Overview
Jay Stampfl in Search Engine Land has set out a clear challenge to how marketers think about budget testing and media allocation: not every channel follows the neat law of diminishing returns. His argument, explored in detail in his piece for Search Engine Land, is simple but uncomfortable – some modern, machine-learning driven channels have a warm-up phase that makes early spend the least efficient, not the most. Read the full column by Jay Stampfl for background. https://searchengineland.com/channels-breadth-commitment-480765
The two shapes that actually matter
Marketing response curves come in two useful flavours. The first is what most of us learned to expect: a C-shaped or concave curve where the first euro buys the most and each additional euro yields less. The second is the S-shaped or sigmoid curve where there is a slow, inefficient start, an inflection point that unlocks much better returns, and then saturation. The crucial point is that you do not allocate against the curve itself but against the marginal curve – what the next euro will buy you. On a C-curve marginal returns fall from the start. On an S-curve marginal returns are U-shaped – poor at the outset, best around the inflection, and worse again as saturation approaches.
Why the difference breaks the industry playbook
Most marketing teams operate on a one-size-fits-all playbook: test small, learn, and scale winners. That logic only works when channels are C-shaped. For S-shaped channels a small test can sit squarely in the warm-up region and produce ugly CPAs. If you kill the channel on that evidence you may never reach the inflection where the channel becomes one of your most efficient marginal buys. Stampfl uses an S-curve worked example to show how a routine €10,000 test can return average and marginal CPAs that look hopeless, while the channel at €20,000 to €25,000 could be operating at CPAs well below target.
Platforms are changing the game
The shape problem is not purely academic. Platform changes – notably Google’s moves towards broader, machine-learning first campaign types – make the warm-up visible again. Performance Max blends a harvesting layer that behaves like a classic C-curve with a prospecting layer that behaves like an S-curve. The early blended numbers can flatter headline performance while hiding the prospecting warm-up beneath branded and retargeting wins. If you want to understand the mechanics of Performance Max for yourself, Google’s product pages explain the approach and the surfaces it touches. https://ads.google.com/home/tools/performance-max/
Measurement, modelling and the trap of averages
This is also a measurement problem. Average CPA and small-scale attribution can mislead. Marginal CPA at your current operating point is the metric that matters for allocation decisions, but most dashboards give you totals and averages. Multi-cell scaling tests, time-sliced marginal CPA analysis, and well-designed holdouts each answer different parts of the question, and no single approach is a panacea. At the aggregate end, marketing-mix modelling can estimate whole curves from programme-level data, but it trades off identifiability and stronger assumptions.
Two cautions Stampfl flags and Irish advertisers should heed
1) Separating a true S-curve from a concave curve with a high half-saturation point is hard. Standard concave models will often fit S-shaped data well enough to hide the warm-up. That makes rear-view mirror analysis deceptive.
2) The learning phase itself can be partly one-time. Some channels carry a fixed training cost – once you feed enough data the marginal behaviour looks concave. But there may also be a persistent minimum volume required to keep performance efficient. Treat any shape diagnosis as provisional and re-test it over time.
Practical rules for allocation
– For C-shaped channels, go wide. Small allocations are productive and the equimarginal rule – equalise marginal CPA across channels – usually gives an interior solution where many channels receive funding.
– For S-shaped channels, treat the decision as binary at launch. Either commit a block that carries the campaign past the inflection or do not fund it. Small, underpowered tests will systematically understate potential.
– Split blended products. With hybrids such as Performance Max, separate harvesting inventory (brand, retargeting, shopping) from prospecting inventory before you judge viability. The blended headline will often flatter you into underfunding the part that creates demand.
– Measure marginal returns at the operating point. If you cannot run a large enough test, consider holdouts or scaled multi-cell tests that expose the marginal slope. MMM can help when attribution is broken, but design and assumptions matter.
What this means for Irish businesses
Irish marketing teams, especially in SMEs and scaleups where budgets are tight, should be cautious about reflexive pruning. The political economy of platforms pushes automation and rapid optimisation, but that is democratically useful only if your account already sits above the learning threshold. If you are trialling AI-driven campaign types without the conversion volume to train them, you will often see poor early results and be nudged to stop – exactly the wrong time to persevere. Be willing to make a deliberate, accountable decision to fund a campaign past its warm-up or to skip it entirely. That is fiscally conservative in the long run; it stops wasteful flirtations with channels that will never scale for you and commits capital where it has a genuine chance to compound.
A checklist for smart commitment
– Estimate the likely inflection zone before you test. Use historical conversion velocity and funnel metrics to approximate the minimum viable spend.
– Run multi-cell scaling tests when possible. They cost more but give the two-dimensional answer you need: how the incremental slope moves as you scale.
– Use holdouts or geo tests to establish incremental lift rather than relying on attributed ROAS alone.
– Treat platform recommendations with healthy scepticism. Automation optimises what it can measure, and platforms have incentives to promote blended products that mask prospecting cold-starts.
Final take
The old marginal-CPA rule still holds. What has changed is how you get there. The shape of the response curve is mostly a launch and evaluation problem. For channels that require a minimum volume to learn, discipline matters: commit a defensible block past the warm-up or do not start. For channels that reward breadth, keep spending lean and reallocate at the margin. Stampfl’s essay is a timely reminder that in a world of algorithmic auctions and automated optimisation, the hand that funds or kills a channel matters as much as the algorithms you let run the campaign. For the original analysis and visual explanation, see Jay Stampfl at Search Engine Land. https://searchengineland.com/channels-breadth-commitment-480765
