Every week we run the same four filters, in the same order, against the same targets on every account, because risk on Amazon gets managed by checklist and not by mood. We have ten years in Amazon e-commerce and we apply this discipline across 15 to 20 accounts at once, so the rules stay simple enough to run at scale and strict enough that nothing bleeding can hide. The system is the keyword risk ladder, and every rung is defined by evidence.
Before any filter fires, we pull each target through six windows: lifetime, year to date, 60 days, 30 days, 14 days, and 7 days, and we add 90 when the account has enough history to make it meaningful. A target can look fine on its lifetime numbers while it has been bleeding for the last two weeks, and the reverse is common, because an ugly lifetime can be converting right now. We check every filter against both short and long windows before we act, so no single flattering or ugly view makes the call alone.
Four filters, top to bottom
Filter one catches the worst offenders, the targets that have burned over $10 with zero sales. We pause them, because at that spend with nothing to show for it the auction has answered the question. This is the only rung where we pause, and we explain below why keeping that restraint matters.
Filter two flags targets still enabled that have spent over $5 with zero sales. We cut the bid by 50 percent, and we take the cut to 75 percent when the account is running above its ACOS target, because a hot account has no room for patience. When the whole account sits over target, every speculative dollar is a dollar the profitable targets have to earn back, so our tolerance for maybes drops with the account.
Filter three works the small end of the ladder, which is spend under $5, zero sales, and at least one click. The action is a bid cut of 15 percent, or 25 on a hot account. The evidence here is thin, since a few clicks without an order proves very little, so the response is a trim rather than a cut, and the target keeps competing while it builds a record.
Filter four is the only rung that looks at converting targets. Anything with an ACOS above 50 percent gets a 15 percent bid cut, or 25 on a hot account. A sale does not make a target innocent, because a target that converts at an unaffordable cost still leaks, and it only leaks more politely than the zero-sale ones.
We keep a companion rule for clicks with no conversions that anchors the new bid to the CPC the auction is charging. When CPC is at $0.02 we set the bid to $0.02. When the current bid sits between $0.02 and $0.10 we set it to CPC minus $0.02. When the formula result comes out above the current bid we leave it alone, because a rule that raises a bid on a target with no conversions has stopped being a risk rule.
Severity is proportional to evidence. We only pause at the top rung, where the spend is large and the auction stays silent, and everything below that keeps a live bid, because a paused target stops producing information while a reduced bid keeps the target reporting on itself every week. Every run gets logged with screenshots in project management, so there is a permanent audit trail of who cut what and when. A question about any bid change months later gets answered by a record instead of a memory.
We built the manual weekly version first and ran it by hand long enough that every threshold earned its place. Once the rules had survived real accounts across all six windows, they moved into software, and our automation now runs the same ladder as code. The order matters, because rules you have never run by hand are guesses, while rules you have run for years by hand are policy, and policy is the only thing we let a machine execute against $10M+ in managed ad spend.
← Back to all posts