Average returns hide the margin
Every marketing report shows averages: total spend divided by total conversions, total revenue divided by total spend. These are honest summaries of what happened. They are poor guides to what will happen if spend changes.
The reason is that conversions are not equally cheap. In search, the cheapest conversions come from the most relevant queries and the people closest to buying; once those are captured, more spend reaches less relevant queries and lower positions. In paid social, delivery systems find the most responsive people in an audience first; more spend goes to people less likely to act and to people who have already seen the ad several times. The pattern holds across most channels: the first share of budget buys the best outcomes, and each additional share buys fewer.
The average blends the cheap early conversions with the expensive late ones, so it always looks better than the last share of spend. A channel can show a healthy average return on ad spend (ROASRevenue generated per unit of ad spend.) while its last share is well below break-even. Cutting that share would raise profit; adding more would lower it. The average cannot show this; only the marginal view can.
This is why forecasts built by multiplying a proposed budget by the current average return are so often wrong. They assume the next share of spend will perform like the average, when it will perform like the margin.
Diminishing returns and saturation
Diminishing returns describe a curve. At low spend, outcomes rise steeply with each increase. As spend grows, the curve flattens: each increase still adds outcomes, but fewer. Eventually it approaches saturation, where more spend buys almost nothing extra because the reachable, responsive audience is exhausted.
The shape differs by channel and context. Search saturates when the volume of relevant searches is captured; spend beyond that buys broader, less relevant traffic or higher prices for the same clicks. Paid social saturates more gradually, as frequency rises and creative fatigues within an audience. Channels that create demand rather than capture it may show slow early returns, because effects build over time, which is why the long-term work on advertising effectiveness published through the Institute of Practitioners in Advertising cautions against judging brand-building spend on short-term response alone.
Signs that a channel is near saturation include cost per click or per thousand impressions rising faster than spend, frequency climbing within the same audience, impression share already high on core terms, and new conversions increasingly coming from people who were going to buy anyway. That last sign is the hardest to see in platform reports, because platforms claim credit for conversions they touched whether or not they caused them. Incrementality (IncrementalityConversions caused by marketing.) tests, which compare exposed and unexposed groups, are the reliable way to check.
Marginal cost per acquisition
Marginal cost per acquisition is the extra spend divided by the extra acquisitions it produced. It is the figure to compare with the break-even and target cost per acquisition (CPACost of each conversion.) from the contribution margin and break-even guide. As long as the marginal figure sits below target, more spend adds profit; once it rises above, the extra spend destroys it, even if the average still looks fine. Platform and budget pages shorten the measure to marginal CPAThe extra spend divided by the extra acquisitions it produced when spend changed; usually above the average, and the figure scaling and cutting decisions need.; the division is the same.
There are three practical ways to estimate it without a statistical model. First, compare periods where spend changed materially and other conditions were similar, and divide the change in spend by the change in conversions. Second, run planned budget steps: hold a campaign at one level for a few weeks, raise it, hold again, and compare. Third, use the platform's own forecasting and simulation tools as a guide, treating them as the platform's estimate rather than a measured result.
Each method has noise. SeasonalityPredictable, repeating changes in demand tied to the calendar, such as months, holidays, weekdays or hours., promotions and competitor activity move outcomes independently of spend, so estimates are best read as ranges. The example below shows the arithmetic for a simple budget step.
Marginal cost per acquisition in a budget step
Lowmoor Insurance spends £20,000 a month on paid search for home insurance and records 400 policies, an average cost per acquisition of £50. Its target cost per acquisition, from contribution margin, is £70. Encouraged, the team raises the budget to £30,000 for a month with similar seasonality and records 520 policies. The new average is about £58, still under target. But the extra £10,000 bought only 120 extra policies, a marginal cost per acquisition of about £83, well above the £70 target. A second step to £25,000 produces 470 policies: the marginal cost of that step is about £71, close to target. The team settles at £25,000 and moves the remaining £5,000 to a channel with a lower marginal cost.
Forecasting outcomes from spend scenarios
A useful spend forecast gives outcomes for a few budget levels rather than one, and builds diminishing returns into each. It does not need a model to be better than the average-times-budget approach; it needs the marginal estimates above and honest judgement about where the curve flattens.
| Scenario | How to estimate outcomes | Main risk | What to tell finance |
|---|---|---|---|
| Current spend held | Recent average outcomes adjusted for seasonality | Market or competitor change | Baseline: most confident |
| Modest increase | Current outcomes plus the increase divided by the estimated marginal cost per acquisition | Marginal cost rising faster than expected | Likely range, with the marginal cost assumed |
| Large increase | Several steps, each at a higher assumed marginal cost; flatten towards saturation | Saturation; creative fatigue; poor-quality volume | Wide range; recommend a staged test first |
| Modest decrease | Current outcomes less the decrease divided by the marginal cost of the last share | Losing outcomes that were cheaper than thought | Outcomes lost are usually fewer than the average implies |
| New channel | Small test budget, then extrapolate cautiously | No history; learning periods | Treat as an experiment, not a forecast |
Notice the asymmetry. Because the last share of spend is the least productive, a cut usually loses fewer outcomes than the average suggests, and an increase gains fewer. Forecasts that ignore this overstate both the pain of cuts and the benefit of increases, which is exactly the bias finance learns to distrust. Money for new channels sits outside these curves, and the 70-20-10 budget rule is one convention for reserving it while marginal returns decide how the proven spend is split.
