Average ROAS hides the marginal number
An account at 4x average can be buying its last £2,000 at 1.5x. The average lags because early efficient spend props it up. Work out marginal ROAS by comparing incremental revenue against incremental spend across the step change.
Do the arithmetic weekly. Week one, £5,000 at 4x. Week two, £7,000 at 3.4x. The extra £2,000 bought £3,800, so the increment ran at 1.9x. If break-even is 2.5x, that last increment lost money and should be pulled.
The signals that arrive before the loss
Frequency climbing past 2.5 to 3 in seven days on prospecting. CPM rising while CTR falls. Conversion rate on site holding while cost per click climbs. Each says the auction is charging more to reach a colder slice of the same audience.
Watch new-customer share too. A campaign that scales by hoovering up returning buyers flatters ROAS and adds nothing. If the proportion of first-time orders falls as spend rises, the extra budget is buying repeat purchases you already owned.
How to scale without resetting learning
Increase budgets by 20% to 30% at a time and leave 48 to 72 hours between changes. Large jumps push the ad set back into the learning phase, which typically needs around 50 optimisation events a week to exit.
When a winner tops out, duplicate into a new campaign with a fresh audience or placement mix instead of forcing more budget through the same ad set. Then iterate hooks on the winning concept to buy another few weeks.