Optimisation is not a burst of activity when something breaks — it is a loop you run on a schedule whether the account is up or down. The operators who compound are not the ones with the best single decision; they are the ones with the most consistent process for turning yesterday's data into today's action, and for doing nothing when the data does not yet justify a move. Most account damage is self-inflicted, done by a nervous buyer reacting to noise.
The loop is simple to state and hard to run with discipline: read the data, decide, act — then wait for the next reading before touching it again. Each pass sorts every unit into one of three verdicts: scale it, hold it, or kill it. This guide lays out the daily and weekly rhythm, the signal-to-action rules that make the verdicts mechanical, how priorities shift across the funnel, how to catch creative fatigue before it drains budget, and — most importantly — how to tell a real signal from random noise. It builds directly on your campaign structure and budget allocation; without those in place, optimisation has nothing clean to read.
Run the loop on two clocks. The daily pass is a quick triage: scan spend, cost per action and any hard limits, catch anything that has blown a threshold overnight, and pause genuine emergencies. It should take minutes and touch almost nothing — its job is to catch fires, not to fiddle. The weekly pass is the real work: with a full week of data you judge which concepts are winning, which are fatiguing, and where the next test budget should go, then make the deliberate scale and kill decisions the daily pass is too twitchy to make well.
The reason for two clocks is statistical. A single day rarely carries enough conversions to trust, so daily decisions should be limited to obvious breakages; the considered moves wait for the week, when the sample is large enough to mean something. Buyers who make weekly-sized decisions on daily-sized data are the ones who churn through winners by accident. Reading the data well is its own skill — start with analytics for beginners if the numbers still feel opaque.
The heart of the workflow is a lookup table from signal to action. When the decision is written down in advance, you stop negotiating with yourself in the moment — you read the metric, find the row, and do what it says. The table below is a starting template; the exact numbers depend on your offer economics and target cost per action, but the structure travels.
| Signal | Reading | Action |
|---|---|---|
| CPA well under target, stable | Proven winner | Scale — raise budget in small steps |
| CPA near target, steady | Marginal but alive | Hold — leave it, watch trend |
| CPA over target, low volume | Not enough data yet | Hold — let it reach threshold |
| CPA over target, past threshold | Confirmed loser | Kill — pause the ad set |
| Frequency rising, CTR falling | Creative fatigue | Refresh — rotate new hooks in |
| CTR healthy, conversions weak | Landing / offer gap | Fix the page, not the ad |
Notice that "over target" is not automatically a kill — it depends on whether the unit has reached its conversion threshold. That single distinction, between losing and not-yet-proven, prevents most premature kills. The strongest media buyers encode these rows as literal automated rules where the platform allows it, so the boring decisions happen without them, and reserve their attention for the judgment calls.
Turning the table into automated rules is what makes the workflow scale past a handful of campaigns. Modern platforms and tools let you chain conditions — pause an ad set when cost per action exceeds a limit and it has cleared a minimum spend, raise budget on a winner that has held under target for two days, alert you when frequency crosses a ceiling. A small set of rules does the bulk of the work: pause underperformers, scale winners, ease budgets by daypart, and flag fatiguing creative. Automating them removes both the delay and the emotion from routine moves.
The discipline that makes rules safe is the guard condition. Never let a rule act on cost alone; always pair it with a minimum-data condition — a spend floor or a conversion count — so the rule cannot kill a good ad set on one expensive hour. Rules without guards are how a bad afternoon deletes a month of learning. Increasingly this is where operator-side AI earns its keep, watching more signals than a person can and surfacing the exceptions; see AI for media buyers.
What you optimise for should change with the funnel stage the campaign serves. At the top of funnel, where you are buying reach and first clicks, the leading indicators are click-through rate, cost per click and hook rate — you are testing whether the creative earns attention at all. Judging a cold prospecting campaign purely on final purchases too early will bury good creative that simply needs more of the funnel to play out.
Deeper in, the metrics tighten toward money. Mid-funnel retargeting and consideration campaigns answer to add-to-cart, lead and landing-page conversion rates; bottom of funnel is judged on cost per action and return, full stop. Match the metric to the stage and you stop making two classic errors: killing a top-funnel winner for not converting like a bottom-funnel one, and letting a bottom-funnel campaign coast on vanity clicks. Clean attribution across these stages is what makes the comparison honest — see attribution models.
Creative fatigue is the slow death of a winner as its audience sees it too often, and it has accelerated. As delivery engines have grown more capable, they find and exhaust the responsive audience faster, so the same creative burns out in less time than it used to — refresh cycles of one to three weeks are now common where they once ran longer. The tell-tale pattern is frequency climbing while click-through rate slides; a widely used rule of thumb flags fatigue when frequency pushes past roughly 2.5 or CTR falls more than about 20% from its peak.
The answer is not to kill a fatiguing winner but to refresh the concept — keep the proven message and change the hook, format or opening visual, feeding three to five fresh executions in before performance craters. That only works if new creative is always in the pipeline, which is why mature operators run a continuous creative supply rather than scrambling when numbers dip. Fatigue is a scheduling problem as much as a creative one; the production side lives in the creative testing framework.
The most valuable optimisation skill is knowing when to do nothing. Small samples swing wildly — a handful of conversions can make a fine ad set look dead or a mediocre one look brilliant, purely by chance. Acting on those swings is how buyers churn winners and chase phantoms. The guard is to insist on enough data before any considered decision: let a unit reach its conversion threshold, judge on trends across days rather than single readings, and resist editing during an active learning phase, since every significant change can restart it.
A useful habit is to ask, before any move, what would I need to see to be wrong? If the answer is "more data than I have," the decision is to wait. Over-reaction is expensive twice — once in the budget it wastes and once in the learning it destroys. This restraint is the operational face of the wider scaling mindset: move decisively when the data is real, and sit still, on purpose, when it is not.
Run a quick daily triage to catch genuine emergencies and a deeper weekly pass for the real scale and kill decisions. The reason is statistical: a single day rarely holds enough conversions to trust, so considered moves wait for the larger weekly sample. Fiddling daily with weekly-sized decisions is the fastest way to churn through winners by accident.
Kill it only once it has both exceeded your cost target and cleared its minimum conversion threshold — that combination means it is a confirmed loser, not just unlucky. If it is over target but still short on data, hold and let it accumulate. Distinguishing "losing" from "not yet proven" prevents most premature kills.
Look at the trend. A bad ad underperforms from the start; a fatiguing one performed well and is now declining, typically with rising frequency and falling click-through rate. Fatigue calls for a concept refresh — same message, new hook or format — while a genuinely bad ad should simply be cut.
Require enough data before any deliberate move: let units reach their conversion threshold, judge on multi-day trends rather than single readings, and avoid editing mid-learning-phase. Ask what evidence would prove you wrong, and if you do not have it yet, wait. Rules with minimum-data guard conditions automate that patience for you.
What media buyers actually do day to day: testing creatives, reading data, managing spend and scaling the campaigns that turn a profit.
Core · 11 min readHow to structure paid campaigns: the campaign / ad set / ad hierarchy, ABO vs CBO, one concept per ad set, naming conventions, and how structure fe...
Advanced · 14 min readThe algorithm handles targeting now; creative is what you control.