Link the scenarios to the budget: each scenario's outcomes, multiplied by contribution per acquisition, gives a contribution forecast finance can compare with the cost. The ROAS calculator and the CPA calculator help check each scenario's implied rates, and a KPI tree shows how the channel outcomes roll into revenue and profit. Splitting a total across channels using these curves is the job of the budget allocation guide. The tutorial on building a planned marketing P&L carries a contribution forecast like the one above through the year by period, with its timing and the assumptions behind it. Organic search needs a different build, because its costs come first and its clicks follow positions rather than spend; the SEO forecasting guide builds its baseline, new opportunity and scenarios from the site's own Search Console data. A business with no spend history has no curve to read yet, and the guide to marketing decisions before you have data shows how its first estimates are set from the margin and replaced as measured points arrive.
Where spend should stop
The economic answer is simple: spend should rise until the marginal cost per acquisition reaches the target cost per acquisition, or equivalently until the marginal return on ad spend falls to the target return. Below that point, each extra amount adds profit; above it, each extra amount reduces profit.
In practice three things complicate it. The marginal cost is estimated, not known, so the stopping point is a range. The target itself may shift: a business prioritising growth may accept a higher marginal cost for a period, provided the trade-off is explicit and finance agrees. And some spend has effects that do not show in the short-term marginal cost, such as brand-building work whose return accrues over longer horizons; for that spend, the stopping rule needs longer-term evidence. A business that wants paid spend to follow this rule rather than a fixed monthly figure can write it into an agreement with finance on a target-led paid media budget.
Lifetime value changes the target, not the logic. If customers acquired at the margin are as valuable over time as the average, the target cost per customer acquisition (CACCost to acquire one new customer.) may be higher than first-order break-even. But marginal customers are often less valuable, because they were harder to persuade. Check before assuming.
Platform scaling advice describes how to raise budgets without disrupting delivery; the Meta Ads guide to scaling campaigns is an example. That is the mechanics of moving along the curve. Deciding how far to move is the economic question this section answers. When the answer is to move money from one line to another mid-period, the budget reallocation memo template sets the marginal estimates for both lines, their caveats and the expected effect as a range beside the decision.
Testing the forecast against reality
A forecast is a hypothesis. Each month, compare actual outcomes with the forecast for the spend that actually happened, not the spend that was planned. Persistent gaps in one direction mean the assumed curve is wrong and needs rebuilding. The customers each spend scenario produces are also the acquisition input to the tutorial on building a growth model in a spreadsheet, which carries them through activation and retention to show what each scenario does to the active base.
Run deliberate tests where the stakes justify it. Budget steps, as in the example, are the simplest. Geographic holdouts, where spend changes in some regions and not others, give a cleaner read of incremental effect. Incrementality tests on individual platforms show how much of the attributed outcome was caused by the spend, which is the difference the comparison of incremental and platform ROAS sets out.
For larger budgets spread across many channels, marketing mix modelling (MMMStatistical model of channel impact.) estimates response curves for each channel statistically, controlling for seasonality, price and other factors. It is the natural next step once simple methods have shown that marginal returns matter. The analytics root's marketing mix modelling basics explains how it works and when it is worth the investment; this guide deliberately stays with the logic that applies before any model is built. The unit economics hub links this guide back to margin and break-even.
When that step is taken, the tutorial on preparing data for a marketing mix model shows how weekly spend, outcomes and controls are assembled so that the curves have real variation to learn from.
When a budget has already been raised and sales did not follow, the diagnosis of ad spend up but sales flat isolates the added money and checks saturation, retargeting absorption, bidding disruption, operational limits and timing.
After a period closes well short of its forecast, the missed marketing forecast diagnosis splits the gap into volume, rate and timing and checks each assumption the forecast rested on.
To size the next budget step at the marginal cost, the marketing budget calculator set up for scaling spend enters only the extra conversions wanted and reads its total as an increment.
Common mistakes
- Forecasting with averages. Multiplying a new budget by the current average return overstates what increases will buy and what cuts will cost.
- Comparing marginal performance with the average target. The marginal cost per acquisition must be compared with the margin-based target, not with the channel's own average.
- Ignoring seasonality in budget steps. A step taken into a peak month looks better than it is; one taken into a quiet month looks worse.
- Trusting platform-attributed increments. Platforms claim conversions they touched; incremental tests show what spend actually caused.
- Scaling until the platform stops spending. Delivery limits are not economic limits; spend should stop well before saturation when the marginal cost passes target.
- Judging all spend on short-term marginal cost. Brand-building effects accrue over longer periods and need a different evidence standard.
