Why some channels reward breadth and others require commitment

Many finances allocation methods assume that each channel follows the identical sample: the primary greenback is the most efficient, and every extra greenback yields a barely decrease return.

The charts beneath present what that sample seems like.

The log form implies that the primary greenback is the most efficient, and every subsequent greenback is price rather less. When each channel seems like that, the sport plan is to unfold the finances to as many channels as doable and equalize the marginal CPAs to maximise revenue.

However not each channel seems like that. Some have a warm-up area the place the early spend is the least environment friendly, not essentially the most. On these channels, the logic above breaks, and so does the “check small, scale the winners” playbook that a lot of the trade runs on autopilot. 

The distinction comes down to at least one query concerning the channel: Is the response curve C-shaped or S-shaped?

The reply can change the way you method channel testing and channel measurement, together with any MMM evaluation. Furthermore, Google has been incorporating extra S-shaped marketing campaign varieties, and after its Google Advertising Dwell bulletins, this development appears set to proceed.

The 2 shapes — and the one half that issues

The response curve plots output (conversions, income) in opposition to enter (spend). This usually ends in two sorts of curves in advertising.

• C-shaped (concave): Diminishing returns from the very first greenback. A log or energy curve. Image the top-left quarter of a circle: steep at the beginning, flattening as you go.
• S-shaped (sigmoid): A sluggish, inefficient begin, then an inflection level the place it will get steep, adopted by a flattening into saturation. A logistic curve.

The response curve itself isn’t what you allocate in opposition to. You allocate in opposition to the marginal curve, the spinoff, which solutions the query: “What did the subsequent greenback purchase me?” That’s the place the shapes diverge in a approach that issues.

• For a C-curve, marginal return is highest on the first greenback and falls in just one course. Marginal CPA rises from the primary greenback onward. If conversions are a*ln(s), marginal conversions per greenback are a/s, so marginal CPA is s/a, climbing in a straight line as you scale. There’s no warm-up. The most affordable conversion you’ll ever purchase is the primary one.
• For an S-curve, marginal return begins low, rises to a peak on the inflection level, then falls. Marginal CPA is U-shaped. It’s costly at the beginning, bottoms out across the inflection level, then climbs into saturation.

That area of accelerating marginal returns is the entire story. It’s the distinction between a channel the place small budgets are productive and one the place they’re wasted.

How this seems in a advertising marketing campaign

Say your CPA aim is $50. Right here is an S-shaped channel, modeled as Conversions = 1000 / (1 + e^(-0.25(s – 20))), with spend within the 1000’s and the inflection at $20,000/month:

Run the $10,000 check {that a} sane particular person runs earlier than committing actual finances. Common CPA comes again at $132, marginal round $94. If these two metrics are all you have a look at, you conclude that this channel can’t hit $50, so let’s kill it.

That verdict is fallacious. At $20,000 to $25,000, the channel is operating at a median of $32 to $40, and the marginal greenback within the $15,000 to $25,000 band prices $18. That’s not “barely viable.” In that band, it’s the very best marginal purchase you may have. The small check fell throughout the warm-up and reversed the conclusion.

In a C-shaped channel, the small check would have proven you the very best the channel can do. On an S-shaped channel, it exhibits you the worst.

That is the lure. The usual playbook is “check small, scale what works.” On S-curves, small assessments systematically condemn channels that will’ve labored at scale as a result of the check is structurally caught within the inefficient area.

The allocation logic, restated

C-shaped channels, go vast

The optimization is convex. There’s one international optimum, the equimarginal rule from the marginal-CPA submit applies cleanly, and the answer is normally inside, which means a number of channels get funded.

Even a small allocation is productive as a result of the primary greenback is the very best greenback. Run many channels lean, reallocate repeatedly on the margin, and pull again the moment marginal CPA crosses your aim.

S-shaped channels, go deep or skip

The optimization is non-convex. A small allocation may be strictly worse than zero as a result of beneath the inflection your marginal return sits underneath your goal, and also you’ve sunk cash to get nowhere.

The choice isn’t “how a lot.” It’s binary: commit previous the edge, or don’t fund it in any respect. There’s an actual minimal viable finances, and it’s typically above regular check budgets. You may’t sprinkle an S-curve and count on effectivity, and you’ll’t consider one on an underfunded check.

These two guidelines can appear like they combat one another, however that’s solely true to a sure level. Previous the inflection, an S-curve is concave, so the equimarginal rule governs it precisely because it governs a real C. The S-specific instruction — commit a block as a substitute of sprinkling — is barely concerning the journey from zero to previous the inflection.

Form is subsequently largely a launch-and-evaluation drawback. Getting a brand new prospecting channel into its environment friendly vary requires a dedicated block and endurance with ugly early numbers. As soon as it clears the inflection, you handle it on the margin like every little thing else, proper up till you take into account slicing it exhausting, the place form issues once more as a result of the draw back is a cliff, not a ramp.

That is the half that’s genuinely counterintuitive, and it echoes the unique marginal-return level: The appropriate transfer isn’t all the time the one that appears most effective at a small scale.

Which channels are which?

The historic default was concave. Simon and Arndt reviewed greater than 100 research and concluded that promoting follows the legislation of diminishing returns, a concave response. 

