Performance Max in 2026: how to actually control it.

You can control Performance Max. The levers are brand exclusions applied before launch, account-level negative keyword lists, asset group architecture built on margin rather than category, deliberate feed-only decisions, and the channel report that shows where the money actually went. What you cannot do is measure PMax the way you measure a Search campaign — three of the questions a good media buyer instinctively asks have no answer at any price. Most PMax accounts are managed badly because those two facts get confused: people accept the reporting as complete and the controls as absent, when the truth is the reverse.

Performance Max is not a campaign type. It is a budget router.

A Search campaign is one channel, one auction, one kind of intent. You can reason about it as a single thing because it is a single thing. Performance Max is one line in your account that buys inventory across six surfaces — Search, YouTube, Display, Discover, Gmail, and Maps — each with a different cost structure, a different level of purchase intent, and a different relationship to whether the sale would have happened anyway.

Then it reports one number for all of it. When your PMax campaign shows an 8.0 ROAS, that figure is an average across six businesses, and an average is the one statistic that tells you least about a mixed portfolio. It is entirely normal for a PMax campaign at 8.0 blended to contain a Search slice at 20-plus and a Display slice below 2, and the campaign will happily report the blend as a triumph while the marginal dollar goes somewhere you would never have funded deliberately.

Three things the reporting will not tell you — ever

There is a meaningful difference between data Google buries in an awkward menu and data that does not exist. Confusing the two wastes months. We pull PMax data through the API directly for our own reporting rather than reading it out of the interface, which is a useful vantage point: when a field is genuinely absent at the API level, no amount of interface archaeology is going to produce it. Three absences matter.

1. There are no keywords and no ad groups, by design

PMax campaigns are built from asset groups. There is no ad group equivalent and no keyword entity, which means a keyword report for a PMax campaign returns nothing — not because a sync failed or a permission is missing, but because the object does not exist in the campaign type. This trips up more audits than it should. If a consultant tells you your PMax keyword data is broken, they have told you something useful about the consultant. The equivalent question — which queries did this campaign actually serve against — is answered by the search terms report, which is a genuinely different and much coarser instrument.

2. Per-asset performance is a label, not a measurement

Google grades individual assets LOW, GOOD, or BEST. Almost everyone reads those as performance data. They are not. They are allocation signals — the system telling you which assets it has decided to serve more often — and Google does not expose a per-asset conversion or revenue figure for PMax at all. So the sentence "this headline drove $40,000" cannot be true of a PMax campaign, no matter who says it or what dashboard it appears in. The honest version is "this asset was labeled BEST, which means the system served it more, which is a claim about the system rather than about the asset."

The practical consequence is that all creative decisions in PMax happen at asset group grain. If you want per-creative measurement, you need a campaign type that provides it, or you need to isolate the creative in its own asset group and read the group. That is slower and less satisfying than a per-asset table, and it is the only version that is true.

3. There is no per-audience performance

Audience signals in PMax are inputs, not segments. Ask the API for audience-level performance on a PMax campaign and the request is rejected outright — not empty, rejected. This means the very common request "show me which audience segment performed best in PMax" has no answer available to anyone. Signals tell the system where to start looking. They do not come back with a scorecard, and any report that appears to give you one has manufactured it.

Sit with what that adds up to. Three of the four instincts a competent buyer brings to a campaign — check the queries, check the creative, check the audiences — return either nothing or something misleading. That is the real constraint of the campaign type, and it is not solved by trying harder in the interface. It is solved by changing what you measure.

The number that lies first: brand

Left ungoverned, PMax finds your branded search. Branded queries convert at very high rates because the intent already exists, so the campaign reports superb efficiency, the bidding leans further into the easy win, and the account ends up with its best-looking campaign doing very little incremental work. This is the single most expensive misreading in Google Ads, so it is worth doing the arithmetic properly rather than asserting it.

Take a month that looks excellent. Substitute your own figures as you read — the point is the shape of the calculation, not these numbers.

1. Reported: $6,250 spend, $50,000 revenue. ROAS 8.0. Every dashboard in the business is happy.
2. Split the brand slice out using the search terms report: brand accounts for $20,000 of revenue on $625 of spend. That slice alone is running at 32.0.
3. Which leaves the non-brand slice: $30,000 revenue on $5,625 spend. ROAS 5.33 — and this is the part of the campaign that is actually finding new demand.
4. Now ask the harder question about the brand slice. In a brand hold-out test, the large majority of that revenue arrives without the ad, because the customer was already searching for you. Assume 85% would have landed anyway: incremental brand revenue is $3,000, not $20,000.
5. Incremental total: $33,000 on $6,250 of spend. True ROAS 5.28.

8.0 reported. 5.28 real. A 34% overstatement — and the overstatement is not the damage. The damage is the decision it flips. If your target is a 6.0 ROAS, the reported number says scale this campaign aggressively and the honest number says you are already below target and should be tightening. Same campaign, same month, opposite instructions. Accounts get scaled into unprofitability on exactly this arithmetic, and everyone involved is looking at a real number the whole time.

You need two inputs to run this on your own account: the brand share of your PMax conversions, and a defensible estimate of brand incrementality. The first is in the search terms report this afternoon. The second requires a hold-out test, and if you have never run one, assume the branded revenue is largely non-incremental until you have evidence otherwise — that is the conservative direction, and being conservative here costs you far less than being wrong.

The control surface, in order of leverage

These are ordered by how much they change outcomes, not by how much attention they get in webinars. Most accounts do the last two and skip the first.

