Performance Max Campaigns: How to Control the AI Black Box

The Problem: PMax Hides the Levers

Performance Max campaigns run across every Google surface — Search, YouTube, Display, Discover, Gmail, Maps — and the algorithm decides where your money goes. That's the pitch. In my experience managing $50M+ in ad spend across Meta, Google, Bing, and TikTok, most PMax campaigns I've audited fall into one of two camps: they either work or they burn budget — no middle ground. The difference? Whether you know the controls Google buries.

PMax is not a black box. It's a box with labels written in tiny font on the back panel. The controls exist. Most advertisers just don't use them.

The levers that actually matter: asset group structure, audience signals, brand exclusions, negative keywords, and budget strategy. Specifics on each below.

Asset Groups: Structure Controls Budget Flow

Asset groups are the single most underrated control in PMax. Most advertisers treat them as creative containers — dump in headlines, descriptions, images, and move on. But asset group structure directly impacts how the algorithm allocates budget across your offerings.

The mechanic: each asset group gets its own learning phase. The algorithm optimizes within each asset group independently. If you cram everything into one asset group, the algorithm picks whichever combination converts cheapest and doubles down — usually your best-known product with the strongest search demand. Everything else starves.

If you fragment into too many asset groups, each one gets thin conversion data and struggles to exit learning. The sweet spot depends on your catalog size, but the principle is consistent:

1 asset group per distinct product category or intent theme. Not per product. Not per SKU. Per category.
Each asset group needs enough conversion volume to exit learning. Google hasn't published an official threshold recently, but after managing $50M+ in spend across platforms, I've found the practical floor is roughly 30 conversions per month per asset group for tCPA or tROAS to stabilize. Below that, the algorithm flails.
Every asset group must be creatively complete. That means 20+ headlines, 5+ descriptions, multiple images, at least one video, and a solid final URL. I've consistently seen better performance from fully populated asset groups — the system seems to favor them.

For an Indian D2C brand selling apparel, a reasonable structure might be: one asset group for men's, one for women's, one for accessories. Not 15 asset groups for individual product lines. Not one asset group for the entire catalog.

The reason this matters for budget control: when you have well-structured asset groups with distinct creative and audience signals, you can see which themes the algorithm is favoring. With one monolithic group, you're flying blind.

Audience Signals: Seeds, Not Constraints

This is the single biggest misunderstanding in PMax management.

Audience signals — your custom segments, customer match lists, remarketing audiences, and in-market audiences — are starting points for the algorithm. They are not hard constraints. The system uses them to accelerate learning, then explores beyond them.

I've seen advertisers obsess over getting their audience signals perfect, as if adding the wrong segment will permanently redirect spend. It won't. Within a few weeks, PMax will have expanded well beyond your signals if the data supports it. The algorithm's exploration is a feature, not a bug.

Most guides miss the nuance: when to constrain and when to let the system explore.

When Audience Signals Help

New campaigns with no historical data. Strong signals (customer match, high-intent custom segments) get the algorithm to relevant users faster during learning.
Niche B2B or high-consideration purchases. If your total addressable audience is small, signals prevent the algorithm from wasting budget on low-relevance impressions while it figures things out.
Campaigns stuck in a performance rut. Refreshing audience signals can nudge the algorithm toward new pockets of converters it hasn't considered.

When to Let the Algorithm Explore

E-commerce with broad appeal. If you sell consumer goods, the algorithm often finds converting audiences you'd never think to target. Don't box it in.
Mature campaigns with stable performance. Once PMax has 50+ conversions per month and stable CPA/ROAS, your signals are mostly redundant. The algorithm has its own data.
When you're scaling spend aggressively. Tight audience signals limit the algorithm's ability to find incremental volume. Loosen them.

The practical approach: start with strong signals in weeks 1-3, then gradually remove artificial constraints as the campaign matures. I typically add 3-5 high-quality audience signals at launch and remove the most restrictive ones after the campaign exits learning.

