A fixed number dates quickly. "Pause when CPM is over 40" is wrong the moment your account's normal CPM moves. Comparing a metric with itself over a longer window asks a better question: is this worse than usual?
Write the condition
On a task condition, the right-hand side can be a number or another metric.
Pick the metric and its time window on the left, for example CPM over Today.
Pick the comparison, for example >.
In the value box, type a multiplier, for example
1.5.Choose the second metric and its own time window, for example CPM over Last 7 days.
That row reads: today's CPM is more than 1.5 times the CPM of the last 7 days. The multiplier is what makes it a comparison rather than a fixed threshold, so 1 means "worse than its own average at all" and 2 means "twice as bad".
Why this is usually the better condition
Fixed value | The same idea as a comparison |
CPM today is above 40 | CPM today is above 1.5 × CPM over 7 days |
ROAS today is below 1 | ROAS today is below 0.5 × ROAS over 14 days |
Frequency is above 3 | Frequency over 3 days is above 1.3 × frequency over 14 days |
The left column needs rewriting whenever your account changes. The right column keeps working, and it travels between ad accounts, which is what makes it worth saving as a strategy.
The two windows are independent
The window on the left and the window on the right are set separately, and that is the point. Comparing Today with Last 7 days spots a sudden change. Comparing Last 3 days with Last 30 days spots a slow drift. Comparing a window with itself is always true and tests nothing.
Give the longer window enough data to be a fair baseline. On a low-spend account a 7-day CPM can still be noisy, so prefer 14 or 30 days there.
Common questions
Can I use two different metrics? Yes, but ask yourself whether the comparison means anything. The same metric over two windows is the common and useful case.
Does this work in the filter? No. Filters take a metric against a number. Metric-against-metric is a task condition. See How to set what a rule does.

