AI Max for ecommerce: the configuration most agencies miss

AI Max is the campaign type Google launched in April that most accounts shouldn't have turned on. Not because AI Max is wrong, but because turning it on without the structure work first makes the existing PMax cannibalise the existing Search, the new AI Max cannibalise both, and the dashboard tell you a story the underlying P&L disagrees with.

We've now run AI Max activations across about 40 ecom accounts in the last six weeks. The configuration work that actually compounds is small and specific. Here it is.

Step 1: lock the brand exclusion

Before AI Max sees any spend, build a brand-exclusion list at the campaign level for AI Max and at the account level for any campaigns you don't want eating brand traffic. Every common misspelling of your brand. Every variant. Every product line name that overlaps with brand search.

If you skip this, AI Max will route brand search through itself because it's the campaign with the loosest match-type constraints. Your reported ROAS will look great. Your incremental revenue will be flat. The branded-search bleed pattern repeats inside one campaign instead of across many.

Step 2: feed-level labelling

AI Max routes Shopping inventory based on signals it reads from the feed. The accounts winning with AI Max use custom_label_0 through custom_label_4 to mark which products belong in which AI Max-eligible bucket, typically by margin tier (high/medium/low), seasonality, hero-SKU flag, and new-arrival status.

The accounts losing with AI Max have the same five custom labels empty or stuffed with the same string for every product. AI Max can't see a difference between products and routes spend toward whichever inventory converts fastest, which is usually the cheapest SKU. Margin collapses in week three.

custom_label_0  =  brand isolation         (brand | non-brand)
custom_label_1  =  margin tier              (high | mid | low)
custom_label_2  =  seasonality              (always-on | seasonal | clearance)
custom_label_3  =  performance tier         (winner | mid | new)
custom_label_4  =  reserved                 (test groups, geo splits, supplier promos)

# AI Max asset-group selectors then filter on these labels:
#   - High-margin always-on group     :  margin=high   AND seasonality=always-on
#   - New arrivals (margin-protected) :  margin=high   AND performance=new
#   - Clearance (capped daily budget) :  seasonality=clearance
#   - Volume movers (low-margin       :  margin=low    AND performance=winner
#     winners on tight POAS targets)

Step 3: audience signal seeding

AI Max accepts audience signals the same way PMax does. Customer match lists go in as positive signals. Recent purchasers go in as exclusions on prospecting-focused AI Max campaigns so you stop paying to acquire customers you already have. Custom segments built from competitor URLs and competitor search queries seed the algorithm's targeting model.

This step is the difference between AI Max learning in 30 days and AI Max never learning. Most accounts skip it and complain about CPA in month two.

• Customer match: upload within the last 90 days, at least 1,000 hashed records, refreshed monthly. Stale lists below the threshold won't activate.
• Custom segments: include your own brand site, the top three competitor sites in your category, and the top three review sites. Search-term variants populate these for you.
• Recent purchasers exclusion: load your last 30 days of purchasers as a negative signal on any prospecting-led AI Max group so the algorithm spends on net-new acquisition.
• In-market segment: pick the closest match to your product category. Avoid parent categories; the granularity is what gives the algorithm a starting point.

Step 4: define the success metric before launch

Pick the metric AI Max is being judged against and write it down before the campaign goes live. Three options work.

The third option, blended MER trend during the AI Max window, is the easiest to set up but the least defensible. It's a useful directional read for week-one sanity checks, then graduate to one of the two cleaner metrics for the real decision.

If you don't define the metric, you'll judge AI Max against whatever number happens to look best at the end of the month, which is exactly how Google Ads optimisation goes wrong on a structural level.

What good looks like at day 60

If your numbers don't look like that at day 60, the problem is almost always one of: missed brand exclusions, empty custom labels, no audience signals, or no success metric defined. The fix is going back to step one.