Performance Max: Take Back Control

The platform's goal is not your goal

Start with the mental model. Google's, Meta's, Apple Search Ads' main goal is to earn more money — that's the plain read from practitioners who've spent careers sitting across the table from platform teams pushing them to spend more. Picture a board meeting where someone asks how to make $100 billion, and the ad product gets designed backwards from that answer.

The pattern isn't new. Google Analytics started as a paid product (Urchin), went free once Google bought it, hooked 90–95% of the market, then started charging again — the same arc as Gmail. In paid ads, Google handed over control for years, then began taking pieces of it back. The conclusion isn't to abandon PMax; it's to never let it run unsupervised.

Never just let it go — especially PMax, when it's running.

The scale is why this matters: roughly 73% of brands run at least one PMax campaign. Independent 2026 tracking agrees in spirit — about 71% of Google advertisers now use PMax, rising to ~93% among retailers running Shopping ads. Be equally wary of industry-average ROAS: e-commerce sits loosely around 3–5x, but Smarter Ecommerce's 2026 Market Observer shows the median target ROAS climbing from ~4.7 to ~6.0 year-on-year — proof these "averages" are moving targets, not numbers to plan a budget around.

Stop grouping by product — group by intent

The most common PMax mistake in the wild is building asset groups around product categories — one for shoes, one for bags — and dumping the full catalogue inside each. The system then spends on irrelevant products because it has no signal for which ones you actually want to sell.

The fix is the demand bucket: group by intent, not catalogue. Running shoes for a training buyer are a different demand pocket than running shoes tied to a summer sale, different again from a high-intent "best basketball shoes" search. Each bucket gets its own asset group, its own tightly selected product list, and its own creative.

The best source for that intent data isn't Keyword Planner, SEMrush or Ahrefs — they mostly draw on the same stale dataset, a month or two old. The real goldmine is Google Search Console plus your own Search campaign query reports: the actual terms driving impressions, sometimes fresh from the last week. One travel client's team caught a surge of searches for discount group-deal travel sites that no keyword tool had, added it, optimized the landing page, and got cheap clicks before competitors noticed. For scale, push the daily bulk export of Search Console into BigQuery (the first 1 TiB/month is free) and cluster the terms with an LLM like Claude.

Don't skip your video assets, either. Leave the slot empty and Google auto-generates video from your product images — and since March 2026 layers on a default AI voiceover for any clip without one. Upload two or three of your own videos across vertical, square and landscape ratios, and nail the hook in the first three seconds.

Feed it segmented signals, and throttle the pace

Structure solves half the problem; the data you feed the algorithm solves the other half. The signal hierarchy, in order: offline conversions and lifetime value first, then Customer Match — segmented, not dumped as one list — then custom-intent audiences (competitor URLs, app-based targeting), then Google's own in-market and affinity segments. LTV look-alikes work far better built in segments than as one undifferentiated list. Use the New Customer Acquisition goal and exclude existing buyers, or PMax quietly becomes an expensive remarketing campaign wearing a prospecting label.

Treat budget and bidding changes as experiments, not universal percentage rules. First validate conversion tracking, value definitions and conversion lag. Then set a review window long enough to observe completed conversions, change one major variable at a time, and log the effect on cost and profitable revenue. A fixed conversion count or a fixed three-day interval is not a guarantee that a campaign has finished learning.

One structural fix worth knowing: budget doesn't distribute evenly across asset groups within a campaign, because Google spends at the campaign level. A star asset group can starve a merely-good one. Break standout performers into their own campaign so they get dedicated budget instead of quietly cannibalizing the group next to them.

Where the inflated ROAS hides

The uncomfortable payoff of all this control: PMax's headline ROAS is often real on the screen and fake in reality. It can spend on your own branded search terms and claim credit for conversions that would have happened anyway.

Adalysis analysed roughly 3,300 non-retail PMax campaigns and 1.2 million search terms to test the overlap directly. Only about 2.8% of all search terms overlapped between PMax and Search — but that touched roughly 67% of campaigns. On the terms where both competed, Search converted better ~84% of the time, delivered higher conversion value ~85% of the time, and a better CTR 65% of the time — while PMax simply grabbed more impressions, 61% of the time. Search usually wins the conversion; PMax often takes the credit.

Brand exclusions help, but they're not airtight. A separate study by hospitality agency three&six found that even accounts with exclusion lists applied still recorded brand-keyword clicks ~94% of the time, with brand terms making up ~23% of total clicks — because PMax keeps triggering on fuzzy brand variations and misspellings no list catches. That's why the discipline is both: apply brand exclusions and negative keywords, and run brand as its own governable Search campaign.

Keep Performance Max and AI Max for Search separate when auditing settings. In Search, opting into AI Max enables Final URL expansion by default, but you can switch it off during setup or in campaign settings. URL exclusions are another control, not a replacement for the off-switch. See Google’s Final URL expansion documentation (checked 4 September 2026). Do not assume that a PMax campaign migrates into AI Max for Search.

The EcomExpo playbook

The four levers convert into a short operating checklist for any PMax account your team runs:

• Audit for brand cannibalization on day one. Apply brand exclusions, add brand negatives, and split brand into its own Search campaign.
• Build asset groups by intent, not by product category, using a demand-bucket map sourced from Search Console.
• Feed segmented signals — offline conversions and LTV first, then segmented Customer Match, then custom intent.
• Document every change — record the hypothesis, conversion lag, observation window and profitability guardrail before changing budget or bidding.
• Govern final URL expansion in each campaign type separately; review the enabled setting and landing-page exclusions before changing it.

PMax is a black box only for those who agree not to look inside.

The platform's incentives are not yours, the controls exist but are often hidden or default-on, and the only reliable edge is to keep testing while everyone else just listens to the platform. Engineer the black box, don't just feed it.