The sentence we hear most about Performance Max: it is a black box, there is nothing we can do. Half of that is true. You genuinely cannot see how the distribution happens inside. But you still decide what goes into the box and what you measure coming out, and those two decide most of the outcome.
When PMax performs badly in an account we take over, the cause is rarely the algorithm. It is the input.
1. Feed quality
On Shopping and PMax this layer decides more than anything else. A product that errors in Merchant Center drops out of rotation and nobody notices, because the campaign keeps running.
What to check: disapproved products, missing GTINs, price mismatches, stock status, category mapping, image quality. The product title is a ranking signal too; if its first words do not describe the product, matching gets weaker.
2. Asset group structure
Putting the whole catalogue into one asset group tells the algorithm that everything is the same thing. If product groups differ in margin or in search intent, they need separate asset groups.
Each group needs its own images, headlines and descriptions. Copying the same text into every group leaves nothing to test.
3. Brand exclusions
The most commonly skipped setting. When brand searches are included, the campaign looks excellent because it takes credit for sales from people already looking for you.
Exclude brand terms and read the campaign on non brand demand only, and the picture changes. That is where real performance shows.
4. Channel and search term reports
The box is not completely dark. Channel distribution, search terms and product level performance reports are all accessible.
They answer two questions: where is the budget going, and which queries turn into sales. Budget drifting into display, in an account that sells on search, is usually not good news.
5. Audience and geography signals
An audience signal is a starting hint, not targeting. Uploading customer lists, site visitors and past buyers helps the algorithm learn faster.
On geography, the common mistake is advertising in regions you do not ship to. The setting takes a minute and the cost repeats daily.
6. Conversion value and targets
The algorithm learns what to maximise from the value you send it. Give every conversion the same value and selling a thin margin product becomes identical to selling a profitable one.
Send conversion values based on contribution margin and the campaign shifts toward products that earn. On its own, this is the single change that produces the biggest jump in most accounts.
A weekly reading routine
Managing PMax is a weekly reading job, not a daily intervention. We look at four lines:
- Spend and the share of non brand sales
- Disapproved products in the feed and the change from last week
- Contribution margin per product: is budget going to items that earn
- Channel distribution: is it drifting away from search
Four lines fit on one page and force decisions to come from data rather than instinct. Filling the same table every week, instead of collecting screenshots, is the only thing that shows what actually worked three months later.
What we see most often
- Brand not excluded, so the campaign applauds itself
- Products disapproved in the feed for months, including the profitable ones
- One asset group, hundreds of products, nothing testable
- Conversions counted twice, so targets are calculated wrong
- Changes every week, so the learning period never completes
Budget and the learning period
After structure, the thing that does the most damage is impatience. Following a structural change, relearning usually takes two to four weeks. Every intervention before that window closes resets the clock.
Three rules work in practice:
- Do not change the target by a large amount in one move; go gradually
- Raise budget in steps instead of doubling it at once
- Do not rebuild asset groups before the learning period ends
When you see a bad week, the first question should be whether you changed something or the market did. If the answer sits in a written change log, panic decisions do not happen.
Running it alongside search
You do not have to run PMax on its own. In a healthy setup, brand searches are captured in a separate search campaign while PMax chases non brand demand.
Two things happen at once: the cost of brand traffic stays controlled, and PMax performance is read against genuinely new demand. Opening a separate campaign for high margin product groups also guarantees budget actually flows there.
When PMax is the wrong tool
If the catalogue is very small, if conversion tracking is unhealthy, or if the daily budget is too low for the algorithm to learn, PMax is a poor choice. In those cases starting with a classic search campaign keeps control in your hands.
Equally, if the product page blocks conversion, no campaign structure compensates for it. Advertising brings traffic; the page makes the sale.
In short
What you lose in Performance Max is control over distribution. Control over the input remains, and that is where the difference is made. With a clean feed, separated groups, brand excluded and conversion values reflecting real profit, the box works considerably better.