Performance Max was designed so you cannot see inside it. One campaign, every Google surface, and a report that tells you it worked without telling you how. Plenty of marketers have made peace with that trade. The results are often genuinely good, the setup is genuinely simple, and questioning a thing that performs feels ungrateful.
I audit it anyway, because opaque and unaccountable are not the same thing, and because the design deserves a moment of clear-eyed appreciation before you trust it. PMax is the first major ad product where the seller of the inventory also controls the campaign structure, the placement decisions, the bidding, and the report card, with no line-item visibility as a matter of design rather than limitation. Google built the black box on purpose, and Google's purposes and yours overlap without being identical. That is not an accusation. It is a description of a counterparty, and you audit counterparties. There are seams in the box, and what leaks through them changes how much credit the campaign deserves.
Three seams to pull
Start with brand, because it is the biggest number and the fastest test. Pull your brand terms into a separate exact-match campaign or exclude them properly, then watch what happens to the PMax numbers. In most audits I run, a meaningful slice of that beautiful ROAS was your own name being sold back to you, which is the self-grading problem wearing a different badge. The campaign did not find those customers. They were walking in the door, and the black box stepped in front of them to collect a commission. Every PMax audit should begin here, and the ones that do routinely revise the campaign's true incremental return down by a third or more.
Then read the channel distribution scripts and placement reports that do exist. Google exposes more than the default interface shows; the developer community maintains scripts that decompose PMax spend by surface, and the split between Search, Shopping, YouTube, and Display tells you what you actually bought. This matters because the surfaces are not equivalent assets. A PMax campaign winning real Shopping auctions is buying demand. One quietly leaning on cheap Display impressions to pad its conversion count is buying attribution, and the two behave completely differently the day your market softens. The report that says both worked is telling the truth about the past and nothing about the future.
Third, feed it honest signals, because this is the one lever inside the box you fully control. PMax optimizes to whatever conversion you hand it, with all the enthusiasm of a system that does not know your margin. Hand it form fills and it will find you form fills, including the ones that never become customers. Hand it qualified revenue events, offline conversions imported from the CRM, cart values with real margin data, and the black box starts working for the business instead of the dashboard. Cheap conversions are the most expensive thing you can teach an algorithm to want.
The posture, stated plainly
Here is the position I hold that the PMax debate keeps missing: the choice was never automation versus control, and the marketers framing it that way are fighting the wrong war. Automation won; the results are too good and the alternative too labor-intensive for the old hand-built structures to return. The live question is whether you operate the automation with verification or with faith. Verification costs a brand-exclusion test, a placement script, and honest conversion data, maybe two days of work a quarter. Faith costs nothing until the quarter the black box's incentives and yours diverge, and by then the divergence has been compounding in the dark for as long as nobody looked.
None of this requires Google to open the box. It requires you to stop treating the lid as a reason not to look. The advertisers getting hurt by automation are not the ones using it. They are the ones who mistook convenience for a receipt.