Search Themes vs. Audience Signals: A Decision Framework

Search themes are a newer PMax feature that lets you add keyword-like phrases to guide the algorithm's Search and Shopping inventory targeting. They sit between audience signals (who) and negative keywords (who not) — they're positive signals for what queries to prioritize.

How I decide:

ScenarioUseWhy Launching a new product categoryAudience signals firstThe algorithm needs to know who buys this type of product before it can match queries Existing campaign missing obvious query categoriesSearch themesYou're telling the algorithm about what intent to capture — direct and efficient Competitor conquestingBothAudience signals for competitor brand audiences + search themes for competitor brand terms Brand defense (your own brand)Neither — use a separate Search campaignPMax will take credit for branded conversions you'd get anyway. Keep branded search in a dedicated campaign Broad awareness / reachAudience signals onlySearch themes are too query-specific for awareness goals

Search themes and audience signals are complementary, not competing. Search themes steer the query matching side. Audience signals steer the user profiling side. Use both when you have clear intent categories to target. Use neither when you want maximum algorithmic exploration.

Brand Exclusion: Your First Move in Any PMax Campaign

Add a brand exclusion list to every PMax campaign before you launch. Not after. Not when you notice cannibalization. Before.

PMax loves branded search. It's cheap, it converts at high rates, and it makes the algorithm's reported performance look great. The problem? Those conversions would have happened anyway through your existing branded Search campaigns. PMax is taking credit for traffic you already owned.

This isn't theoretical. Across the $50M+ I've managed, I routinely see 30-50% of PMax conversions attributed to branded queries when no exclusion is in place. The campaign reports a fantastic CPA, but your overall account CPA doesn't improve because you're just shifting conversions from one campaign to another.

How to set it up:

Create a brand exclusion list in your Google Ads account settings. Add your brand name, common misspellings, and any branded product names.
Apply it to every PMax campaign. This prevents PMax from showing ads on branded queries.
Keep your branded Search campaign running separately. You want full control over branded search — exact match keywords, specific ad copy, dedicated budget. PMax can't give you that granularity.
Monitor Search Impression Share on your branded campaign. If it drops after launching PMax without an exclusion, that's your cannibalization signal.

For Indian D2C brands, this is especially critical. Many have strong brand search volume from offline awareness or social media presence. PMax will happily vacuum up those queries and claim the credit.

Negative Keywords and Placement Exclusions: The Safety Net

Unlike brand exclusions (which are campaign-level), negative keywords in PMax operate at the account level. This means you need to be strategic — a negative keyword that protects one campaign might hurt another.

My approach:

Add obvious negative keywords at the account level: terms like "free," "cheap," "jobs," "salary" — anything that signals zero purchase intent across all your campaigns.
Use campaign-level placement exclusions aggressively. If you see PMax spending heavily on specific Display or YouTube placements with no conversions, exclude them. The Placement report in PMax is buried but it exists — find it under the "Insights" tab.
Review the Search Terms report weekly. PMax doesn't show you every query, but it shows you some. Look for patterns: irrelevant query clusters, competitor terms you don't want to bid on, or informational queries with no purchase intent.
Don't over-negativize. Every negative keyword reduces the algorithm's exploration space. If you're too aggressive, you choke off the system's ability to find new converting audiences. Add negatives for clear waste, not for "maybe waste."

The repeatable system: check the Search Terms report every Monday. Add 3-5 negative keywords maximum per week. Check the Placement report every other week. Exclude placements with 100+ clicks and zero conversions. This cadence keeps waste down without constantly disrupting the algorithm.

Budget Caps: Campaign-Level vs. Portfolio Bidding

Budget control in PMax is where most advertisers feel the most pain. The algorithm can overspend your daily budget by up to 2x on any given day (though it averages out over the month). During the learning phase, spending can be erratic. And if you're using portfolio bidding strategies, the budget allocation across campaigns is entirely algorithmic.

