When Performance Max quietly moves budget away from Search

A Performance Max campaign reports as one line: one CPA, one ROAS, one blended number. What that number hides is where the spend actually went. PMax can allocate budget across Search, Shopping, Display, Video, Discover, Gmail, and Maps, and a campaign showing a stable, on-target CPA can still be quietly reallocating toward inventory that converts far worse than the average suggests.

The check

The concrete check is a network-level split of spend against conversions. When the browse and discovery inventory -- Display, Video, Discover, Gmail, Maps combined -- absorbs a disproportionate share of cost relative to the conversions it generates, that is a real signal, even while the campaign-level average still looks healthy: the strong-converting slice of the campaign is quietly subsidizing the average.

A worked example

Illustrative case: a PMax campaign reports a stable, on-target CPA for two straight months. Split by network, Display and Discover together have grown from roughly 20 to 45 percent of spend over that window, while their share of conversions has barely moved. Search's share of spend has shrunk to compensate. The campaign-level CPA still looks fine because Search's strong conversion rate is carrying the average -- the shift only becomes visible once spend and conversions are split by network and compared over time, not read as a single snapshot.

The opposite failure mode

The opposite failure mode matters just as much and is easier to miss: too little of the budget actually reaching Search within a PMax campaign means the campaign leans heavily on lower-intent inventory instead of the channel most likely to convert. Neither direction shows up in the headline number Google surfaces by default.

Why this stays invisible for months

Google's own reporting does not put network-level cost and conversion split next to the campaign's headline metric by default, so the drift compounds quietly until someone deliberately goes looking for it.

Why a cadence matters more than a one-time check

A single network-mix check catches the drift at the moment it is run, but PMax keeps reallocating continuously as it learns -- a mix that looked healthy last quarter can shift again without any single change triggering it. Treating this as a check to run once, rather than a number to watch on a cadence, is how the same drift gets rediscovered from scratch every few months instead of caught early each time.