The short version (3 lines)
- Whether you're increasing, reallocating, or cutting budget, the #1 criterion is the same — the marginal metric (the ROAS/CPA of the next or last dollar).
- Add where marginal efficiency is highest, shift from low to high, cut where it's lowest. Same curve, you're just pushing it in a different direction.
- One exception — the curve isn't symmetric. Scaling that blows up and cuts that don't bounce back share the same cause. That's the core of this post.
Why ranking by average efficiency misleads you
Rank channels by average ROAS (or average CPA) and you keep pouring into already-saturated winners while starving channels that still have headroom. What matters isn't the average — it's the efficiency of the "next dollar" = marginal efficiency. Every channel has a response curve where efficiency falls as spend rises, and every budget decision is just moving a point along that curve. The equimarginal principle: allocation is optimal when marginal efficiency is equal across channels.
One metric solves all three directions
Scaling up — add only where marginal ROAS is above your target line
Having headroom means you're still pre-saturation. Add only to channels whose next-dollar ROAS clears the target line; add to a channel that's dropped below it and you lose from that dollar on.
So why does a winning campaign break the moment you scale it? It entered the flattening part of the response curve. Spend↑ → frequency↑ → creative fatigue↑ → CTR↓ → CPA↑ — a domino. Add a learning reset on top and it gets temporarily worse. So scale in 20–30% steps, let learning stabilize, then take the next step — not one big jump.
Reallocating — move until marginal efficiency equalizes
With a fixed total, pull from the low-marginal channel and push to the high-marginal one. The moment they equalize is the optimum; moving more won't lift total performance.
Cutting — remove the lowest-marginal dollar first
"Told to cut 30% — what goes first?" lives on the same curve. It's not an even, across-the-board trim; you build a cut order ranked by marginal (incremental) contribution. Reversible things first; brand and always-on, which are hard to walk back, last.
The curve isn't symmetric — why scaling and cutting both fail
Just as scaling breaks in the flat zone, cutting doesn't simply rewind the curve. Both betray you because the curve is non-linear and path-dependent. Three asymmetries sit on the cut side:
- Learning-phase reset — cutting rewinds the auto-optimizer's learning. Recovery costs time and money, so the removed dollar costs more than its value on the curve.
- Minimum viable spend floor — below a threshold, some channels don't decline linearly, they collapse. That's why "trim everything evenly" is dangerous.
- It may not reverse — re-adding budget doesn't guarantee the old performance comes back.
Simulate all three with the tool
The Budget Allocation Simulator is marginal-efficiency based, so it shows where to add and where to pull in a single view. Demo data auto-loads, so increases and reallocations you can judge straight from it.
⚠️ One honest caveat. This simulator assumes budget moves are reversible. Asymmetry #3 (reversibility) is not in the model. In reality, pulling then re-adding may not restore performance — creative fatigue accumulates and learning resets in the meantime. So for cut order, don't trust the reallocation output alone; pair it with Incrementality Analysis to see "how much actually disappears if I pull this channel."
Try it on your own data
No signup · data processed in your browser only (zero server transfer) · check instantly with demo data.
👉 Open the Budget Allocation Simulator
Related
- Tools: Campaign Saturation Diagnosis — is there headroom to scale · Creative Fatigue Analysis — why cuts don't bounce back
- Glossary: Marginal CPA · Marginal ROAS · response curve — the marginal-metrics hub