Last Updated on August 14, 2026 by Taya Ziv
Picture a board meeting that is going extremely well.
The founder puts up the growth slide. Trial conversion is up. Revenue per customer is beating every consumer subscription benchmark anyone in the room has seen. Somebody says the words “category leader” out loud and nobody laughs. The deck is honest, the numbers are real, and everyone leaves believing the same thing.
Now run that same company forward one year and pull the cohort report on its annual subscribers. Roughly four out of five of them are gone.
That is not a hypothetical. That is the middle of the largest public dataset on consumer subscription apps, and it has been sitting in the open since the complete report landed on 14 March 2026 while most of the industry kept quoting the flattering half of it.
Two numbers that should not belong to the same product
On 14 March 2026, RevenueCat published the complete 2026 State of Subscription Apps report, a benchmark that runs past 330 pages, built on more than 115,000 apps and upward of a billion transactions.
Before the numbers, the thing you need to hold in your head while you read them. RevenueCat sells subscription billing and monetization infrastructure to app developers, which means it sells tooling aimed at the exact problem this report measures. The dataset is not a census of the app market. It is the apps that run on RevenueCat’s own platform, a point TechCrunch made explicitly in its coverage: the analysis covers the subscription app providers that use RevenueCat’s tools, over 75,000 developers generating more than $11 billion in revenue annually. The $16 billion figure that travels with the report is the revenue total across the whole dataset, not an annual run rate, and the two get used interchangeably in the coverage. RevenueCat also says it processes about a billion dollars a month, which it puts at roughly a fifth of global subscription app revenue. That last one is the company’s own estimate of its own market share, so hold it loosely.
None of that makes the data wrong. It makes it a large, self-selected sample published by a vendor with a stake in the conclusion, which is a different thing from a market census and should be read as one.
Here are the two numbers.
AI-powered apps generate 41 percent more realized lifetime value per paying customer in year one than non-AI apps. Thirty dollars and sixteen cents against twenty one thirty seven at the median. In the first month the gap is 39 percent, $18.92 against $13.59. AI apps get people to start a trial 52 percent more often, 8.5 percent against 5.6. Download to paid conversion runs 20 percent higher, 2.4 percent against 2.0. On every measure of getting a human being to open their wallet, AI is winning by a distance that would normally make a category unfundable to compete in.
Now the second number. Twelve month retention on annual plans is 21.1 percent for AI apps against 30.7 percent for everyone else. On monthly plans it is 6.1 percent against 9.5 percent, which means about 94 percent of monthly AI subscribers are gone inside a year. Refunds run 4.2 percent against 3.5 percent, with an upper bound of 15.6 percent against 12.5.
One correction before we go on, because it cuts against a headline you may have already read, and because I put the wrong version of it in an earlier draft. Weekly plans are no exception to any of this. RevenueCat’s own key takeaway is one blunt sentence: retention is lower for AI apps across all subscription durations. TechCrunch reported the opposite, that AI led on weekly at 2.5 percent against 1.7, and I believed it long enough to write it into an earlier draft of this piece. The report publishes two different weekly pairs in two different places, 1.2 against 1.7 in one chart block and 2.5 against 3.4 in a chapter summary, and it looks like AI’s 2.5 got crossed with the wrong comparison number somewhere on the way out. Both pairs point the same direction. PPC Land’s write-up had it right at 1.2 against 1.7, and I only caught that by going back to the report page instead of trusting the coverage. Keep that in mind for everything below. Every number here came out of a document that a lot of people wrote about and fewer people opened.
Your funnel is measuring curiosity
Here is the part that took me a while to see, and it is the whole reason I wanted to write this down.
The obvious reading is that the early advantage decays. It does not. That is the first thing I got wrong. The revenue gap actually widens across the year, from 39 percent at month one to 41 percent at month twelve. The AI apps that keep a customer keep a more valuable one.
Which makes the retention number worse, not better. Both things are true at once: AI apps lose customers far faster, and the survivors pay enough to hold year-one value per payer ahead anyway. The churn is not showing up in your revenue per payer because price is covering for it. You are running a leakier bucket and a higher tap, and the tap is what your dashboard is built to show you.
So you have a dashboard that is technically accurate and upside down where it counts. It is loudest exactly where the signal is weakest. I have written before about how the metric Silicon Valley loves most quietly became the thing it lies to itself with, and this is the same disease with better graphics. The number is not wrong. The number is incomplete.
And the tell is in why people leave. In RevenueCat’s Google Play churn survey, the reasons voluntary cancellers give cluster into two buckets: cost, at 25 to 45 percent depending on category, and not enough usage, at 26 to 40 percent. Technical problems account for 3 to 7 percent. Almost nobody churns because your product broke. They churn because they stopped opening it. That is not a reliability problem you can engineer around. It is a habit problem, and habits are much harder to build than features.
There is a second tell in the plan mix, and it is more interesting than the version I first wrote. Monthly plans are 59.8 percent of subscriptions sold by AI apps against 26.2 percent for non-AI. My first instinct was that AI buyers are refusing to commit at more than double the rate. That is not what the data says. Non-AI apps spread across weekly at 30.3 percent, monthly at 26.2, and yearly at 41.8, so a big chunk of the non-AI mix is a shorter commitment than monthly, not a longer one. The comparison that survives is the annual one. Non-AI apps put 41.8 percent of their subscriptions on yearly plans. AI apps put 24 percent. The report sizes that gap at 17.8 percentage points. Yearly is where an app gets to prove itself over a real horizon, and AI apps are largely not selling it.
The machine may be fooled too, and it is spending your money
This is the part that should worry anyone running paid acquisition, and I want to be careful about how strongly I put it, because it is the one place where I could not get to a primary source.
