- The black box argument is out of date. Channel-level reporting, full search terms and asset-level metrics all exist now, and most accounts still are not looking at them.
- Performance Max will find the cheapest conversions available. If branded search is in scope, that is where it goes, and your reported ROAS becomes fiction.
- Asset groups are the only structural lever you have. Treat them the way you would treat ad groups, not as a dumping ground for every image you own.
- Signals are hints, not targeting. Exclusions are the actual controls: brand exclusions, campaign negative keywords, placement and content exclusions.
- Everything here depends on trustworthy conversion data. If tracking is wrong, PMax will optimise confidently toward the wrong thing.
The black box argument is out of date
For its first two years Performance Max deserved its reputation. You handed over budget and creative, and got back a single row of aggregated numbers. Diagnosing anything meant running third-party scripts and inferring the rest.
That has changed substantially. Channel-level reporting, full search terms visibility and asset-level metrics have all arrived, campaign negative keyword capacity expanded dramatically, and search themes per asset group increased. The constraint on most accounts is no longer what Google shows you. It is that nobody has gone looking.
| What you can now see | Where | What it answers |
|---|---|---|
| Channel performance | Insights and reports | How much went to Shopping, Search, Video, Display, Discover |
| Search terms | Insights and reports | Which queries triggered your Search and Shopping placements |
| Asset group performance | Asset groups, table view | Which creative theme is carrying the campaign |
| Asset-level metrics | Asset details | Impressions, clicks and cost per individual asset, not just Low or Best labels |
| Placements | Insights and reports | Where display and YouTube impressions actually landed |
| Listing group performance | Asset group, retail campaigns | Which products are absorbing the budget |
It is still less granular than a Search campaign. You cannot see which specific creative drove which conversion on a non-search channel, and a share of low-volume search terms remains hidden. But the gap between what is knowable and what most advertisers know is now almost entirely a question of effort.
What Performance Max actually is
Structurally, PMax is thinner than it looks. One campaign, a budget, a bid strategy, and beneath it a set of asset groups. That is the whole hierarchy, and asset groups are your only real structural lever.
Decide what PMax is actually for
The most expensive PMax mistake is not a setting. It is running PMax and Search over the same intent, then reading the combined result as growth.
Both campaign types can serve the same query. The auction resolves by Ad Rank rather than by a rule you control, which means a PMax campaign with broad scope will absorb whatever your Search campaigns used to capture, report it as its own performance, and leave you concluding that Search stopped working.
For retail with a large catalogue and a healthy feed, PMax is usually the right primary vehicle. For lead generation with a small set of high-value services, Search should normally lead and PMax should support. For accounts under roughly $3,000 a month in spend, PMax often struggles to gather enough conversion data across five channels to optimise any of them well.
Fix the data before you touch the campaign
PMax is more dependent on conversion data quality than any other campaign type, because it is making allocation decisions across five channels simultaneously with no manual override. Feed it a conversion action that counts phone number clicks and it will find you an enormous number of phone number clicks, on the cheapest inventory available.
- One primary conversion action, reflecting the outcome the business is paid for.
- Values attached, so target ROAS means something. Without values, PMax treats a $200 order and a $2,000 order identically.
- Offline outcomes imported for lead generation, otherwise the algorithm optimises toward form fills rather than customers.
If any of that is missing, stop here and read the conversion tracking guide first. Nothing in the rest of this page will compensate for a campaign optimising toward the wrong number.
Build asset groups deliberately
The default instinct is one asset group with everything in it. That produces a campaign optimising a single blended average across products with wildly different margins and audiences.
Split on economics, not aesthetics
The useful question is not "do these look similar" but "would I be happy for budget to flow freely between these". Split when the answer is no.
| Split by | When it applies | Why |
|---|---|---|
| Margin band | Retail with variable margins | A blended ROAS target overspends on low-margin lines and starves high-margin ones |
| Product category | Distinct catalogues | Creative and search themes can actually be relevant |
| Service line | Lead generation | Emergency work and planned work need different messaging and different lead values |
| Best sellers versus long tail | Large catalogues | Stops the algorithm ignoring 80 percent of the feed |
| Image dimensions or ad format | Never | Not an economic distinction; just fragments your data |
Do not over-fragment
Each asset group needs enough conversion volume to learn. Splitting a campaign with 40 conversions a month into six asset groups gives you six groups that never leave learning. As a rough floor, aim for at least 15 to 30 conversions per month per asset group before splitting further.
