Anyone who has run Performance Max for a while has lived the same scene. The campaign is built, the budget goes in, the first two weeks look great, week three gets shaky, and the answer is that the system is still learning. Ask what the team can do about it, and the remaining advice is to add assets and wait. Others don't buy it and run PMax like a manual campaign, editing assets and settings daily, until the learning period never finishes and performance gets worse.

Both sides share one problem, nobody has spelled out which variables still sit in human hands. The official documentation actually lists them in detail, just scattered across a dozen articles. This piece gathers them into a single setup-to-optimization path, and answers the harder question on the side, whether PMax really brought the gains attributed to it.

PMax automation takes over how and where money is spent. The inputs and boundaries stay in human hands. Goals and conversion definitions, feed data, assets and asset groups, audience signals, exclusions, and reporting and experiments each decide what the system treats as correct. It is neither a launch-and-forget campaign nor one that needs daily manual rescuing. The workable posture is to stand up the six control points in order, then verify results with controlled experiments. PMax does not promise a ROAS lift, and incrementality has to be proven. The broader account strategy sits in the optimization strategy guide.

PMax Campaign Setup Starts With Goals, Conversions, and the Feed

Every optimization in PMax orbits the goals and conversion values you set, so the first job is nailing the conversion definitions. Which actions count and what each is worth decides the direction the system learns, and inaccurate tracking means every automation downstream magnifies the error. The conversion tracking and attribution guide covers the full method, so this piece moves on.

For retail, the prerequisite is product data. With a Merchant Center feed in place, PMax can build assets and delivery around products, and the data path of account verification, sources, and field consistency is laid out in the Shopping and Merchant Center guide. This article starts once the feed is ready. Bid goal selection and budget guardrails belong to the bidding and budget guide, and stay out of scope here.

One priority rule deserves an early note. Search campaigns with an exact match keyword are prioritized over PMax for matching queries, and search themes carry the same priority as phrase and broad match. To defend specific terms, use a Search campaign rather than hunting for a switch inside PMax. That is lesson one in understanding where the controllable boundaries are.

PMax Asset Groups Decide the Ceiling

The asset group is PMax's creative container, made of images, logos, headlines, descriptions, videos, sitelinks, and audience signals, whether uploaded by you or generated by Google AI. The documentation suggests building each group around a single theme or audience, and a campaign can add more asset groups after launch to cover different themes.

One starting point worth clarifying, advertisers with a Merchant Center feed can launch PMax without supplying extra creative, but they may see a Poor ad strength rating, and the documentation recommends supplying assets so the campaign can perform across surfaces. The automation also leaves a control hatch. When text customization is on, AI generates headlines, descriptions, images, and videos from your landing page, domain, and existing creative, and those generated results appear in the asset report, where unsuitable text assets can be paused or deleted.

The asset combinations report drives creative iteration, listing the six best combinations per category, text, images, and videos, for each asset group, with previews of how those combinations look across channels. Beyond the asset group sits one more layer, on-site promotions get pulled automatically into Shopping and Performance Max ads, so promises on the page need to be real and redeemable, and the automated promotions checklist walks through the decisions.

Audience Signals Are Guidance, Not a Fence

The official definition is measured. Audience signals are suggestions added to help Google AI optimize toward your chosen goals, and they are optional. Available inputs include your own first-party data, customer lists, website and app visitors, custom segments built from keywords, URLs, or apps, demographics, and additional segments covering interests, habits, and active research.

One official note has to stay in view, PMax may show ads to audiences outside your signals when they show a strong likelihood of converting. Signals point the model in a direction rather than fencing delivery, and their value is measured by whether they help the system find conversions faster, not by whether impressions stay inside the list.

The steadier practice, feed the system the high-value audiences you already know, repeat customers, cart abandoners, custom segments built around core search terms, so it starts with decent clues, and accept that it will eventually reach beyond the list.

Exclusions Are the Few Hard Instructions

Only a few controls in PMax behave as hard instructions, and negative keywords and brand exclusions are the two that matter most, along with their boundaries.

Negative keywords apply to Search and Shopping inventory only, the query-driven surfaces. They can sit on a single campaign or at the account level, where the list automatically applies to relevant campaign types including PMax, capped at 1,000 entries per account, with a tool for estimating their impact. Brand exclusions follow the same logic, applying only to Search, Shopping, and YouTube search inventory to keep your ads away from brand queries you want to avoid.

The sequence runs opposite to instinct. Don't stack negatives before launch. Let the campaign produce real data, find the wasteful patterns in the search terms and asset reports, then decide what to exclude. Account-level settings and structure map to the account setup and structure guide. One reminder to close, these controls only cover query-driven inventory. Suitability problems on YouTube, Display, and Discover need other settings, and blaming negative keywords for waste on those channels sends the diagnosis in the wrong direction.

PMax Optimization, Read Reports, Run Experiments

PMax reporting covers assets, asset groups, campaigns, and combinations, with the recommended starting metrics being conversions, conversion value, conversion rate, and cost per conversion. The asset report carries an Added by column separating human uploads from AI-generated or enhanced assets, so iteration can start with what you control. One technical detail matters, conversions at the asset level are not split across components, so summing all assets produces a number larger than campaign totals, and that arithmetic should not drive decisions.

The real answer to whether PMax caused the results is experiments. The official toolkit has three uses, uplift experiments measuring the conversion or conversion value uplift from adding PMax alongside other campaign types, upgrade experiments simulating a budget shift from Standard Shopping, Dynamic Search Ads, or Display into PMax, and optimization experiments testing asset variations and settings such as Final URL expansion. Some experiment entry points only appear once an account has enough conversion volume, and thin data means tracking comes first.

When experiments are out of reach, incrementality still has workable routes. Split traffic into control and treatment and compare the difference between groups instead of reading one campaign's report. Run a geo split, PMax in one region and not in a comparable one, and compare total conversions. Use time splits, switching delivery on and off in blocks, while avoiding promo seasons and natural swings. And peel brand queries and direct traffic away from the numbers, since those conversions would largely happen anyway. Budget and cost math lives in the budget guide, and this piece stays on the judgment framework.

Common Misconceptions

Launching and walking away. Automation takes over execution, while the inputs and boundaries stay with you. Goals, assets, signals, exclusions, and reports all drift if unmaintained, and the system accelerates in the wrong direction.

Treating audience signals as targeting. The documentation says delivery goes beyond the signals, and demanding that impressions stay inside the list leads to wrong conclusions.

Using negative keywords against waste on every channel. They cover Search and Shopping inventory only, and mixing up their scope produces no results and no diagnosis.

Reading campaign-level conversions as the verdict. Brand queries and direct traffic mix in, asset-level numbers do not sum to totals, and the wrong reporting basis bends every optimization decision after it.