Here is what is documented. Google’s own guidance for target ROAS bidding in App campaigns tells you to set your target against your biddable in-app event’s conversion window, and to evaluate performance over a 14 to 30 day window. The measurement horizon these systems are tuned around is short by construction, because it has to be. Nobody can run a bidding algorithm on a signal that takes twelve months to arrive.
Here is what is inference, and whose. PPC Land’s read of the RevenueCat data is that higher early revenue per payer makes AI app users look attractive to value-based bidding models anchored to short LTV windows, and that predicted-value inputs for these apps need to account for a steeper decay curve than non-AI benchmarks would suggest. That is an ad-tech trade publication’s analysis, not a RevenueCat finding and not a documented platform mechanism. I went looking for Google or Meta documentation stating a specific 30 or 60 day lifetime value prediction window and did not find one.
Treat it as a hypothesis worth testing on your own account rather than a law of physics. The shape of it is plausible enough to check: in your first 30 days you look spectacular, because that is where your entire advantage lives. If the algorithm reads that as a high value user and bids up, it goes and buys you more of exactly that person. Then month twelve arrives and the annual cohort is at 21 percent.
The failure mode, if it is real, is that you did not just misread your own dashboard. You handed the misreading to a machine with a credit card and told it to go faster. That is how a company posts good blended numbers for three quarters and then discovers its unit economics were negative the whole time, which is roughly the story of the AI startup that was growing fast right up until the autopsy found the problem in its margins.
Rebase everything to Month 3
The useful fix comes from a different dataset and a different kind of company, and both of those matter.
In September 2025, a16z published a piece called Retention Is All You Need by Santiago Rodriguez and Alex Immerman, arguing that the leading AI companies do not have a retention problem so much as a measurement problem. a16z is a venture firm with money in the AI companies whose retention it is benchmarking, and the thesis lands in a direction that suits that position. I think the framing is slightly generous and mostly right, and knowing who is doing the framing makes that easier to judge, not harder.
Note the switch, because the article you are reading just made it. RevenueCat measures consumer mobile and web subscription apps where the median payer is worth about $30 in year one. a16z looked at dozens of mostly self-serve AI companies above $1 million in ARR, which is a smaller, richer, more business-flavoured population. The M3 framework below was built on the second group. I think it travels, but it is a framework crossing a boundary, and you should know that before you apply it.
Their finding, after analyzing hundreds of AI companies, is that you should stop anchoring retention and acquisition math to Month 0 and rebase it to Month 3. The reasoning is clean. Every AI cohort has what they call tourists, people who signed up to look around, and those tourists mostly wash out by month three. The curve most often begins to flatten there. So M3 is not an arbitrary checkpoint, it is the first honest headcount of your actual customer base.
From that, three changes worth more than any growth experiment you have queued up:
Report M12 divided by M3, not M12 divided by M0. This one ratio tells you how the people who genuinely chose you behave over a real year. a16z calls it an early predictor of long-term retention quality and a leading indicator of net dollar retention above 100 percent. Put it on the board slide. If it is ugly, you have found the problem while you can still afford it.
Track cost per retained customer at M3, not cost per acquisition. CAC at signup is the most flattering and least informative number in your company. Cost per user still standing at month three is the number that tells you whether more spend produces a business or a bigger leak. It is the metric a16z says it tracks on its own portfolio.
Stop compressing the trial. The data here is close to comedy. Trials of 17 days or more convert at 42.5 percent against 25.5 percent for short ones, a 17 point gap, and yet nearly half of all apps now run trials of four days or less, a share that rose 4.4 percentage points year over year. Meanwhile 55 percent of three day trial cancellations happen on day zero. The industry has collectively decided to shrink the window for proving value to the exact moment when a new user understands the product least. If your thing takes two weeks to become a habit, a three day trial is not a growth tactic. It is a decision to sell only to people who were already sold.
The honest counterargument
I do not want to hand you a doom narrative, because the data does not support one.
Some AI cohorts show what a16z calls a smiling curve, where churned or quiet users come back as the product gets better. a16z says ChatGPT’s curve is the perfect example. That is a genuinely new behaviour and it did not exist in older subscription software, where a lost customer was a lost customer. It is plausible that AI companies end up with better long term retention than anything before them, because the product a user rejected in March is not the product waiting for them in September.
The catch is that this is an argument for patience, not for ignoring the curve. A smiling curve and a dying curve look identical for the first several months. The only way to know which one you are on is to measure past the tourists, which almost nobody is doing, because the Month 0 numbers feel so good that looking further seems pessimistic.
And there is a supply problem making all of this louder. According to Appfigures data cited in RevenueCat’s report, new subscription app launches went from roughly 2,000 a month in January 2022 to 14,700 in January 2026, about seven times, with the steepest climb from early 2025, coinciding with the rise of AI-assisted development tools. Yet apps launched before 2020 still earn 69 percent of all subscription revenue, while everything launched in 2025 or later earns 3 percent. Building got easy and mattering did not, which is the same lesson as the fastest way to build a startup nobody wants, now with a number attached.
What I would actually do on Monday
Open your analytics and pull one cohort from twelve months ago. Find how many of them are still paying. Divide that by how many were still paying at month three. Write that single number on a sticky note and put it on your monitor, because it is the closest thing you have to the truth about your company.
If it is strong, you have something durable and you should spend into it much harder than your CAC math currently allows. If it is weak, you have not failed. You have found out early, in a market where most of your competitors are going to find out in a Series A data room instead.
The uncomfortable summary is that AI gave every founder the best first impression in the history of consumer software, and first impressions were never the hard part. Getting someone to try you was always the cheap problem. Getting them to still be there in a year is the whole company.
