Why feeling isn't enough
There's no room for "I think this ad isn't that good." Managing traffic means looking at numbers and comparing them against a defined target before opening the ads manager. If two people look at the same data with the same pre-established criteria, they should reach the same conclusion.
The first step is always defining the metric guiding the campaign: cost per lead, cost per acquisition, or ROAS, depending on the goal. Without that defined metric, the data says nothing on its own, because good or bad always depends on what you're trying to achieve.
To prepare your next step, explore Artificial intelligence in digital marketing.
Don't decide with too little sample
A common mistake is pausing ads that have barely spent anything. You need enough sample before drawing any conclusion about an ad set. I mentioned using a language model to ask, feeding it data from several ad sets, whether there's already enough sample to declare a winner.
That helps avoid hasty calls, but reaching a spend threshold on its own doesn't prove statistical significance; it's just a practical guardrail against deciding on too few events.
Explore this decision further: Did the Ad Fail, or Did It Never Get a Real Chance to Be Tested?.
The CPA multiple as a practical benchmark
One way to give an ad room before pausing it is letting it spend a multiple of your target cost per acquisition. In the example I gave for a long-term campaign with a 20€ CPA target, I only decide from 40€ spent without result, a multiple of two.
If by 60€ there's still only one isolated conversion, that's when pausing is justified.
That multiple isn't fixed. It depends on the project's risk profile and the value of the target. Low targets, like a 2€ lead, can tolerate larger multiples (say three times, so 6€) without much cash-flow impact. High targets, like 100€ per lead, don't justify waiting until 300€ spent; there the multiple should sit much closer to one. Short campaigns also call for tighter multiples, since there's no time to let an ad breathe the way you would in an evergreen campaign.
The learning phase isn't an excuse
If a campaign has spent a significant amount and still produces zero results, waiting for the platform's so-called learning phase solves nothing. The algorithm only learns to optimise delivery when it receives positive conversion signals; without them, it's simply gathering data about auctions that don't work for your goal.
In my experience, campaigns that start badly rarely correct themselves over time on their own. The recommendation is to give the auction the minimum necessary time to test itself, then act as soon as the predefined spend limit is reached without traction.
Secondary signals that anticipate wear
Even when accumulated CPA still looks acceptable, it's worth tracking second-level metrics: CPM, frequency, CTR and CPC. If frequency rises and CTR drops week after week, that's a sign the audience is saturating, and conversion cost is likely to climb soon.
Breaking performance down by segment (placement, platform, age) can reveal that a single placement is draining budget without return, letting you fix the issue without pausing the whole creative.
Always compare several time windows
A mistake is pausing a profitable ad just because it had one weak day. The recommendation is to look at yesterday, the last 7 days, the last 14 days (covering two full weeks, including two weekends) and, when possible, 30 to 90 days.
An ad might have been a top performer over a 90-day window and only look bad because yesterday it didn't sell anything.
When a campaign is hitting its target, the best decision can be to leave it alone. I like to say the team that's winning doesn't change formation: if you want to test new hypotheses, do it in parallel ad sets, not inside the campaign that's already working.
Rules that protect account history
Two practical rules make a difference. First: I never delete ads, ad sets, or campaigns with delivery history, even when I decide not to continue with them. Deleting destroys traceability of past conversions; you can only delete something that never registered a single impression. What you do instead is pause it.
Second: I never fully overhaul an old campaign by swapping copy, image, and landing page inside the same structure. That mixes old and new data and prevents honest comparisons over time. The correct path is to pause the old structure, duplicate it, and apply the changes to the new copy, which starts with a clean history.
Explore this decision further: What to Record When a Campaign or Partnership Ends.
Optimising is about reinvesting, not cutting spend
The true point of optimising is freeing up budget from ineffective ads to reinforce the ones that already work. Sound growth logic focuses on being able to invest progressively more while keeping solid profitability. You direct every available pound straight toward the variations that deliver real returns.
If an ad goes from a ROAS of 5 to a ROAS of 14, the sensible move is to scale that investment, not shrink it.
Record the hypothesis before making the change
Imagine two ads promoting the same offer. One received much more spend because campaign delivery concentrated on it. Before comparing sales, check how much each ad spent, how long it ran and whether the destination and offer were equivalent. This helps establish whether you are comparing sufficiently similar situations.
Then write down the hypothesis you want to test. You might give an under-delivered variation more opportunity, revise a message that has started losing response or investigate a problem on the destination page. Choose a change you can connect to that hypothesis and decide when to return to the data.
If you change the budget, creative, audience and page simultaneously, understanding what contributed to the result becomes harder. Some situations require several corrections, but record that limitation. The purpose of the history is to retain a useful explanation for your next decision, including when a test does not support a clear conclusion.
Finally, compare what happened with what you expected. If the hypothesis was not confirmed, keep that information. It can prevent you from repeating the same change weeks later without remembering that you already tried it in that account under those conditions.