How to Kill a Losing Ad Using a Written Decision Rule

Put the decision order in writing before you launch: delivery, hook, hold, outbound CTR, then CPA against your allowable. The stack is what you will read in that sequence, and the thresholds are what you will act on. If the rule only exists in someone's head, the meeting will negotiate it every time an ad they like is under water. Write the numbers on the brief, not in a later Slack thread.

Use the same working floors as the rest of the canon, not a private-group average. Hook rate is 3-second views divided by impressions: target 30%+ on cold traffic, treat below about 20% as a weak open. Hold rate is 15-second views divided by 3-second views, healthy at 12–20%. Outbound CTR target is 1%+ on cold. CPA is judged against the break-even you already calculated from contribution after COGS, shipping, fees and expected refunds — not against a round number that sounded tidy. Write the window too: diagnostics at 48–72 hours or about 1,000 impressions per variant; economic kills only after purchases have had time to arrive.

Also write the things you will not kill on. Do not kill on CPM. Do not kill because the other avatar looked better in the preview. Do not keep a dog because it has comments. Do not abort the whole 6-creative grid because day-one CPA looked mean — that is learning-phase weather, and teams that abort at hour ten never accumulate a control. Paste the stack into the brief so a media buyer who was not in the kickoff can still execute it on a Thursday without a vote or a fresh debate in that Slack thread.

At 24 hours, look only at whether each ad is actually delivering. A file with near-zero impressions is usually unapproved, broken, or so uncompetitive it never entered the auction — none of those is a hook-rate verdict. Even spend is nice; it is not required. What is required is enough delivery to later trust the diagnostic. Do not crown or bury a variant on overnight CPA.

Open the delivery diagnostics before you open the performance columns. Check disapprovals, learning-limited warnings, and whether one variant ate the budget because it won the first Reels burst. In a consolidated ad set the algorithm is trying to get conversions, not to run your science fair, so uneven spend is expected. You still get a read on a starved ad if it reached a few hundred to a thousand impressions; below that you do not. Catch-up budget mid-test usually contaminates the CPA you were waiting for, so leave the apparatus still unless something is actually broken.

Policy-violating and clearly broken files are the exception: kill those immediately. A rejected ad has no CPA. A 404 destination has no CPA. A mute render or a missing pixel is a production incident, not a creative test. Log the reason code if it was a rejection so the next batch is written clean rather than appealed clean. Everything else waits for the diagnostic window you already printed on the brief — 48–72 hours or about 1,000 impressions — even if the room is already picking a favourite from the previews.

After 48–72 hours or about 1,000 impressions per variant, kill or rewrite any cold ad still under about 20% hook rate. That number is a thumbstop failure. The body script, the avatar you preferred, and the landing page cannot save an open that never earned the look. This is the first performance kill, and it is the one teams skip because the file still 'might convert'.

Read hook at ad level, on cold traffic, with 3-second views over impressions — not thruPlay, not campaign roll-up, not a mix of retargeting and prospecting. A 45% hook on a 1% conversion audience is recognition, not a prospecting grade. Break Reels out from Feed when the gap is large. Pair the number with hold and outbound CTR so you do not crown a curiosity trick: a 40% hook with hold well under 12% is a false thumbstop, and a high hook with outbound CTR under 1% on cold is a pause you bought, not attention you earned.

Grey-band ads around 25% are where waste happens. Above the weak line, below the 30%+ target, they are salvageable if hold is a healthy 12–20% and CTR is 1%+ with an acceptable early CPA — keep the body, replace the first two seconds. If hold and CTR are also soft, treat 25% as a warning that the open is not doing enough work. Do not scale a grey-band file just because it is not the worst in the set, and do not keep it running beside a 34% hook on equal budget once the window has closed. Iterate it; do not defend it.

Purchases arrive slower than 3-second views. After the hook read, let survivors collect conversions before you kill on cost. A directional CPA needs roughly 10–15 conversions; a scale decision wants more. Killing on four purchases is how you fire a good hook and keep a noisy one. Use the spend checkpoints you wrote on the brief, not a mood in the daily standup.

A practical early-kill line that matches the rest of the playbook: at 1.5× target CPA spent with zero conversions, kill. At 2× target CPA spent with under two conversions, kill. At 3× target CPA spent with CPA still 1.5×+ over allowable, kill. Those lines assume the pixel is firing and the destination works — if they are not, you are killing a tracking bug. Day-one CPA is not one of the lines. Editing the ad, splitting the ad set, or raising budget mid-window to 'help a loser catch up' resets learning and guarantees you never see a stable number.

If 72 hours still is not enough conversions, do not force a CPA ranking on six purchases total. Drop the obvious dogs on hook, hold and CTR, and let the rest run. Forcing an economic ranking on a handful of events will pick a random winner and you will scale it. Match the conversion event to the one you will actually spend for; testing on landing-page views and scaling on purchases trains two different jobs, then you wonder why CPA moved. Write that event name on the same brief as the kill stack.

Not every rising CPA is a losing ad. Falling hook rate with flat frequency is fatigue — refresh the first two seconds, do not pause the whole account. Rising frequency in the 2.5–3.5 band on cold with a flat hook is an audience ceiling. Strong hook and hold with weak CTR is a click or offer problem. Strong click and terrible CPA is usually the page, the claim, or tracking. Kill the right thing.

Walk the funnel in order so the meeting cannot skip to taste. Weak hook (under ~20%) with decent hold and CTR on the people who stay is a hook problem: new first two seconds, same body. Strong hook (30%+) with hold well under 12% means the open over-promised and the next ten seconds did not pay it off. Strong hook and hold with weak outbound CTR is creative-to-offer or creative-to-page. Strong click and terrible CPA is not a reason to keep iterating faces. If several decent hooks and avatars all land on the same bad CPA after a fair read, stop rendering and diagnose the offer.

Winners get a different kill rule. Retire them when fatigue is real — frequency in the 2.5–3.5 band on cold, hook sliding toward the weak line, CTR off 1%+, CPA rising — and you have a replacement ready. Killing a winner because a new test launched is a mistake; overlapping is cheaper. Killing a winner because you are bored of it is also a mistake. Winners pay for the tests. Keep a graveyard note of hook family and face so you do not relaunch the same dead open next month and call it a new idea.

When you kill, pause the ad, write one sentence on why, and harvest the lesson into the next 3 × 2 batch. Do not leave losers delivering 'just in case'. Do not dump six untested hooks into the only scaling ad set on a Monday. New creative is allowed to fail. Scaling creative is not. The test lane stays open; the scale lane only takes files that cleared the stack.

The log is the unglamorous part that saves the most money. Capture the hook family, the face, the spend, the hook rate, the CTR, the CPA, and the reason code — diagnostic kill, economic kill, policy, fatigue, or offer. Six weeks later nobody remembers whether the ad died from a broken landing page or a 14% open, and without that note the next brief is a guess. Teams that keep a simple retired-creative sheet routinely find that their best new ad of the quarter is a hook family they already killed for the wrong reason, or a face they discarded on preview taste.

Scale what survives gradually: 20–30% budget steps every two to three days rather than an overnight double that sends a healthy CPA back into learning. Keep replacements rendering, because frequency 2.5–3.5 arrives sooner as you spend. Copies of the same open are not diversification; they are frequency with extra filenames. Genuine next tests are a new first line, a new face, or a new proof beat — not twenty clones of the file that won sitting in twelve lookalike ad sets on the same afternoon. Scale spend, or scale genuine variants.