Which event should a subscription app optimize for?

Meta wants roughly 50 events per ad set a week. Google wants ten users a day. Your purchase event fires 17 times. Here is what to do about that.

The event you optimize for decides which users you get. Send installs and the platform finds people who install. Send trial starts and it finds people who start trials. Send purchases and it finds buyers, but only if it sees enough of them to learn.

That last clause is where most subscription accounts break. The purchase event is the one that matches your economics. It is also the one that fires least often. So teams either pick it and starve the algorithm, or drop back to installs and fill the funnel with people who never open the paywall.

There is a rule that resolves this. Before the rule, the numbers.

What each platform needs before it can learn

Checked 23 September 2026 against each platform’s own help pages. These change. The date matters more than the number.

Two things fall out of that table.

The thresholds are not the same across platforms, so “50 a week” is a Meta number, not a law. I use it as the floor for how many campaigns an account can carry, and it is a floor for learning, not for trusting a result: an ad needed 50 or more purchases before its ROAS predicted the next week, and even then only moderately.

And Apple Ads sits outside this decision entirely. The event question is a Meta, Google and TikTok question.

The event you pick is the user you get

The platform does exactly what it is told. Move an account from purchase optimization to trial optimization because purchase volume per ad set is too low, and cost per trial falls. So does trial to paid. Cost per paying subscriber goes up. The platform found people who start trials, and some of those people start trials everywhere.

RevenueCat’s 2026 data shows why this bites hardest on short trials. Among apps with three day trials, 55.4% of trial cancellations happen on day 0, and 84% by day 1. A trial optimized campaign can look excellent on Tuesday and be worthless by Thursday, and the platform never sees the difference because the cancellation is not an event it optimizes against. Whether the trial belongs in the funnel at all is its own question.

Why the deepest event is usually the wrong first choice

Purchase is the event that matches your P&L. If you could always optimize to it, you would.

You usually cannot, for two reasons.

Volume. Take an illustrative subscription app doing $80K a month at a $60 annual price. That is around 1,300 first payments a month, or roughly 300 a week. Split across four ad sets on Meta that is 75 per ad set, which clears the bar. Split across ten ad sets and two platforms it does not. Structure decides whether the deep event is usable, not only spend. The arithmetic is made up to show the shape; your own numbers go in its place.

Delay. A seven day trial’s first payment lands on day 7 or 8. On iOS, SKAdNetwork and AdAttributionKit report that in the second or third postback window as a coarse value, days later. The platform is learning from a signal that describes what you bought last week. Meta’s predicted value parameters and Google’s value bidding exist to shorten that lag, and they only work if the prediction has been checked against cohorts that have actually matured. That is where value optimization goes wrong when it is switched on early.

So the deepest event is the target, not the starting point.

The rule

Optimize to the deepest event that is frequent enough to clear the platform’s threshold at your budget, timely enough to arrive inside the learning window, and predictive enough that improving it improves paying subscribers.

All three conditions. A frequent event that does not predict payers is a trap. A predictive event that fires 17 times a week is a stall.

In practice that means a ladder.

1. Install when you have no downstream data yet, or when downstream events are broken. Use it to learn costs, not to buy customers. Move off it as soon as you can.
2. A qualified mid funnel event (onboarding complete, a key action, or a trial start with a condition attached) when it demonstrably predicts paid conversion in your own data. This is where most subscription accounts sit on Meta and TikTok until purchase volume clears the threshold.
3. First payment or Subscribe with value when it clears the threshold per ad set, your MMP and store data agree on the count, and refunds flow back.
4. Value or predicted LTV only after you have compared early predictions against matured cohorts and they still hold. Which bidding mode fits which app is a separate decision from the event.

The test for step 2 is simple. Pull 90 days of cohorts. Compare trial to paid and 60 day revenue per install between users who fired the candidate event and users who did not. If the gap is small, the event is not qualifying anyone. Pick a different one. On one account day zero value predicted D28 ROAS better than cost per purchase did, which is the kind of check that tells you whether a signal deserves to be optimized on.

The mistake that looks like success

Cheap trials are the classic. A $10 cost per trial with an 85% day 0 cancellation rate is worse than a $30 cost per trial at a normal cancellation rate. The dashboard says the first campaign is winning. The bank account says the opposite, six weeks later.

The second mistake is quieter. Teams pick purchase optimization, watch it sit in Learning Limited for a month, and conclude that Meta does not work for them. It does. It never got enough data to try.

The third is consolidating ad sets to clear the threshold and then re-fragmenting them a week later for a creative test. Every significant edit resets the clock.

What this page cannot tell you

There is no published controlled experiment that holds audience, creative and spend constant and compares install, trial and purchase optimization on mature subscription revenue. What exists is platform guidance, the research on short term proxies for long term outcomes, and vendor benchmarks. The rule above is a defensible principle. It is not a law of nature, and the thresholds in the table will be different in six months.

What you can do is run the ladder against your own cohorts, compared at the same cohort age, so a week old cohort is never judged against a three month old one. Store fees change the CPA ceiling you are optimizing toward, so the ceiling comes from net revenue, not the list price.

Where your account sits

If your purchase event is not clearing the threshold and you have been on trial optimization for more than a quarter, that is usually a 90 minute conversation. A paid 90 minute session covers your setup and what to change in which order, booked through the same calendar as the intro call. The full picture of paid UA for subscription apps and what signal engineering fixes sit on their own pages.

Sources and scope

Platform thresholds checked 23 September 2026. Meta: About the learning phase, Meta Business Help Center; Meta’s page could not be re-read by script on that date, so the 50 events in seven days figure stands on Meta’s published guidance as I last read it. Google: Best practices guide: Setting up your App campaigns and Set up an App campaign for target return on ad spend, Google Ads Help. TikTok: Tips for Value-Based Optimization for app, TikTok Business Help Center, updated May 2026. Apple: Maximize Conversions best practices, Apple Ads. Trial cancellation timing: State of Subscription Apps 2026, RevenueCat, more than 115,000 apps, calendar 2025, produced by RevenueCat; a distribution of cancellations, not a share of all trialists. The surrogate outcome argument follows Athey, Chetty, Imbens and Kang, NBER Working Paper 26463, a working paper. The $80K example is illustrative arithmetic, not an account.