Performance Max done right (and when to abandon it)

The principle: signal density beats granularity

PMax learns from the conversion data inside each asset group. The more conversions an asset group has per week, the faster the model converges and the better it bids. Split the group too early and every sub-group starves. Split it too late and the model averages incompatible buyer patterns and bids poorly.

The right rule: each asset group needs 30+ conversions per week to feed a healthy bid model. If you're considering splitting and either side would drop below 30, don't split. Wait for total volume to rise first.

When to split: the four legitimate reasons

If your reason for splitting isn't on this table, don't split.

Split execution checklist

1. 1

   Verify volume

   Each side of the split must average 30+ weekly conversions for the past 4 weeks. If not, wait.

2. 2

   Mirror creative pre-split

   Build the new asset group structure with the existing creatives mirrored across before changing budget. Switch traffic over once both groups have 14 days of learning.

3. 3

   Distinct audience signals per group

   Each asset group gets its own custom segment seeded against the relevant buyer pattern. Same audience signal across groups defeats the purpose of splitting.

4. 4

   Seven-day learning window

   Don't touch budgets for the first seven days. The bidder is relearning. Touching budgets resets the learning phase.

Frameworks behind this chapter

PMax fails for one of four reasons

When a PMax campaign underperforms, the bidder usually gets blamed first. Almost always wrongly. The bidder is consuming the inputs you give it. If those inputs are broken, the bidder bids poorly.

Four leaks account for the vast majority of PMax failures: branded-search bleed, catch-all search themes, weak audience signals, and no Search backstop. Run the diagnosis below before changing the bid strategy.

Leak 1: branded-search bleed

1. 1

   Symptom

   Dashboard ROAS is 30-50% higher than blended (Shopify) ROAS. PMax's reported conversions include branded queries that would have converted anyway.

2. 2

   Diagnosis

   Pull the PMax search-terms report. Filter to brand-name terms. Sum their conversion value. If it's more than 15% of total PMax conversions, you have brand bleed.

3. 3

   Fix

   Carve branded search into a dedicated Search campaign with a tight POAS target. Add brand terms to the account-level negative list. Watch dashboard ROAS drop 25-40% in week 1; blended ROAS holds steady.

Leak 2: catch-all search themes

1. 1

   Symptom

   Search themes set to broad concepts like 'skincare' or 'supplements'. The bidder picks any query loosely related and the spend dilutes across the entire vertical.

2. 2

   Diagnosis

   Open the campaign, view search themes. If any theme is a single broad word (skincare, fitness, jewellery), it's a catch-all.

3. 3

   Fix

   Replace each catch-all with 3-5 specific themes (e.g. 'retinol serum for dry skin', 'vitamin C with hyaluronic acid'). Specificity teaches the model what the asset group is actually for.

Leak 3: weak audience signals

1. 1

   Symptom

   Audience signal is a generic affinity (e.g. 'beauty enthusiasts'). The model treats it as low-signal and falls back to broad targeting that takes longer to learn.

2. 2

   Diagnosis

   Check the audience signal panel. Custom segments built from your CRM or in-market intent feeds outperform generic affinities by 3-5x.

3. 3

   Fix

   Build a custom segment per asset group: search-history terms specific to the routine, plus a CRM-data segment of recent buyers in that routine. Combined signal teaches the model fast.

Leak 4: no Search backstop

1. 1

   Symptom

   PMax is the only campaign in the account. High-intent search queries (your category + commercial intent) get whatever asset group PMax decides, often the wrong one.

2. 2

   Diagnosis

   Look at your campaign list. If you only have PMax campaigns, you're missing the backstop.

3. 3

   Fix

   Build a thin Search campaign with your top 20-30 commercial-intent keywords (category + 'buy', 'best', 'reviews'). Tight POAS target. PMax handles cold-funnel and demand creation; Search handles bottom-of-funnel intent that converts at a higher rate.

Templates for the rebuild

AI Max is an extension, not a replacement

AI Max for Search uses the same Smart Bidding model as PMax to expand keyword matching beyond explicit keywords. It's powerful when the underlying account is structurally healthy. It accelerates a bad account toward the cliff faster.

The mistake we see: brands turn on AI Max hoping it'll fix poor performance. It doesn't. It scales whatever signal the account already has. If the signal is broken (any of the four PMax leaks), AI Max amplifies the brokenness.

AI Max deployment decision matrix

Pre-deployment readiness checklist

• Branded search lives in a dedicated Search campaign, not PMax
• Brand terms are on the account-level negative-keyword list
• Asset groups split by intent, routine, or margin (not by category alone)
• Each asset group has 3-5 specific search themes
• Each asset group has a custom audience signal (not just a generic affinity)
• Your top 20 commercial-intent search terms run in a backstop Search campaign
• Conversion value reflects margin or LTV, not gross revenue
• Last 14 days of dashboard ROAS within 25% of blended (Shopify) ROAS

Reference reading