Learning phase

Automated ad delivery works by trial. When a campaign is new, the system has no idea who responds, so it experiments. Costs swing and results are unstable until it has seen enough conversions. Meta's widely cited guidance is roughly 50 optimisation events per ad set within a week; Google's bidding has a comparable learning period after changes. Large edits to budget, bid, targeting or creative can restart the process.

This shapes how arbitrage campaigns are built. A purchase-style goal is too rare, so buyers optimise for events that happen often in a two-click flow, such as a keyword click or a feed ad click, to exit learning quickly. They also avoid constant tinkering and raise budgets in steps when scaling.

The phase costs real money: early clicks are often bought at a loss while the system learns, and the feed's revenue reporting delay slows the feedback further. Budgeting for this "tuition" is part of honest unit economics. An ad set that cannot gather enough events stays stuck and is usually merged or stopped.

Think of it like this

It is a new waiter's first week: a few wrong orders while he learns the regulars, and every time you move him to a different section he starts again.

An example

Say an ad-click conversion costs $0.50. Reaching about 50 of them takes roughly $25 of spend per ad set, so testing ten ad sets means about $250 spent before any of them is judged fairly.