My rule:

Campaign-Level Budgets (For Control)

Use campaign-level budgets when you need to cap spend on a specific PMax campaign. This is the right choice when:

You're testing PMax for the first time and want to limit exposure.
You're running PMax alongside Search campaigns and need to enforce a specific budget split.
You have a PMax campaign targeting a specific product category that shouldn't absorb budget from others.

The downside: campaign-level budgets can limit the algorithm's ability to capture spikes in demand. If your daily budget is $500 and there's a surge of high-intent traffic, PMax stops spending at $500 regardless.

Portfolio Bidding Strategies (For Scale)

Portfolio strategies (like a shared tCPA or tROAS across multiple campaigns) let the algorithm shift budget between campaigns based on where it sees the best return. This works when:

You trust the algorithm's optimization across your account.
You have mature campaigns with stable conversion data.
Your primary goal is total conversions or total revenue, not budget allocation by channel.

The risk: the algorithm will always favor the campaign with the easiest conversions. If your branded Search campaign is in the same portfolio as PMax, the algorithm will pump budget into branded because it's cheap and high-converting — exactly the cannibalization problem we discussed earlier.

My recommendation for most accounts scaling past $100k/month: Run PMax on its own campaign-level budget. Run branded Search on its own campaign-level budget. Use portfolio bidding only for non-branded Search campaigns that share a similar intent profile. Never mix PMax and branded Search in the same portfolio.

The 30-Conversion Threshold

This is the root cause of most PMax failures I've audited.

PMax needs conversion data to optimize. Without it, the algorithm is guessing. The commonly cited threshold is 30 conversions per month for tCPA or tROAS strategies to function properly. Based on what I've seen across accounts, this is roughly accurate — campaigns below this threshold show wildly unstable CPA and frequent re-learning phases.

What this means in practice:

If your PMax campaign gets fewer than 30 conversions per month, switch to Maximize Clicks or Maximize Conversions without a target. Let the system gather data before constraining it with a CPA or ROAS target.
If you're restructuring asset groups, make sure each new group will still clear the 30-conversion bar. Splitting a 50-conversion/month campaign into three asset groups means each group gets ~17 conversions — below threshold. That's asset group fragmentation, and it's worse than having one slightly bloated group.
If you're a low-volume advertiser, consider whether PMax is the right campaign type at all. Standard Search and Shopping campaigns give you more control when data is thin. PMax's strength is scale, not precision at low volume.

Before and After: Controlling a Runaway Campaign

Same story, different accounts. I've seen this pattern repeat across e-commerce brands scaling on Google:

Before

A D2C brand launches PMax with one asset group containing their full catalog, no brand exclusion, broad audience signals, and a tCPA set at their account average. Within two weeks:

PMax is spending 40% of budget on branded queries
Reported CPA looks great ($12) but overall account CPA hasn't improved
Branded Search campaign has lost 30% of its impression share
PMax is showing ads on Display placements with 200+ clicks and zero conversions
Monthly spend is $45,000 with no incremental revenue vs. pre-PMax baseline

After

Same account, same budget. Changes made:

Added brand exclusion list (brand name + 5 variations)
Restructured from 1 asset group to 3 (by product category)
Added account-level negative keywords for clear non-buyer intent terms
Excluded 12 Display/YouTube placements with high clicks, zero conversions
Switched from portfolio tROAS to campaign-level budget with tCPA
Kept audience signals but removed the broadest (least relevant) custom segment

Results over the following 6 weeks:

PMax CPA increased from $12 to $22 (because branded conversions were removed — this is expected and healthy)
Overall account CPA dropped from $28 to $21 (because branded Search was back under control with its own budget)
Incremental revenue from PMax (non-branded) went from near-zero to $85,000/month
Branded Search impression share recovered to 92%
Monthly spend remained at ~$45,000 but was now generating actual incremental conversions

PMax's reported CPA going up was the good outcome. The artificially low CPA was branded search cannibalization, not genuine efficiency.

