enthusiasm fades before cancellation happens. At the earliest stage, behavioral distance — not product flaws — is usually the first signal.

The earliest churn signal is disengagement, not cancellation

Most founders look for churn in their billing system. By that point, it's already over. The real signal appears weeks earlier: a user who was active daily starts logging in weekly, then stops asking questions, stops giving feedback, stops complaining. Complaints are actually a green flag — a user who emails you about a bug is still invested. The ones who go quiet are the ones leaving.

This is why staying close to individual users during the early stage is so operationally valuable, not just emotionally satisfying. Paul Graham's observation in 'Do Things That Don't Scale' is that the feedback you get from directly engaging your earliest users is the best you'll ever receive — and that holds for churn signals too. When you're visiting users in their offices or watching them use your product, you notice the hesitation, the workarounds, the quiet frustration that never makes it into a support ticket.

The practical implication: before you build a churn-prediction model, build a habit of weekly one-on-one conversations with a rotating set of users. Not surveys — conversations. Ask what they did in the product last week, not how they feel about it. Behavior is harder to falsify than sentiment.

Segment your users early and watch each cohort differently

Not all users churn for the same reason, and conflating them leads to interventions that help no one. A useful early framework is to split users into three buckets: power users who get immediate, repeated value; casual users who signed up out of curiosity; and obligated users who adopted your product because someone above them required it. Each cohort churns differently and at different rates.

Power users rarely churn silently — when they leave, it's usually because a specific feature broke a critical workflow, a competitor solved a concrete pain point better, or your pricing crossed a threshold they couldn't justify to finance. These are the churners worth a personal call, because the reason is almost always recoverable information.

Casual users churn by drifting. They never developed a habit with your product because you never gave them a reason to return unprompted. The fix here is activation, not retention — they churned at the moment they didn't reach their first 'aha' moment, even if they didn't formally cancel for months. Obligated users are a different problem: they're retained by org chart, not value. Their churn is a lagging indicator of a failed enterprise sale, and no amount of feature work fixes it.

The behaviors that reliably predict departure

Across product categories, a few behavioral patterns show up repeatedly as leading indicators of churn. First is declining frequency in the core action — whatever the single thing is that users do when they're getting value (sending a message, running a report, publishing a post). When that frequency drops more than 50% week-over-week without an external cause like a holiday, it's a warning sign. Second is failure to complete setup or integration steps that unlock deeper value. A user who never connected their data source, never invited a teammate, never set up their first automation — they're in a pre-churn state from day one.

Third is support silence combined with low product usage. This sounds counterintuitive, but users who submit zero support tickets and use the product minimally are more at risk than users who complain loudly. Loud users are engaged. Silent minimal users have mentally already left.

Fourth — and this one founders often miss — is when a user starts asking questions that imply they're evaluating alternatives. 'Can you export everything in CSV?' or 'How do I get all my data out?' are questions a satisfied user doesn't ask. These are migration-prep questions, and they're an extremely reliable signal that you have a short window to change their mind.

What to do once you've identified a likely churner

The biggest mistake is sending an automated 'we miss you' email. That signals to the user that your system noticed their absence but no human did — it confirms their sense that they're just a row in your database. The intervention that actually works is a personal message from a founder or a real human, referencing something specific about how they used the product.

The goal of that outreach is not to make a sale. It's to run a diagnostic. Ask what's changed in their workflow, whether the problem they originally signed up to solve has changed, and what would have to be true for the product to be genuinely useful again. Even if they churn anyway, that conversation is worth more than any retention dashboard because it tells you whether you have a product problem, a positioning problem, or a customer-fit problem.

Garry Tan has noted the trap founders fall into of imitating big companies' indifference to individual users in the name of seeming 'professional.' That impulse is expensive. At the early stage, personally chasing a churning user is not a sign of desperation — it's a competitive advantage that larger companies structurally cannot replicate. Use it aggressively while you still can.

Building a lightweight churn radar without a data team

You don't need a machine learning model to catch most churn at the early stage. What you need is a simple weekly ritual: pull the list of users who were active two weeks ago and aren't active this week, and personally reach out to the ones who matter most to your business. At fewer than 500 users, this is entirely manual and takes less than an hour. At 500–2,000 users, a simple SQL query against your event log surfaces the at-risk segment quickly.

The metric you're tracking is not 'days since last login' in isolation — it's deviation from that user's own baseline. A user who normally logs in twice a week and hasn't logged in for ten days is more alarming than a user who only ever logs in once a month and just logged in yesterday. Normalize engagement to individual user patterns, not aggregate averages.

As you scale, the habit of close user observation should be institutionalized, not abandoned. Assign specific users to specific team members. Run monthly calls with power users. Create a customer advisory group from your most engaged cohort. The founders who scale past churn are usually the ones who never fully delegated user understanding to a tool.

“The feedback you get from engaging directly with your earliest users will be the best you ever get.”

— Paul Graham, source

The one thing to do

This week, pull every user who was active 14 days ago and hasn't returned, personally message the five who matter most to your business, and ask one question: what changed?

Frequently asked questions

What's the single most predictive metric for churn?

Declining frequency in the core action — the one thing users do when they're getting value — is the most reliable leading indicator across most product categories. When that drops sharply relative to a user's own baseline, intervention is warranted within days, not weeks.

Should I build a churn prediction model?

Not until you have enough users that manual monitoring is impossible and enough historical churn data to train on. Before that point, a weekly review of disengaged users and direct outreach will outperform any model because the interventions require human judgment anyway.

How do I tell the difference between a user taking a break and a user who's churning?

Reach out and ask. A user on a genuine break will respond quickly and give a plausible reason. A user who's mentally left will either not respond or give a vague, non-committal answer. The response behavior itself is the diagnostic.

Is it worth trying to win back users who have already canceled?

Sometimes, but the more valuable outcome is the exit interview. Understanding exactly why they left — in their own words, not a multiple-choice survey — gives you information that improves retention for future users, which compounds over time far more than any single win-back campaign.

Sources

More playbook answers · Growth Prophet home