Why your best network degrades as you scale it

A growth lead finds a network that returns well above target at a few thousand dollars a day, triples the budget, and watches the return slide toward the target and then under it within a month. The usual conclusion is that the network's inventory quality fell off, or that a competitor moved in, or that the model broke. The more accurate conclusion is that nothing about the network changed. The team moved along a curve that was always there and was reading it through an attribution lens that flattered the network at small scale and punished it at large.

The thesis is that channel degradation under scale is mostly two effects the team controls: the gap between average and marginal cost, and the way last-touch attribution redistributes credit as one channel grows. Both are predictable. Neither shows up in the network's own reporting.

Marginal cost is the number you never see

A network's model, given a small budget, spends it on the users it is most confident about. Those are cheap to convert and often high value. Increase the budget and the model has to reach further down its confidence ranking. Each additional dollar buys a slightly worse user than the previous one. The average CPI you see in the dashboard moves slowly, because it blends the excellent early users with the ordinary later ones. The marginal CPI, the cost of the last user you bought, moves fast.

This is why a network that "worked" at a small budget and "stopped working" at a large one is describing one curve at two points. More of the market's growth is bought than it was, which means more teams are pushing further along these curves than before, and more of them are meeting the marginal user for the first time.

Marginal cost is not reported anywhere by default. You compute it: take the change in spend between two periods and the change in installs or revenue across the same periods, and divide. If your average ROAS is at target and your marginal ROAS is well below it, you have already scaled past the point where the last increment paid for itself.

Attribution share moves in the wrong direction

The second effect is measurement. Under last-touch attribution, whichever channel delivered the final click before install gets the credit. At small scale, a network is rarely the last touch in a multi-exposure journey; it wins the credit only for the users it genuinely closed. Its attributed installs are a fairly honest count.

Scale the budget and the network's ads are now in front of far more users, including users who were also going to install via a competitor channel or organically. It becomes the last touch more often simply because it is present more often. Its attributed installs rise faster than its real contribution. Reported CPI looks better than it is, for a while.

Then the bill arrives. The users the network is now claiming include a growing share who were going to install anyway, and their downstream revenue would have arrived without the spend. Cohort ROAS on the scaled network falls, not because the network is buying worse users, but because more of its attributed users were never its users. Organic installs drop at the same time, which the team attributes to seasonality.

The second-order effect worth stating: scaling one network degrades the reported performance of the others. Their attributed installs fall as the big network takes the last touch, so the team cuts them, concentrating further in the network that is already past its marginal point. The portfolio narrows exactly when it should widen.

A step test that separates the two effects

A budget step test with a holdout distinguishes marginal cost from attribution drift, and it costs two weeks.

Pick the network, pick a geo or geo cluster where it runs, and split that geo into two comparable halves. Raise the network's budget materially in one half and hold it flat in the other. Everything else stays the same. After two weeks, compare three things across the halves:

• Total installs, paid and organic together. This is the incremental read. If the scaled half shows only a small lift in total installs against a large lift in attributed installs, the network is claiming installs that were arriving anyway.
• Marginal CPI, computed from the difference in spend and total installs between halves.
• Attributed installs on the other networks in each half. If they fell in the scaled half, you are watching credit reallocation, not performance.

As an illustrative example, a match-3 title doubling a network's budget in half of a tier-one market might see attributed installs on that network rise around 90 percent, total installs across all sources rise about 30 percent, and organics fall by a fifth. The network dashboard will call that a success. The marginal CPI computed from total installs is roughly three times the average CPI shown. The figures are invented, but the gap between the two readings is the finding.

What to do with the curve once you can see it

Once you have a marginal read, scaling becomes a question of where on the curve you want to sit, not whether the network is good. The decision rule: increase budget in steps of no more than a third at a time, re-run the marginal calculation after each step, and stop when marginal ROAS crosses your target even though average ROAS still looks comfortable. Average ROAS is what you report. Marginal ROAS is what you decide with.

This piece builds on the ground covered in "A cohort quality checklist before you scale UA spend" from last June, which dealt with whether the users were good. The point here is different: the users can be good and the increment can still be worthless.

The uncomfortable implication for the portfolio

If every network follows this curve, the optimal allocation is rarely the one that concentrates spend where average ROAS is highest. It is the one that equalises marginal ROAS across channels, which almost always means more networks at smaller budgets than the average-ROAS ranking suggests. That is harder to manage and looks worse in a summary deck. It is also where the money is, and the network you are about to cut because its attributed installs fell may be the one whose marginal return is currently the best in the stack.

Related archive reading

These articles provide related context and remain subject to their stated review status.

• Set UA budgets by marginal return, not share
• Frequency is a bidding problem, not a creative one