The dissent got here later: Vakratsas, Feinberg, Bass, and Kalyanaram discovered that threshold results do exist and that response shouldn’t be essentially globally concave. Their rationalization for why thresholds have been so exhausting to search out is the helpful half. Mature accounts already function contained in the efficient vary, so the warm-up by no means exhibits up within the information, and most research match a concave mannequin (the double-log) that may’t reject an S-curve even when one is current.

The platform shift has made the edge seen once more. Here’s a fuller map, ordered roughly from C to S. The form column is an inference from how every system targets and learns, not a measured fixed, and the appropriate form on your account nonetheless must be measured.

Two rows do a lot of the work.

AI Max is the dwell instance of a channel migrating from C towards S. Swapping specific key phrases for broad and keywordless matching means it wants conversion quantity to be taught which queries convert, so beneath a knowledge threshold it explores badly (Search Engine Land, ‘AI Max for Search,’ 2025). The combined unbiased outcomes match that: Google experiences about 14% extra conversions on common and as much as 27% for exact-match-heavy campaigns, whereas unbiased testing experiences 84% of advertisers seeing impartial or detrimental outcomes (ALM Corp, January 2026). A lot of that unfold is accounts that turned it on with out the conversion quantity to clear the educational area.

Efficiency Max is the lure, as a result of its curve is a composite. It blends a harvesting layer (branded, retargeting, Purchasing in opposition to present intent) with a prospecting layer (keywordless growth throughout surfaces). The harvesting layer is an inexpensive C that pays off on the primary greenback. The prospecting layer is the S beneath.

Blended, the early effectivity seems nice, since you are largely skimming demand you already had, and the typical hides the prospecting warm-up fully. That can also be why the platform is glad to optimize it for you: the mix flatters the headline quantity. You may’t learn PMax or run the form evaluation on it till you break up the harvesting from the prospecting.

The throughline runs in two layers. Guidelines-based auctions seize the very best stock first, which yields concavity; machine-learning programs have to be fed earlier than they’re environment friendly, which introduces a threshold. Beneath each, harvesting present demand is concave and largely non-incremental, whereas creating new demand is the S-shaped half the place the true development and the true warm-up value each sit.

Common versus marginal: whole over spend, or the slope the place you stand.

What you allocate in opposition to is marginal incremental return, the slope of the incremental curve at your working level. A holdout fixes the primary axis solely. Time-sliced marginal CPA on attributed information fixes the second solely. A multi-cell scaling check will get each, at a price. 

MMM (technique 1) estimates the entire curve from mixture information and sidesteps click on attribution fully, however pays in identifiability and modeling assumptions as a substitute. Most arguments about ‘what’s working’ are two individuals standing on completely different axes.

There are two main cautions, and I might flag each as genuinely unsettled relatively than settled info. 

• Separating a real S-curve from “concave with a excessive half-saturation level” is tough, as a result of a concave mannequin will match S-shaped information effectively sufficient to cover the inflection (that is the Vakratsas level, and it applies to your personal dashboards as a lot as to tutorial research). 
• The educational section could also be a one-time mounted value to coach the mannequin relatively than a everlasting characteristic of the steady-state curve. Whether it is transient, the channel could behave concavely on the margin as soon as it’s skilled, and the S you measured was a startup artifact. The reality might be a combination: a one-time coaching value, plus an ongoing minimum-volume requirement to remain environment friendly. Deal with each form name as provisional and re-check it.

Another failure mode, and this one shouldn’t be unsettled science however a matter of the place you might be standing on the curve. An S solely seems like an S in case your information spans the inflection. 

Above the inflection, an S is concave, mathematically equivalent to a C. Have a look at solely the $20,000-and-up rows of the desk above: marginal CPA rises monotonically from $18, a textbook C-curve, and the convex warm-up is invisible since you are not working in it. 

Established accounts normally sit previous the inflection, which is strictly why Vakratsas discovered thresholds so exhausting to detect, and why you possibly can run an S-shaped channel for years, appropriately, whereas believing it’s concave. The inform arrives the day you chop exhausting and fall off the inflection as a substitute of easing down a slope.

When to go vast and when to go deep

The marginal-return submit advised you to equalize marginal CPAs throughout this system. That rule continues to be right, however the form of the curve tells you the way you’re allowed to get there. 

• On C-shaped channels, you will get there by sprinkling, as a result of each greenback is productive and breadth is the pure reply. 
• On S-shaped channels, you must commit a block of finances previous the inflection earlier than the channel earns its place, after which focus relatively than unfold.

Lay the harvest-versus-create lower on high. Harvesting channels (branded, retargeting, non-brand search) are your C-curves: fund the primary {dollars}, then cap them early, as a result of they saturate quick and a lot of the tail isn’t incremental, regardless of how robust the attributed ROAS seems. 

Prospecting channels (Meta, YouTube, LinkedIn, the growth half of PMax) are your S-curves and your solely actual supply of incremental development: commit previous the warm-up or don’t begin, and decide them on incremental elevate relatively than attributed CPA, otherwise you’ll kill the factor that was working.

Traditional search rewards going vast. PMax, AI Max, and Meta prospecting reward going deep on fewer bets and giving every sufficient quantity to clear the warm-up. Run an S-curve like a C-curve and also you’ll starve it, learn the underfunded consequence, and kill a channel that will’ve been one in every of your greatest.