Customer Churn: How to Calculate It, Model It, and Keep More Customers

Imagine your medical spa or clinic gains 100 new clients this quarter. Great news! But what if 80 existing clients stopped coming back in the same period? Many business owners focus on attracting new customers while losing just as many through the back door. That loss is customer churn, and it quietly caps how fast you can grow no matter how much you spend on acquisition.

Here is the short answer. Churn is the share of customers who stop doing business with you in a period, and the right formula depends on your model: subscription businesses count cancellations, while repeat-purchase businesses have to define churn as inactivity past a set window. A single blended churn number hides most of what matters, so track churn by cohort, estimate survival properly when many customers are still active, and use a simple prediction model to decide who gets attention. Then test whether your fixes actually reduce churn before you scale them. Every number below is a made-up example so you can follow the arithmetic.

1. What is the right churn formula for your business?

For a subscription or membership business, churn has a clean event: a cancellation. The basic formula is customers lost during the period divided by customers at the start of the period. For example, if your med spa membership program had 250 members on March 1 and 10 of them cancelled during March, monthly churn is 10 ÷ 250 = 4%. If you are growing fast, new members who sign up and cancel in the same month distort that formula. Say you also added 100 members in March and 4 of them cancelled before month end. A cleaner measure is churn per member-month: 14 cancellations divided by roughly 300 member-months of exposure (250 existing members plus 100 new members counted as present for half the month) gives 14 ÷ 300 = 4.7%.

Also, never multiply monthly churn by 12. Monthly churn of 4% means 96% of members stay each month, and 0.96 to the twelfth power is about 0.61. So annual churn is about 39%, while multiplying by 12 would have told you 48%.

For a business without subscriptions, such as a clinic, a med spa without memberships, or an ecommerce brand, nobody cancels. They just stop coming back. You have to define churn as a lapse. Look at repeat customers and measure the gap between their purchases or visits, then pick a window that covers most of those gaps, for example the point by which 90% of second visits have happened. If that point is 150 days, a client with no visit in 150 days counts as lapsed. Then, as an example, if 1,200 clients visited in the first quarter and 660 of them had no visit in the 150 days after their last first-quarter visit, the lapse rate is 660 ÷ 1,200 = 55%. Write the window down and keep it fixed. In healthcare, base it on the normal recall or follow-up interval, and count patients who formally transferred out as lost.

All of this is an extreme over-simplification. It is a helpful starting place, but to gain real insight you need modeling that tells you the effect sizes of the variables that drive your churn rate. Which of those can you control? Which levers can you actually pull? That is the true feedback.

2. Why does a single monthly churn number mislead you?

A blended churn rate is an average of very different customers, and averages move when the mix moves. New customers almost always churn faster than customers who have stayed a year. Here is an example. Long-tenure customers churn at 2% a month and customers in their first three months churn at 10% a month. In month A you have 900 long-tenure customers and 100 new ones, so blended churn is (18 + 10) ÷ 1,000 = 2.8%. In month B you run a big acquisition push and have 900 long-tenure customers and 400 new ones. Blended churn is now (18 + 40) ÷ 1,300 = 4.5%.

Churn looks like it got much worse, but neither group changed at all. You simply have more new customers in the mix. The reverse trap is just as common: cut acquisition and churn appears to improve. The fix is to measure churn within groups that share a starting point, which is what cohorts do.

3. How do you read a cohort retention curve?

A cohort is every customer who started in the same period. You track what share of each cohort is still active one month later, two months later, and so on. Here is an example with three monthly cohorts of 200 new members each, where the March cohort started after an onboarding change:

• January cohort (200 members): 160 active after month 1 (80%), 140 after month 2 (70%), 130 after month 3 (65%)

• February cohort (200 members): 170 after month 1 (85%), 154 after month 2 (77%), 146 after month 3 (73%)

• March cohort (200 members): 176 after month 1 (88%), 162 after month 2 (81%), month 3 not yet observed

Read across a row to see the shape. The January cohort loses 40 of 200 in month 1 (20%), then 20 of the remaining 160 (12.5%), then 10 of the remaining 140 (7.1%). Losses shrink each month, so the curve is flattening, and where it flattens tells you the size of your loyal core. A curve that never flattens points to a product or value problem; one that drops hard early and then flattens points to onboarding.

Read down a column to see whether things are getting better. Month-1 retention went from 80% to 85% to 88%, which is evidence the onboarding change is working. Before you celebrate, check that the cohorts did not also differ in channel, offer, or season, because those can move a column too.

4. How does churn connect to customer lifetime value?

If churn were constant, the expected customer lifetime would be 1 divided by the churn rate, and lifetime value would be monthly gross margin per customer times that lifetime. Here is an example. A membership costs $120 a month at a 60% gross margin, so each member contributes $72 a month. At 4% monthly churn, expected lifetime is 1 ÷ 0.04 = 25 months, and LTV is $72 × 25 = $1,800. Cut churn to 3% and lifetime becomes 1 ÷ 0.03, about 33.3 months, so LTV is $72 ÷ 0.03 = $2,400. A one-point drop in churn added $600, or 33%, to the value of every member. If it costs $450 to acquire a member, the LTV to CAC ratio moves from 4.0 to about 5.3.

