Recurring revenue and the problem of customers who leave

Suppose a coffee roaster in Can Tho sells a monthly delivery box. It signs up 100 new subscribers every month, yet after a year the base still hovers around 400. If 25% of subscribers leave each month, 100 newcomers only just replace the leavers at 400. The roaster and every number here are invented, for arithmetic only. This post reads McCarthy, Fader and Hardie (2017) on valuing subscription businesses to ask two things: where does churn show up in the data, and when is keeping a customer worth more than finding a new one?

What is recurring revenue, and where does churn show up in the data?

It is revenue from customers who pay a repeated fee for access to a product or service. McCarthy, Fader and Hardie (2017) call such firms subscription-based businesses. They describe the model as long the home of newspapers, magazines and telecoms, now spreading to consumer software, food preparation (Blue Apron) and business-to-business software as a service, where firms want their revenue to be more predictable (McCarthy et al., 2017). The library has a subscription model term and a recurring revenue framework. Its three forces framework adds technology and customer preferences; in the paper I read, only predictable revenue is stated outright.

Churn shows up in one table. Rows are the month customers were acquired, columns are calendar months, and each cell is how many customers from that cohort still pay. The monthly churn rate is customers lost divided by customers at the end of the previous month (McCarthy et al., 2017). If you are weighing pay-per-use rather than a flat fee, see the post on Rolls-Royce charging per running hour.

How much does a coffee roaster lose when customers leave? An invented example

At 25% monthly churn, each new subscriber only just breaks even. Every figure below is invented. A box sells for 250,000 dong a month, with a contribution margin of 100,000 dong after beans, packing and delivery. The acquisition cost of a new subscriber, covering advertising and samples, is 400,000 dong, so four months of margin are needed to recover it.

If monthly churn is a constant c, a subscriber pays for 1/c months on average. If A customers join each month, the base heads for A/c. The table ignores discounting, tax and the fact that real churn is not constant.

The last column assumes 100 new customers a month. At 25%, more advertising raises revenue but not profit. Cutting churn from 25% to 10% gives a base 2.5 times larger from the same 100 sign-ups. This is a very simple form of customer lifetime value; Gupta and colleagues (2006) summarise models of that kind as tools for allocating marketing resources between acquisition, retention and cross-selling (Gupta et al., 2006).

Is one churn rate for every customer wrong?

It can be, and it tends to undervalue existing customers. McCarthy and colleagues note that earlier models, including the pioneering work of Gupta, Lehmann and Stuart (2004), assume a constant retention rate. By their account, Fader and Hardie (2010) show that this ignores the way retention within a cohort rises over time, so the remaining value of current customers is underestimated (McCarthy et al., 2017). I have not read either paper.

McCarthy and colleagues build a model in which customers differ and churn changes with tenure, and apply it to Dish Network, a US pay-TV provider. They compare two imaginary customers: Robin, who has just joined at the end of the first quarter of 2015, and Larry, a customer of ten years. The model puts Robin's expected remaining lifetime at 5.5 years and Larry's at 9.4 (McCarthy et al., 2017).

Source: table 8 of McCarthy, Fader and Hardie (2017), accepted version (McCarthy et al., 2017). These are model estimates from aggregated figures, not customer-level data. What a business owner can take away is the method: do not report one blended churn rate, split it by tenure.

Keep or acquire: what do Dish and Sirius XM say?

It depends, on these two examples. The authors apply one model to two different companies. At the end of the first quarter of 2015, Dish earned 88.72 USD per customer per month and spent 716.46 USD to win one; Sirius XM earned 16.72 USD and spent 82.06 USD (McCarthy et al., 2017). In the authors' description, Dish is mature and Sirius XM is growing fast; the two also differ in cost structure and customer profile.

On sensitivity, they hold two processes fixed, redraw the parameters of the third 500 times by bootstrap, and take the 95% interval for Dish's estimated share price. It is ±0.2% when the acquisition process varies, ±3.4% for retention and ±3.0% for revenue per subscriber (McCarthy et al., 2017). That measures uncertainty in the estimate, not "raise retention by 1% and value rises by this much".

In their decomposition, the authors write that Dish's value depends on, and is more sensitive to, its ability to retain existing customers than to acquire new ones, while Sirius XM's value depends more on acquiring customers and earning a high return on that investment (McCarthy et al., 2017). For an owner, there is no universal answer to "keep or acquire". Sirius XM's acquisition cost per customer is roughly one ninth of Dish's, though each customer also pays far less per month.

Is the Starbucks loyalty programme recurring revenue?

No. The fiscal 2025 10-K (year ended 28 September 2025) describes the Starbucks Card as designed to raise how often cardholders visit stores, in part through Starbucks Rewards; in nearly all markets, including the US and Canada, customers who register a card are enrolled automatically (Starbucks Corporation, 2025). The accounting note calls it primarily a spend-based loyalty programme: purchases earn Stars, enough Stars earn a free product, and Stars generally expire after six to twelve months depending on the market (Starbucks Corporation, 2025). Stored value cards and the loyalty programme together carried 1,751.7 million USD of deferred revenue at year end (Starbucks Corporation, 2025). See also the library's Starbucks case.

The difference from a subscription is that the 10-K describes no repeated fee or cancellation step, so churn cannot be observed directly: customers simply stop visiting. McCarthy and colleagues say their framework fits only subscription businesses, where leaving is observed; for non-contractual settings they admit it is unclear what a company would need to disclose for an outsider to estimate the cohort table (McCarthy et al., 2017). In the parts of the 10-K I searched, I found no membership count or lapse rate (Starbucks Corporation, 2025).

My own reading, not the source's: the Stars expiry date is the one clock that turns a customer drifting away into a measurable event. A small café has to set its own rule, say sixty days without a purchase counts as gone, and then track by cohort. For the same company, When history fails reads how Starbucks separated transactions from average ticket during the pandemic; the general idea sits under loyalty programme.

What should you check before moving to recurring fees?

It suits you if customers need the product regularly, you can see who is still paying, and the break-even churn rate sits well above what you actually experience. It does not suit you if demand is erratic or you cannot tell a lapsed customer from one on a break. That is my inference.

1. Write down what counts as leaving: cancelling or not renewing, or no purchase within a set number of days.
2. Build a cohort table by month of acquisition and split churn into three bands: the first three months, months four to twelve, and after a year.
3. Compute break-even churn as contribution margin divided by acquisition cost and compare it with your real figure.

Going deeper: what I read and what I did not

The library has a framework called "Six recurring-revenue models", but the published version holds only a question and an origin, with no models listed, and McCarthy, Fader and Hardie contains no such classification. So this post neither counts nor compares models by name; it works from what the two sources do say: a repeated fee, churn and cohort data.

I read McCarthy, Fader and Hardie (2017) in full, in the accepted version held openly at LBS Research Online, so I have no print page numbers and figures may differ slightly from the printed article. The Dish and Sirius XM numbers are model estimates using data to the first quarter of 2015, for one US pay-TV firm and one satellite radio firm. Table 5 gives Dish's acquisition cost per customer as 716.46 USD, while the Robin and Larry comparison uses 854 USD; I use table 5.

For Gupta and colleagues (2006) and Dorotic, Bijmolt and Verhoef (2012) I read only the abstracts. For Starbucks I read only the fiscal 2025 10-K.