You tripled the budget. A week later cost per trial sits where it always sat, so you call the scale a success and plan the next increase. That reading is not wrong yet — it is early. The number holding steady is the only number that had time to arrive.
Two clocks are running, and they are not the same length. One belongs to the ad platform. The other belongs to your business. Almost every argument about whether a scale worked is really an argument about which clock someone is reading.
A flat cost per trial after a budget increase is not yet evidence the scale worked. Meta's attribution window closes long before your payback period does, so the number that stayed flat is the only one that had time to arrive — the quality of the cohort you just bought shows up weeks later. Gate the next increase on the trial-to-paid rate of the previous cohort, not on cost.
Meta's clock
Conversion campaigns optimize against an attribution setting you choose at the ad set level, and the longest click window on offer is seven days: Meta's documented options are 1-day or 7-day click, with an optional 1-day view. Anything that happens outside that window is invisible to the system doing the optimizing. This is a measurement horizon, not a defect — an optimizer has to close its books at some point in order to learn from them, and seven days is a reasonable place to close them for most of what runs on the platform.
Two caveats on that sentence, both from Meta's own documentation. The window above is the one the delivery system optimizes against, which Meta states plainly is a different setting from the attribution window you read in reports. And it describes standard attribution: Meta now also offers incremental and custom attribution models, and the time-window framing here does not apply to those.
Your clock
If you sell a subscription with a trial, the event that pays for the ad is not the trial. It is the second month, or the third. Payback measured properly often lands somewhere between thirty and sixty days. Which means the outcome you actually care about happens long after Meta has closed the books and moved on.
Why a flat CPA hides the damage
More budget buys a wider audience. That is the mechanism, and it is the point — you cannot spend three times as much on exactly the same people. But a wider audience is, on average, a less qualified one, and the effect does not land evenly across the funnel. The front end absorbs it quietly; the back end takes it.
- Cost per trial: flat. Wider targeting is cheaper per click, which offsets the lower intent.
- Trial volume: up, roughly with budget. This is the number that makes the scale feel successful.
- Trial-to-paid: down. Nothing in the ad account reports this, and it is the one that decides whether the money came back.
So the dashboard can be entirely truthful and still tell you the opposite of what happened. Nobody is hiding anything. The damage simply lives on the far side of the window.
What Ads Manager cannot tell you
Ads Manager reports the segment it handled: impression, click, and whichever conversion event fired inside the window. Whether those people were still paying in week six is not a fact it has access to. That fact lives in your billing system, and it is the only place the question can be answered.
Gate the increase on the previous cohort, not on cost
Before raising budget again, look at the trial-to-paid rate of the cohort your last increase bought. If it has fallen more than about 20% relative to the cohort before it, hold the budget where it is until you know why.
Relative, not absolute — a two-point drop is nothing at a 40% base rate and catastrophic at 5%, so a fixed number of points is not a threshold, it is a coin flip that varies by product. And the 20% is not arbitrary: a 20% relative fall in trial-to-paid raises your true cost per paying customer by 25%, because you are dividing the same cost per trial by a smaller rate. That is a real move, large enough to act on and large enough to sit clear of ordinary week-to-week noise — provided the cohort is big enough for a rate to mean anything at all. At single-digit trial counts you are reading noise no matter which threshold you pick.
How long to wait before comparing
Give each cohort at least half your payback period before you judge it. Compare a two-week-old cohort against a six-week-old one and the younger group loses every time — not because the traffic got worse, but because it has had a third of the time to convert. Group by the week the trial started, let each week mature the same amount, and only then put them side by side.
The short version
Flat cost per trial after a budget increase is not evidence the increase worked. It is evidence that the fast number arrived first. The slow number decides, it lives in your billing data rather than your ad account, and it needs a few weeks before it will say anything honest. We built our own optimizer to hold budget increases behind exactly that check, which is the only reason we have opinions about it.