That shortcut is fine for a first pass, but real cohort curves are not constant. They drop early and flatten. A better LTV sums the cohort curve: margin in each month times the share of the cohort still active, discounted for the time value of money. That is the number to use for acquisition cost targets. If you have to defend those targets, my post on how to explain marketing ROI to your CFO walks through the conversation.

5. What is survival analysis, and why does Kaplan-Meier matter?

The hard part of measuring churn is that many customers have not churned yet. A member who joined last month has only one month of history. You do not know how long she will stay. Statisticians call her "censored." If you treat censored customers as if they stayed forever, you overstate retention. If you throw them out, you understate it. Survival analysis exists to handle exactly this, and the Kaplan-Meier estimator is the plain-English workhorse: at each month, look only at the customers who were still around and observable at the start of that month, and calculate what share of them made it through.

Here is a tiny example with 10 customers. In month 1, 2 of 10 churn, so survival is 8 ÷ 10 = 80%. One customer joined recently and has only one month of history, so she drops out of the count without churning. In month 2, 7 customers are at risk and 1 churns, so survival is 80% × 6/7 = 68.6%. In month 3, 6 are at risk and nobody churns, then 2 more customers reach the end of their observed history. In month 4, 4 are at risk and 1 churns, so survival is 68.6% × 3/4 = 51.4%. The remaining 3 are still active. A naive count that treats the 3 censored customers as survivors says 6 of 10, or 60%, made it through month 4. A naive count that drops them says 3 of 7, or 42.9%. Kaplan-Meier lands at 51.4% because it uses each customer for exactly as long as you observed them.

Once you have survival curves, you can compare them by group, such as customers from paid search versus referrals, and a Cox regression goes a step further by estimating how much each factor raises or lowers the risk of churning at any point in time while holding the others constant.

6. How do you predict which customers will churn?

The most useful starting model is a logistic regression. The outcome is simple: did this customer churn in the next 60 or 90 days, yes or no. The inputs are signals you already have. The ones that tend to matter are tenure, days since the last visit, purchase, or login, the trend in usage (dropping is worse than low), failed payments, support tickets or complaints, whether the customer finished onboarding, the plan or service they bought, the acquisition channel, and whether they came in on a deep discount. The model gives every customer a probability of churning and tells you how much each signal moves the odds.

Odds are easy to misread, so here is an example. Say the model finds that no login in the last 14 days triples the odds of churning. If a customer's baseline churn probability is 5%, her odds are 0.05 ÷ 0.95, about 0.053. Triple that and the odds are about 0.158, which converts back to a probability of 0.158 ÷ 1.158, or 13.6%. Tripling the odds took her from 5% to 13.6%, not to 15%. Always test the model on customers it did not learn from. A useful check is lift: if the riskiest 10% of customers contain 35% of the people who actually churn next quarter, the model concentrates churners 3.5 times better than picking at random. Tree-based models such as gradient boosting can predict a bit better, but logistic regression is easier to explain to your team.

One caution: a prediction model tells you who is likely to leave, and it does not prove that changing a signal will keep them. That takes a test.

7. What should you do once you know who is at risk?

Start with the cohort curve. If most losses happen in the first 30 to 60 days, fix onboarding and the first experience: a clear first-week plan, a check-in after the first visit or order, appointment and replenishment reminders, and a reason to come back for the second visit. In healthcare, keep this to scheduling, recall reminders, and service follow-up, and keep clinical guidance with the clinician.

Next, target by both risk and value. A high-risk, high-value customer deserves a personal call or a meaningful save offer. A high-risk, low-value customer may only justify an automated message. Then test the save offer with a holdout group. Here is an example. Take 2,000 high-risk members and randomly split them. The 1,000 who get the offer churn at 15% and the 1,000 control members churn at 20%, so the offer saved 200 minus 150 = 50 members. If 600 members accept an offer that costs you $40 each, the cost is $24,000. If each saved member is worth the $1,800 LTV from earlier, the return is $90,000. Without the control group you would have credited the offer with members who were never going to leave.

Finally, look at churn by segment. My guide to customer segmentation with data shows how to build segments you can measure this way.

8. When do you need an expert to model churn?

You can calculate a basic churn rate yourself, and you should. Bring in help when the number moves and nobody can say why, when your customer mix is shifting because of acquisition pushes, when you need an LTV your finance team will sign off on, when most of your base is still too new to judge, or when you want to know which drivers are actually worth fixing. That is the core of my marketing analytics work, and for clinics, med spas, and telehealth brands it ties directly into my healthcare and telehealth marketing practice.

If churn is eating your growth and you want a clear read on where customers are leaving, why, and what to fix first, you can book a free 30-minute introductory call here.

Dr. Linda Orr, PhD, is the founder of Orr Consulting, a fractional CMO and marketing analytics practice serving brands from $5M to $500M in revenue. She built Upwork's first enterprise marketing mix model, covering an $88M annual media portfolio, and spent about 20 years as a marketing professor before moving fully into consulting.