Fill every slot
Asset groups serve across five channels, and missing formats simply reduce where you can appear. Supply the full complement of headlines, long headlines, descriptions, and images in every aspect ratio, and supply your own video. If you do not, Google will auto-generate one, and auto-generated video is rarely something you would have approved.
Take branded traffic out of the equation
This is the single highest-impact control in the campaign, and the one most often left untouched.
Performance Max is designed to find conversions efficiently. The most efficient conversions in any account are people searching your brand name, who were going to buy anyway. Left unchecked, PMax will consume that traffic, report a spectacular return, and mask whatever is happening on the incremental spend.
Use brand exclusions, not negative keywords
Google provides a dedicated brand exclusion control, and it is materially better than a literal negative keyword for this job because it is built to recognise misspellings, variants, subsidiary brands and non-Latin script versions of a brand name. A negative keyword blocks the string you typed; a brand exclusion understands the entity.
Use both, for different jobs. Brand exclusions for your brand and, where appropriate, for competitor brands you do not want to pay for. Campaign negative keywords for query patterns: "free", "jobs", "how to", "DIY", "cheap", or whatever your search terms report actually shows.
Then run a real brand campaign
Excluding brand from PMax does not mean abandoning branded search. It means putting it in a dedicated exact-match Search campaign where you can see its cost, control its budget, and stop it flattering everything else.
Understand what signals and themes actually do
Two features look like targeting and are not. Getting this wrong wastes weeks.
Audience signals
An audience signal tells the algorithm where to start looking. It does not restrict who sees your ads, and PMax will serve well outside the signal once it finds patterns it prefers. That makes signal quality a matter of accelerating learning rather than controlling reach.
Rank your signals by how close they are to real purchase behaviour: your own customer lists and converter data first, then site visitor segments, then custom segments built on competitor and high-intent search behaviour, and only then Google's in-market and affinity audiences.
Search themes
Search themes tell PMax which query territory you want covered, particularly where you have no historical data to draw on. The per-asset-group limit was raised substantially, so there is room to be specific. They are directional input rather than keywords, and they do not override your negatives or exclusions.
Signals will not save a bad feed
For retail campaigns the product feed carries more weight than any signal you can supply. Titles, product types, GTINs, image quality and custom labels do more to determine where PMax spends than the audience list you attached. If the feed is thin, fix the feed first.
Get the numbers out and look at them properly
Native reporting covers most of what you need now. For anything beyond it, or for reporting across multiple campaigns at once, the API is straightforward. This query returns asset group performance for the last 30 days, which is the view the interface makes tediously slow to assemble.
SELECT
campaign.name,
asset_group.name,
asset_group.status,
metrics.cost_micros,
metrics.conversions,
metrics.conversions_value
FROM asset_group
WHERE segments.date DURING LAST_30_DAYS
AND campaign.advertising_channel_type = 'PERFORMANCE_MAX'
ORDER BY metrics.cost_micros DESCWhatever route you take, the review sequence is the same every month, and in this order:
- Channel split. Where did the budget go, and what did each channel return? Video and Display absorbing spend at a fraction of the campaign average is the most common finding.
- Search terms. How much is branded? How much is irrelevant? Every junk term here is a negative keyword you have not written yet.
- Asset groups. Is one group consuming most of the budget? Is another getting nothing, and if so is that a data problem or a genuine verdict?
- Listing groups. For retail, are your best-margin products actually getting impressions?
- Placements. Are display impressions landing on apps and content you would not have chosen?
Apply exclusions like you mean it
Exclusions are where the actual control lives. This is the full set worth configuring on every campaign.
| Control | Level | Use it for |
|---|---|---|
| Brand exclusions | Campaign | Your brand, and competitor brands you do not want to bid on |
| Negative keywords | Campaign, and shared lists | Query patterns: free, jobs, DIY, second hand, research intent |
| Placement exclusions | Account | Junk apps, made-for-advertising sites, unsuitable channels |
| Content suitability | Account | Inventory type, sensitive content categories |
| Location exclusions | Campaign | Areas you cannot service, and presence versus interest targeting |
| Listing group exclusions | Asset group | Out of stock, discontinued or unprofitable products |
Set location targeting to presence rather than presence or interest unless you have a specific reason. "Interest" includes people merely researching your area from elsewhere, and in service businesses it is a reliable source of enquiries you cannot fulfil.
