A target has been running for three months. Over ninety days it sits at 38% ACOS, comfortably inside your ceiling, and every rule you own leaves it alone. It has also produced nothing for the last fortnight while spending at exactly its usual pace, and nothing in your account is going to tell you.
What averaging does to a break
Here is the same target, cut two ways. First as one window, which is how a rule reads it:
Ninety days, as one block
Spend $2,880 · Sales $7,580 · ACOS 38%
Inside every threshold. No rule fires.
Now the same ninety days cut into consecutive periods, none of them overlapping:
| Period | Spend | Sales | ACOS | Reading |
|---|---|---|---|---|
| Days 90–71 | $640 | $2,010 | 32% | Behaving |
| Days 70–51 | $640 | $1,950 | 33% | Behaving |
| Days 50–31 | $640 | $2,020 | 32% | Behaving |
| Days 30–15 | $512 | $1,600 | 32% | Behaving |
| Last 14 days | $448 | $0 | — | Breaks the pattern |
Nothing about the target's history changed. What changed is that the last period stopped matching the ones before it — and that is only visible because they were read separately.
Seventy-six days of consistent performance is enough to absorb a fortnight of nothing and still average out at 38%. The longer the window, the more it can absorb. This is not a flaw in your threshold; it is what an average is for.
Two mechanisms, two questions
Confusing these causes most of the surprises in an automated account.
- A fixed window is a rule's question. You give it a set number of days and it evaluates thresholds across that whole block. It is predictable, which is exactly the property you want in something that changes your account while you sleep. Its failure mode is the one above: a window containing a good period and a bad one shows you the average.
- Consecutive periods are an analyst's question. A block of time read against the blocks before it, running one after another and not overlapping. That is what makes a break visible instead of averaged away.
"This target is losing money over ninety days" is a rule's kind of statement."This target stopped converting two weeks ago" is the other kind. Do not read one as the other.
Why not just shorten the window?
Because you would be trading one blindness for another. A fourteen-day window on a low-volume target contains too little to conclude anything, so the rule starts acting on noise. And Amazon is still settling attribution on recent days, so the newest data is the least trustworthy data — which is why a rule can be told to ignore the most recent days entirely.
The window length is not the real variable. Whether time is read as one block or as a sequence is.
Rules that read more than one window
Some of this can be encoded into a rule, and where it can, it should be — a break caught on the day it happens beats a break caught when you next open the account. Two shapes matter:
- A long baseline against recent pace. Take a 90 to 180 day baseline that says this target genuinely sells, then check order pace at 60, 30 and 14 days against a floor you set. The long record establishes that the target is worth defending; the checkpoints catch the slide while it is still sliding.
- Deterioration confirmed twice. A term that used to convert over 90 days or more, that has gone bad at 60 days, and gone bad again at 30. Both recent checkpoints have to fail on their own, so one bad fortnight is not enough to condemn a term that has earned for a year.
Both exist for exactly the reason this article exists: a single averaged window cannot see a change in direction.
What to do with this
- Pick three targets your rules have never touched. Cut the last ninety days into consecutive fortnights and read the columns down rather than across.
- Where a target is defended by a long record but slipping recently, that is a baseline-and-pace problem, not a threshold problem.
- Keep the two questions separate. Rules run on a predictable window because they act on their own; asking what changed is a different job and should not be forced into the same shape.