The same campaign can read 0.3 at seven days and 2.1 at ninety. Here is why, and the single figure that every bid decision actually follows from.
ROAS is revenue divided by ad spend over some window. For a game selling a one-off unlock it is a fair summary of a week. For a subscription app it is mostly a summary of how long your window was, and if you manage to it you will turn off the campaigns that were working.
A subscriber pays you over months. The ad that found them was paid for on a Tuesday. Any window shorter than the time it takes a user to pay you back will therefore report a loss — not because the campaign lost money, but because you measured before most of the money arrived.
Then there is the trial. An app with a seven-day trial has, by design, almost no revenue in the first week from users acquired in the first week. Day-one ROAS for that app is a measurement of the trial length. Day-seven is a measurement of the trial length plus the weekend. Neither is about the ads.
And Apple's commission comes off the top. A subscription at ten a month is not ten a month to you, and a ROAS computed on gross revenue is describing a business you do not own.
Ask a different question: what is the most I can pay for one install and still come out ahead? That is a single figure, it does not move with the window you chose, and every bid decision follows from it.
Work it out in this order:
Almost nobody knows theirs, which is the main reason app advertising is guesswork. It is also why "is a two dollar install expensive?" has no answer: it is cheap for an app that makes five and ruinous for an app that makes one.
One line, one threshold. Above it you are buying users at a loss; below it you are buying them at a profit, and the only question left is how many you can get.
Take the people who installed in one month and follow what they paid over the months after. That tells you when a campaign becomes profitable rather than whether it was profitable in a window that ended arbitrarily.
New campaigns lose money while they find out which searches convert. That cost is real and unavoidable; what is avoidable is paying for the same lesson twice. Judge the learning period on whether the expensive terms are being turned off, not on its ROAS.
When the window is longer than your payback period and the revenue is net of Apple's cut. At that point it is a reasonable summary for somebody who wants one number — an investor, or you in a quarterly mood. It is a rear-view mirror, and it is a poor steering wheel.
For the daily decisions — raise this bid, pause that keyword, open this country — the number that decides is what an install is worth to you, and ROAS cannot tell you that.