1. Brand exclusions, applied before the campaign goes live. Both account-level and campaign-level exclusions exist; use both. The step-by-step is in our brand exclusions walkthrough.
2. Account-level negative keyword lists, which now apply to PMax. This is your only defense against the obviously irrelevant traffic the system finds while exploring.
3. Asset group architecture built on margin tiers, not product categories. A hero product at 60% margin and clearance stock at 8% do not deserve the same bidding behavior, and putting them in one group asks the system to average two businesses you actively want treated differently.
4. Search themes and audience signals used as direction rather than targeting. They nudge where exploration starts; they do not constrain where it ends. Treating them as targeting is the most common source of "but I told it to target X" confusion.
5. Feed-only versus feed-plus-assets, decided per product set rather than inherited. Feed-only behaves far more like Shopping and gives you a narrower, more predictable placement mix — often the right answer for a catalog where the product images are the persuasion.
6. Geographic, language, and location-intent settings audited rather than copied from a template. Presence versus interest targeting quietly changes who you are buying, and it is almost never checked.

Decision rules, so this is actionable rather than merely true

• →Apply brand exclusions unless branded search is a trivial share of account clicks and you have no separate brand campaign to catch that demand. In practice: apply them.
• →Split asset groups when the margin spread across your catalog exceeds roughly 15 percentage points. Do not split because you have many products — SKU count is not a reason, margin dispersion is.
• →Before diagnosing a performance change, read the channel split. If YouTube and Display together hold more than about 30% of spend and the conversions there are view-through heavy, treat the campaign ROAS as unproven rather than bad.
• →Choose feed-only when the product image does the selling and you want a predictable placement mix. Add assets when you need PMax to reach demand that is not already shopping.
• →Change one structural thing at a time and leave it for two to three weeks. Not because learning is sacred, but because you cannot attribute an outcome to a change you made alongside three others.
• →If a proposed PMax change cannot be stated as "this should move that number in that direction by roughly this much," it is not a change, it is fidgeting.

Make the average decompose

The channel performance report is the most important thing Google has added to PMax reporting, and most accounts have never opened it. It shows the split of spend and conversions across Search, YouTube, Display, Discover, Gmail, and Maps — which converts the single blended number into something you can actually reason about. Read it before you form any theory about why performance moved. Roughly two thirds of the "PMax suddenly stopped working" cases we look at are a shift in channel mix rather than a change in campaign quality, and the fix is structural rather than a bid adjustment.

Alongside it: the search terms report weekly, placement exclusions maintained as a live list rather than set once, and asset group level performance read as the true creative grain. The discipline is not the looking. It is acting on what you see within the same week, because a placement report reviewed and not acted upon is theatre with extra steps.

The objections worth taking seriously

"Google says excluding brand starves the algorithm of conversion data." This is a real mechanism and mostly a misapplied one. The system does learn from conversions, and branded conversions are the least informative ones it can learn from — they teach it that people who already want you will buy from you. On a campaign with adequate non-brand conversion volume, removing brand improves the signal by removing noise. On a campaign genuinely starved of conversions, the honest answer is that PMax was the wrong choice at that volume, not that you should feed it brand to keep it fed.

"You are just moving brand revenue into a brand Search campaign — same money, different label." Correct, and that is the point. In a dedicated brand campaign that demand costs a fraction as much, it is reported separately so it stops flattering your prospecting numbers, and it no longer teaches the router that easy wins are what you want more of. The revenue does not change. Your ability to see what your prospecting is worth changes completely.

"Governance itself has a cost — every change resets learning." True, and it is why the rules above specify one change at a time with a wait. But note which way the cost runs: the expensive resets come from panicked reactions to a number nobody decomposed. Accounts that read the channel split and the brand share make fewer changes, not more, because most of the changes people make are attempts to fix a misreading.

Where this advice stops working

• →Brand exclusions are not surgical. They work on a list of terms, close variants leak, and a brand with many misspellings or a generic-word name will never get a clean separation. Verify with the search terms report rather than assuming the setting did what it says.
• →Everything above about creative is bounded by the per-asset measurement gap. We are not withholding a better method — there is not one. Any tool promising per-asset PMax ROAS is showing you allocation labels dressed as measurement.
• →The incrementality arithmetic needs a hold-out test to be more than inference. Without one, the 85% figure in the worked example is our working assumption from accounts we run, not a constant. Treat it as a starting point to be replaced by your own number.
• →The channel report tells you where money went, not what each channel was worth. There is no per-channel conversion attribution inside PMax, so a Display-heavy month is a flag for further work rather than a verdict.
• →Below roughly $3,000 to $5,000 a month of PMax spend, most of this governance is premature. There is not enough conversion volume for structural segmentation to mean anything, and the honest advice is a simpler campaign type until there is.

What good governance actually looks like

• →Before launch: brand excluded and verified, negative lists applied, asset groups segmented by margin, feed-only decision made deliberately, geo and language settings checked rather than inherited.
• →Weekly: channel split, search terms, placements. Act inside the week or do not look.
• →Monthly: recompute incremental ROAS with the brand slice removed, and compare against target. This is the number that governs budget, not the one in the campaign row.
• →Quarterly: revisit the margin tiers behind your asset groups, because catalogs drift and last quarter’s hero product is this quarter’s discount line.

When we scaled Awesome Books across millions of titles and multiple regions, this governance loop was a large part of how sales grew while ROAS and CPA targets held. The campaign type did not change. What changed was that nobody made a budget decision on a number that had not been decomposed first. The same discipline is what our Performance Max management work consists of, and the asset group structure guide covers the architecture layer in more depth.

The one sentence to take away

Performance Max is not a campaign you control, it is a budget router that reports one average and calls it performance — so the entire job is making that average decompose. Brand share, channel split, margin-tiered asset groups: three decompositions, and every one of them turns a number you have to trust into a number you can act on. Advertisers who do this outperform advertisers who do not, and the gap has almost nothing to do with knowing a clever setting.