Making the most of the fabled recurring revenue
Dialing in churn, CAC, and CLV are existential. The highest levers are new users and churn/retention. A company growing new subscriptions at 50% y/y will be net neutral on overall growth with a 5% monthly churn rate, normal for startups.
OpenAI, Anthropic, Starlink, Cursor, Samsara, Trimble
Churn/retention is the most sensitive, non-linear lever for CLV growth. It’s also a canonical metric for PMF. See here for more on how to boost growth.
Overall churn rate is at best misleading for understanding underlying trends in retention and CLV. Use cohort-specific measures.
Monthly churn rate data from Recurly’s 2,200 merchants that use subscriber models (covering ~58M unique subscribers):
Churn dynamics of 25 young subscription-based brands:
Subscription- vs Usage-based Pricing in Frontier AI
One mistake/untenable equilibrium that hot AI startups like Cursor, Anthropic, Windsurf have recently been running is the unlimited usage subscription model that’s deeply unaligned with their underlying business.
It’s a fixed revenue, highly variable costs (which are mostly deeply negative). Their most profitable users (i.e. those that barely use the service) churn first. Their remaining power users make tokens go burrrrrr for no additional cost to themselves.
In the long-term, such startups must switch to usage-based pricing, hike the subscription pricing above costs, get acqui-hired, or find other monetization streams (e.g. Cursor selling access to their proprietary coding data to frontier labs or OpenAI figuring out monetization of free users with advertising).
https://ethanding.substack.com/p/ai-subscriptions-get-short-squeezed
https://ethanding.substack.com/p/windsurf-gets-margin-called
https://hypersoren.xyz/posts/smart-squeeze/
https://abovethecrowd.com/2012/09/04/the-dangerous-seduction-of-the-lifetime-value-ltv-formula/
https://recurly.com/research/churn-rate-benchmarks/
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4722115
https://www.linkedin.com/pulse/blue-aprons-q3-18-results-cac-moves-higher-retention-trends-mccarthy
CAC is cheaper for market share leaders, but retention isn’t effected by positioning
https://www.michaeldempsey.me/blog/2025/10/03/sequencing-vs-equal-odds-applied-research/
Young customers: it also really matters when you want to predict well for young cohorts, who may be affected by that seasonality for the first time, making it hard for the model to know whether that seasonality is indeed seasonality or simply a part of the baseline goodness of the cohort, or relatedly, whether it is a seasonal effect that will recur each year or a non-seasonal effect that will not. The latter issue is just as problematic because in truth, statistical identification of seasonal versus non-seasonal effects is not possible unless you have observed at least two years' worth of data. At many growing businesses, the vast majority of the active customer base is less than 2 years old, compounding the issue.
What is typically done to account for these issues? Probably the most common solution is to simply pool all the customers who are 2 years or younger together into a single model. The problem with this is that such pooling can often lead to ill-fitting projections when there are cross-cohort dynamics, as it means we’re pooling “apples with oranges.” Separating out these various effects while still allowing young cohorts to be fundamentally different from older ones is a real challenge. And an important one.
Half of the auto-renewal contract takers continue to a full-price subscription while rarely using it. However, consumers anticipate their own inertia. When shown an auto-renewal promotion, 24–36% who are apparently on the fence for trying it don’t sign up. Over the subsequent two years, auto-renew contracts lower the overall subscriber share by ~10%.
Overall, auto-renew briefly earned more, but by ~8 months the revenue difference wasn’t statistically distinct from zero difference and by 24 months the difference completely disappeared.
Tenure effect (dominant): Within any cohort, high-risk customers leave early; the remaining base is hardier, so per-period hazard falls with tenure. As survivors accumulate, the overall, tenure-weighted average churn drifts down. In the paper’s data, hazard drops steeply from early months to month 12 across most brands.
Acquisition growth effect (short-run bumps up): Big new cohorts inject lots of low-tenure, high-hazard customers, briefly nudging overall churn up next month—but this is usually temporary and outweighed by the tenure effect.
Cross-cohort effect (often worse over time): Later cohorts can be less sticky (worse product-fit or more promo-driven signups), so their cohort-level month-one/month-three churn can rise over calendar time even while the overall rate falls. The Hubble example shows exactly this: aggregate churn trends down while month-one churn trends up.
ChatGPT users spend 3x more time a day on the app than 21mos ago.
ChatGPT revenue has grown 10x+ annually do $3.7B
I analyzed the engagement of leading AI, consumer and enterprise apps over the last decade. Turns out increasing engagement is incredibly rare (kudos Duolingo moving up from 50 to 60%)
The most engaged apps are those with the most distribution
https://apoorv03.com/p/why-meta-and-google-may-win-consumer
