Day three. The client has been in the dashboard since launch, the cost per lead is double what anyone hoped, and the message arrives: should we turn this off?
Most campaigns that get killed are killed on a day like that, before there was anything to read. A smaller number run for months past the point where the answer was obvious, because nobody wrote down what "not working" would look like. Both mistakes come from one gap: the decision was never defined before the spend started.
The takeaway up front: don't decide on elapsed days, decide on two clocks. The volume clock asks whether enough of the outcome you actually care about has happened to be readable. The learning clock asks whether the platform has finished exploring. Until both have run, a bad number is usually noise — with a short list of exceptions that mean act now, today.
Clock one: outcomes, not days
The question people ask is "how long should I run it?" The useful version is "how many outcomes have I seen?"
A week is not a unit of evidence. A week with four conversions and a week with four hundred tell you different things, and the calendar cannot distinguish them. What makes a result readable is the count of the thing you optimise toward — purchases, qualified leads, booked calls — not clicks, and not days elapsed.
As a working rule of thumb rather than a statistical threshold: you want at least a couple of dozen of the target outcome before you compare anything to anything. Below that, the gap between two ad sets is mostly randomness wearing a costume.
Two consequences fall out of that, and both are more useful than the rule itself.
If the campaign cannot produce that many outcomes in a reasonable window, you are measuring too deep. A business with a handful of sales a month will never accumulate enough purchases to judge creative quickly. Optimise and judge on a reliable earlier step — a qualified enquiry, a booking started — and reconcile against real sales monthly.
Budget determines how long the wait is. A campaign spending a tenth of what readable volume requires is not being tested cheaply, it is being tested slowly, and it will produce an unreadable answer in three weeks instead of a readable one in one. The sizing methods behind that are in how to set an advertising budget.
Clock two: the platform is still exploring
Modern ad platforms run an exploration period after a campaign launches or changes materially — the phase where delivery is unstable while the system works out who responds. Costs in that window are routinely worse than what the campaign settles at.
The important part is what resets it: meaningful edits restart the exploration. Changing budgets substantially, swapping the optimisation event, editing targeting, adding or removing creative — each buys you information and costs you the stability that was about to produce a clean read. A campaign edited every second day is permanently in its worst-performing state, and its manager concludes, reasonably enough, that it never worked.
That is the trade-off nobody names when they say "just test it": every change resets the clock you are waiting on. Batch your changes, and make them at a review point you set in advance rather than the morning after a bad day.
Signals that genuinely mean act now
A small set of problems are not noise and should not wait for either clock. All of them are breakages, not disappointments:
- Spend is running with zero of the tracked outcome, and tracking is confirmed working. Zero is different from bad — and verify the tracking first, because a broken conversion event looks identical to a dead campaign.
- The traffic is going somewhere broken. A landing page erroring, timing out, or unusable on mobile. Nothing downstream is measurable until that is fixed.
- The targeting is wrong in an obvious way. Wrong country, language, audience, or product feed. That is a build error, not a performance result.
- Delivery is misconfigured against pacing. A daily budget that will exhaust the month in five days, or a bid cap so low the campaign only wins the least relevant inventory.
- The creative is a brand risk. A claim that shouldn't be live, an offer that has ended. Off, immediately.
- Cost per outcome is many multiples of the ceiling you can afford, with real volume behind it. Not slightly over. If every unit loses money at a rate you cannot absorb, that is a decision, not a fluctuation.
Five of the six can be verified in the account within the hour, without waiting for anything. That is what separates a genuine emergency from an anxious one.
Signals that only look urgent
These generate the day-three message, and none is a reason to stop:
- One bad day. Daily results swing on small numbers, auction competition, and who happened to be online. One day is not a trend.
- A weekend or overnight dip. Most businesses have a weekly rhythm. Compare like periods — this week against last — not Monday against Sunday.
- Day-one cost per acquisition. Almost always the worst number the campaign will ever post, because it comes entirely from the exploration phase.
- A dashboard that disagrees with the CRM. Attribution differences are normal and are not evidence of failure. Reconcile before reacting; the honest version of this problem is in marketing analytics and ROI.
- One ad in a set underperforming. That is the set doing its job. Judge the set; let the platform allocate within it.
- Frequency ticking up in week one. Fatigue is real but takes repeated exposure to develop; early frequency movement is just delivery finding its audience. The actual signature is in ad fatigue and how to fix it.
When a bad number sits on this list, write down what you saw and when you will look again. That is not passivity — it is refusing to trade a readable result for a reaction.
Kill is not the only verdict
"Kill or keep" is a false pair, and reaching for it destroys history that has value. There are four verdicts:
| Verdict | Use it when | What it costs |
|---|---|---|
| Leave it alone | Both clocks are still running and nothing is broken | Patience, and the discomfort of doing nothing |
| Reduce | Results are poor but readable, and you want it alive while you fix something else | Slower learning at the lower budget |
| Restructure | The diagnosis points at the build — wrong audience, wrong optimisation event, wrong offer | Resets the exploration clock; treat it as a new test |
| Kill | The offer or channel itself cannot work at a cost you can afford, with volume to prove it | The account history and any learning inside it |
Pausing sits between reduce and kill: it stops the spend without deleting the structure, though on most platforms a long pause means delivery restarts from exploration anyway. Use it for genuine breakages, and a budget reduction when you simply want to slow down.
Before choosing between restructure and kill, run the symptom through a diagnosis rather than an instinct — why ad campaigns underperform matches symptoms to causes, and the cause decides which verdict is available.
The move that prevents all of this: decide before you launch
All of it is easier if it was written down when nobody was anxious. Before a campaign goes live, put four things in the plan:
- The outcome that counts, named precisely — not "leads", but which event, recorded where.
- The volume threshold at which you will consider the result readable.
- The review date, with no material edits before it unless something on the act-now list occurs.
- The kill criterion, written as a number: the cost per outcome above which this cannot work.
Four lines. They convert every future dashboard panic into a reference to a document, which is among the most useful things an agency can do for a nervous client — and the same discipline that makes a whole plan defensible, as in our advertising strategy guide.
The exception worth naming: short seasonal windows. If the campaign only has two weeks to exist, you cannot wait for volume. Decide in advance which leading indicator you will act on — click-through rate, add-to-carts, cost per landing-page view — and say plainly, to whoever reads the report, that it is a weaker signal.
FAQ
How long should I run a new campaign before judging it? Long enough to clear the exploration phase and accumulate enough of your target outcome to be readable — a function of budget and conversion rate, not of the calendar. If you need a default, review at two weeks, but review against outcome volume rather than the date.
What if the client insists on turning it off in week one? Show the volume: how many outcomes have occurred, and how little that count can settle. Then offer the middle path — reduce the budget and keep it alive rather than killing the structure. A pre-agreed review date makes this conversation much shorter.
Does pausing a campaign hurt performance when I restart it? Often, yes. A long pause typically means delivery re-enters the exploration phase, so you pay the unstable period twice. Reducing the budget usually preserves more than pausing does.
Is a rising cost per acquisition always a reason to stop? No. It can mean fatigue, a seasonal auction shift, or scaling into a less responsive audience. Diagnose which before acting — each has a different fix, and only one is "stop".
Decide on evidence, not on the calendar
The discipline is unglamorous: define the outcome, fund the test enough to produce it, leave the campaign alone while both clocks run, and act instantly on the short list of things that are genuinely broken. Everything else waits for the review date you already set.
If you would rather someone else held that line, find and compare advertising and marketing agencies on AgencyList by specialty and city — and ask each one how they decide when a campaign is finished.