My Take: Human + Automation Beats Pure AI Bidding

I learned this the hard way on Meta before I ever touched PMax. I scaled Meta Ads accounts from $10k/month to $500k/month by combining manual bidding strategy with custom automation pipelines I built in Python. The automation handled budget pacing, audience refresh, and creative rotation — things Meta's native tools couldn't do well enough at scale. Advantage+ alone would have hit a ceiling around $100k/month. The custom automation pushed past it.

PMax is the same pattern. The algorithm is excellent at finding converting users within the boundaries you set. It's terrible at deciding what those boundaries should be. That's your job. The algorithm doesn't know that branded search should be protected. It doesn't know that certain placements damage your brand. It doesn't know that your asset group for premium products shouldn't compete for budget with your clearance category.

I saw this from the monitoring side too. I migrated reporting for multi-platform accounts from manual spreadsheet exports to a real-time dashboard pulling from 5 ad platforms via API, normalized in Python, visualized in Looker Studio. Cut reporting time from 4 hours/week to near-zero. But the real value wasn't the time saved — it was catching PMax cannibalization and budget spikes the same day instead of discovering them in a weekly report. If you can't see what the algorithm is doing in near-real-time, you can't course-correct. The advertisers winning with PMax aren't the ones who "set it and forget it." They're the ones who treat PMax like any other campaign type: structured intentionally, monitored regularly, and adjusted based on data the algorithm can't see.

The controls exist. Use them.

Want more tactical playbooks like this? I share advanced PPC automation strategies, scripts, and real campaign breakdowns in my newsletter. Sign up here to get the next one.

Frequently Asked Questions

How do I stop Performance Max from spending on branded searches?
Add a brand exclusion list to your PMax campaign. Go to Account Settings → Brand Exclusions, create a list with your brand name and common variations, then apply it to every PMax campaign. Keep branded search in a dedicated exact-match Search campaign with its own budget. Monitor your branded Search campaign's impression share — if it drops after launching PMax, your exclusion isn't comprehensive enough.
What is the ideal number of asset groups for a PMax campaign?
It depends on your catalog, but the rule is one asset group per distinct product category or intent theme — not per SKU. Each asset group needs roughly 30+ conversions per month to maintain stable tCPA/tROAS optimization. If splitting into more groups would push any group below that threshold, keep groups broader. For most e-commerce brands, 2-5 asset groups per campaign is the practical range.
Do audience signals actually control who sees my PMax ads?
No — audience signals are seeds, not constraints. The algorithm uses them to accelerate learning in the first 2-3 weeks, then explores beyond them. If you need hard audience restrictions, PMax is the wrong campaign type. Use audience signals to point the algorithm in the right direction at launch, then let it expand as it gathers conversion data.
How many conversions does PMax need to exit the learning phase?
Google hasn't published an exact number recently, but the practical threshold across most accounts is roughly 30 conversions per month per asset group for tCPA or tROAS strategies to stabilize. Below this, expect frequent re-learning phases and volatile CPA. If you're below 30 conversions, consider switching to Maximize Conversions without a target CPA until volume increases.
Can I set a hard budget cap on Performance Max campaigns?
PMax uses daily budgets like other Google Ads campaigns, but the system can spend up to 2x your daily budget on any single day (averaging out over the month). There's no true hard cap within a day. For tighter control, use campaign-level budgets rather than portfolio bidding, and monitor spend daily during the first two weeks after launch or any major restructuring.
Why is my PMax campaign cannibalizing my Search campaigns?
PMax runs across all Google inventory including Search, and it will bid on queries your Search campaigns already target — especially branded terms. The fix: add a brand exclusion list to PMax, keep branded search in a dedicated campaign, and never put PMax and branded Search in the same portfolio bidding strategy. Check your branded Search impression share as a leading indicator of cannibalization.

Google Performance Max: How to Actually Control the AI Black Box (Asset Groups, Audience Signals, Budget Caps)