The marginal install costs more than your blended CPI
Blended CPI is an average over installs you already bought. The next install costs more, and the gap widens exactly when you decide to scale.
A UA manager who reports a blended CPI of 2.40 and asks for budget to double the campaign is making a claim they cannot support: that the next tranche of installs will cost roughly what the last tranche did. It will not. The figure on the dashboard is the average price of installs already won, most of them at the cheap end of the auction. The price of the next install is the marginal cost, and in every functioning auction the marginal cost sits above the average. The gap between the two is the real price of scale, and almost nobody measures it.
The pushback is predictable. Buyers will say that algorithmic bidding smooths this out, that target-cost campaigns hold CPI flat as budget rises, and that the dashboard CPI has been stable for months. All of that describes the average. None of it describes the margin. A platform that holds your average CPI flat while you scale is doing so by winning cheaper impressions elsewhere to offset the expensive ones it is now buying at the top of your budget. The marginal install still cost more; the platform hid it inside a blend.
Why the marginal cost rises
An ad auction ranks the available impressions by what they cost you to win. At low budget you win the ones where your bid clears easily: users with high predicted conversion for your game, in placements with weak competition. As budget rises, those run out. The platform has to reach for impressions where more advertisers are bidding, or where its model is less sure the user will install, so it needs a higher effective bid to win the same install.
The result is a supply curve. Each additional install is bought a little further up it. The average lags the margin because the average still contains all the cheap installs from the bottom of the curve. A campaign whose blended CPI moves from 2.40 to 2.70 as spend doubles has not seen a 12 percent rise in cost. Illustratively, if the first 10,000 installs cost 24,000 and the next 10,000 lifted the blend to 2.70, those next 10,000 cost 30,000, a marginal CPI of 3.00, 25 percent above the old average. Doubling spend raised the marginal price by a quarter while the dashboard reported a rise of an eighth.
This is arithmetic, not a platform quirk. Any time the average rises as volume rises, the marginal unit costs more than the new average, not just the old one.
The bit the dashboard hides
Average CPI is a lagging, smoothed number by construction. It gets worse when the platform is optimising to a target it can meet by blending.
A target-CPI or target-ROAS campaign is told to hold an average. The cleanest way to hold an average while budget grows is to buy some very cheap installs alongside the expensive ones. Cheap installs in most genres come from lower-intent placements and geos, so the platform obeys the instruction by quietly changing the mix. The buyer sees a stable CPI and a slowly deteriorating D7 retention and often blames creative fatigue.
The second-order effect is that marginal cost and marginal quality move against you at the same time. The expensive incremental installs are the ones the platform was unsure about, and the cheap offsetting installs are lower intent. Both ends of the new volume are worse than the middle you already had. Blended CPI reports none of this, and blended ROAS, which an earlier UA Ledger piece on blended ROAS calculation took apart, reports it late.
Measuring the margin
The marginal cost of an install is the change in spend divided by the change in installs between two budget levels, holding everything else as constant as a live campaign allows. That is a measurement problem, and it has practical solutions.
The cleanest is a budget step test. Run a campaign at budget B for a week, then at 1.5B for a week, with creative and targeting frozen. The marginal CPI is the extra spend over the extra installs. Repeat the step and you have three points on the supply curve, which is enough to see its shape. Seasonality and day-of-week effects contaminate this, so run the steps on matched days or use a geo holdout as a control.
Cheaper but noisier: plot daily spend against daily installs over a period where budget varied naturally and fit a curve. The slope at your current spend is the marginal cost. Platforms that have been pacing you hard produce very flat scatter, which is itself informative.
The output you want is a number per channel: the marginal CPI at current spend and the rate at which it rises per increment of budget. That is the figure a scaling decision should rest on.
A decision rule for scaling
Compare the marginal install's cost to the marginal install's value, not the average to the average.
Illustrative case. A channel shows a blended CPI of 2.40 and a predicted 12-month LTV of 3.20 on the blended cohort. Average economics say scale. A step test shows the marginal CPI at 3.00, and the incremental cohort from the step retains 15 percent worse at D7, which the LTV model translates to a marginal LTV of roughly 2.70. Marginal economics say the next increment loses money, and the blended numbers will only reveal that after the cohort has aged and the money is gone.
Put into a rule:
- Scale a channel only while marginal predicted LTV exceeds marginal CPI by your required margin.
- Treat a flat blended CPI under rising spend as a mix change to investigate, not as evidence of headroom.
- Re-run the step test whenever the platform changes its optimisation model, because the supply curve moves with it and the average will not tell you it moved.
What the marginal view changes upstream
Once a team measures marginal cost, the budget conversation changes shape. Instead of "we have 200,000 more, where does it go", the question becomes "at what spend level does each channel's marginal CPI cross its marginal LTV", and budget flows to the channel with the most room before the crossing. That is usually not the channel with the best blended ROAS. It is often a mid-tier channel with a flatter supply curve, where the tenth thousand installs cost about what the first thousand did. Such a channel is easy to spot once you have the scatter: the fitted line has a shallow slope, the platform has not been pacing you against a daily cap, and the D7 retention of the incremental cohort from the step test sits close to the base cohort rather than below it.
The last point is about incentives. A buyer judged on blended CPI is rewarded for keeping the average pretty, and the easiest way to do that is to let the platform pad the blend with cheap volume. A buyer judged on marginal CPI against marginal value has no such escape, and will start pushing back on scaling requests from above. That friction is the measurement working.
