In-flight media optimisation: what to change, when, and why

Optimisation is a cadence, not a reflex

The failure mode is tinkering: daily budget nudges, audience swaps on a hunch, bid changes because a chart dipped. Every change resets platform learning and muddies attribution of what caused what. The alternative is a cadence - a fixed rhythm at which evidence is reviewed and decisions made, with defined triggers for acting between reviews. Weekly is the working default for most always-on activity; short flights compress it. Between reviews, pacing alerts handle the genuinely urgent.

The signal hierarchy: delivery, efficiency, outcome

Diagnose in order, because each layer explains the one below it.

1. Delivery. Is each line spending to plan, at sane frequency, in the right geos? A line that cannot spend has a delivery problem, and no efficiency metric from a starved line means anything.
2. Efficiency. At what cost is delivery arriving - CPM, CPC, CPA - and in what direction is it trending? Rising CPMs with flat outcomes point at auction pressure or fatigue; the fix is creative or audience, not budget.
3. Outcome. Is the spend producing the KPI the plan set, at the target rate? Only here do reallocation decisions live - and only when the two layers above have been ruled out as the cause.

Most bad optimisation decisions are category errors: solving an efficiency problem with a budget lever, or reading a delivery constraint as an outcome failure.

Diminishing returns and the marginal question

Every channel-audience combination has a response curve, and past a point the next pound buys less than the last one did: deeper into the auction, higher frequency against the same people, weaker remaining inventory. The number that matters is therefore marginal cost, not average cost. A channel can show a healthy average CPA while its marginal CPA - what the most recent budget increment actually paid per outcome - has drifted well past the point where another channel would use the money better.

Estimating marginal cost by hand means comparing efficiency across spend-level changes, which is tedious and usually skipped. It is also exactly the pattern problem models are for: Medusa's ML recommendations learn each channel's response from connected account data and surface moves like lowering a daily budget to the diminishing-returns threshold, ranked by predicted impact, with the reasoning attached.

Reallocation rules that survive scrutiny

Write them before launch, into the plan. A workable rule has four parts: the trigger (what signal, e.g. CPA divergence between two lines), the threshold (how much, for how long - say 20% for seven days with adequate volume), the action (move up to X% of the weaker line's budget), and the cap (no line moves more than Y% per review, no channel drops below its minimum viable level). Rules written in advance turn the weekly review into execution. Rules improvised in the moment turn it into a negotiation.

Creative fatigue: the efficiency leak with a spend costume

Fatigue announces itself as gently rising frequency, sagging engagement and creeping response costs - often misread as an audience or budget problem. The tell is divergence between a creative's delivery cost and your effective target: when an ad's cost of delivery runs at a multiple of your effective CPA, refreshing creative beats any budget move. Watch frequency alongside it; the same audience seeing the same asset nine times is not a targeting insight, it is a rotation failure.

When not to optimise

• During learning phases. Conversion campaigns need a stable period post-launch or post-edit; every material change restarts the clock.
• On small samples. A CPA computed from nine conversions is noise wearing a decimal point. Set minimum volumes for decisions and hold to them.
• Inside normal variance. Daily wobble within an expected band is weather, not climate. The threshold-and-duration test exists to filter it.
• During externally noisy windows. Auction-wide CPM spikes in retail peaks are the market moving, not your campaign breaking.

The weekly loop, on one table

Write down what you did, and why

The decision log is the least glamorous row on that table and the one that pays longest. A dated line per change - the evidence, the move, the expected effect - is what makes the post-mortem an analysis instead of a memory contest, and it is the raw material the next budget allocation learns from. Optimisation without a record improves this flight; optimisation with one improves every flight after it.