Bidding and budget without the whiplash
Start on Maximise Conversions or Maximise Conversion Value with no target while the campaign gathers data. Once you have a stable cost per acquisition or return, set a target that reflects reality rather than ambition. A target set 40 percent above current performance does not make the campaign more efficient; it makes it stop spending.
Give any significant change two to three weeks before judging it. Changing the target, the budget and the asset groups in the same week guarantees you will never know which one moved the numbers.
Test it properly, or you are just guessing
PMax reports its own performance using its own attribution, which makes "is this incremental" almost impossible to answer from inside the campaign. Two methods actually work:
- Campaign experiments. Split traffic between a control and a variant to test structural changes, targets or asset group configurations with a clean comparison.
- Geo holdouts. Pause the campaign in a set of comparable regions and compare total business outcomes, not platform-reported conversions. This is the only method that answers the incrementality question honestly, and it is worth the disruption once or twice a year.
Both need a fixed test window agreed in advance. The failure mode is stopping a test in week two because the numbers look bad, which is exactly when a re-learning campaign always looks bad.
Common failure modes
| Symptom | Likely cause | Fix |
|---|---|---|
| Spectacular ROAS, flat revenue | Campaign absorbing branded search | Apply brand exclusions, run brand separately, re-baseline |
| Search campaigns collapsed after launch | PMax winning the same auctions | Define the boundary with exclusions and search themes |
| Most spend on Display and Video | Cheap inventory offering cheap conversions | Check channel report, exclude placements, tighten conversion definition |
| One asset group takes everything | Uneven asset quality or conversion history | Rebuild the starved group's assets, or separate into its own campaign |
| Campaign stopped spending after a target change | Target set beyond achievable performance | Revert, then move targets in 10 to 15 percent steps |
| Leads arrived, none were qualified | Optimising toward form fills with no offline data | Import qualified and won outcomes, attach real values |
| Best products get no impressions | Listing group structure or feed quality | Segment listing groups, fix titles, product types and custom labels |
Frequently asked questions
Can I see which channels Performance Max is spending on?+
Yes. Channel-level reporting shows how spend and conversions break down across Search, Shopping, YouTube, Display, Discover and Gmail, alongside search terms, asset-level metrics and placement reports. It is less granular than a Search campaign, but the old black box argument no longer holds.
Should I exclude my brand from Performance Max?+
In almost all cases, yes. Otherwise PMax absorbs traffic that would have converted anyway and reports it as its own performance. Use brand exclusions rather than negative keywords, since they recognise misspellings and variants, and run branded search as a separate exact-match campaign.
Do audience signals control who sees my ads?+
No. Signals tell the algorithm where to start looking, and it will serve outside them once it finds patterns it prefers. They speed up learning rather than restrict reach. Exclusions are the actual controls.
How many asset groups should I have?+
As few as the economics allow. Split when budget should not flow freely between two sets of products or services, usually by margin band, category or service line. Each group needs enough conversion volume to learn, so roughly 15 to 30 conversions a month is a sensible floor before splitting further.
Is Performance Max worth it for a small budget?+
Often not. PMax allocates across five channels at once, and below roughly $3,000 a month there is rarely enough conversion data to optimise any of them well. Small accounts usually do better concentrating on Search and Shopping first.
How long before a change shows real results?+
Two to three weeks for most structural changes, longer if conversion volume is low. Change one thing at a time, and resist judging a re-learning campaign in its first fortnight.
Where this fits
Performance Max rewards the accounts that have already done the unglamorous work. Trustworthy conversion data comes first, then a clear boundary between PMax and Search, then disciplined exclusions and a monthly reporting habit. Get those right and PMax becomes a genuine growth lever rather than a place budget disappears.
If you would rather have this reviewed than rebuilt from scratch, our Google Ads consulting engagements start with a full account audit, and the free audit will show you exactly where your PMax spend is going.