Decision brief · SEM
AI Max changes the experiment from a copied campaign to a divided budget
The new experiment keeps control and AI Max inside one Search campaign. That removes a source of configuration drift, but it does not remove the shared auction clock or the eligibility limits around the test.
- The split moved inside the campaignTraffic and budget are diverted between control and AI Max conditions without creating a trial campaign copy.
- Cleaner configuration is not a frozen marketBoth conditions still experience the same calendar, auctions and demand shocks, so timing remains part of interpretation.
- Eligibility defines where the result can travelShared budgets, portfolio bidding, bidding exploration and active experiments exclude a campaign from this test format.
Most Google Ads experiment guidance begins with a copy. A custom experiment duplicates a campaign into control and trial branches, then asks whether the changed branch performs differently. AI Max experiments alter that architecture: Google diverts traffic and budget within the existing Search campaign instead of creating a second campaign to maintain.
That sounds like an implementation detail. It is actually a change in the unit of comparison. The operator gains a cleaner match between settings, history and budget ownership, while accepting a stricter eligibility boundary and a result that still unfolds inside one moving auction market.
The unit of comparison is one campaign, not two copies
Experimental unit
One budget owner, two conditions
The same Search campaign carries both conditions. The experimental question becomes whether internally allocated AI Max traffic changes outcomes, not whether a separately maintained campaign copy can reproduce the control.
In a copied custom experiment, unnoticed differences can accumulate between the base campaign and its trial: a budget edit, a targeting change, a new asset or a maintenance action may land on one branch first. The within-campaign design reduces that configuration-drift surface. It does not prove that every observed difference was caused by AI Max, but it removes one avoidable source of mismatch.
Cleaner configuration does not freeze the auction clock
Both conditions run through the same period, which is useful: a holiday, competitor launch or demand shock is not assigned to a different calendar merely because a trial campaign began later. But simultaneity is not immunity. Auction composition, conversion lag and learning can still change while the experiment is active.
Configuration comparability
One campaign lowers the chance that settings, ownership or routine maintenance diverge across control and trial copies.
Market and measurement movement
Auctions, demand, attribution delay and conversion maturity still move while both conditions collect evidence.
Read the result with complete date ranges and mature conversion windows. Google also recommends allowing at least two weeks after AI Max is enabled before making changes such as adding negative keywords. That waiting period is not proof of statistical sufficiency; it is a practical warning against interpreting an adapting system too early.
Eligibility exclusions define the experiment’s external validity
The cleanest AI Max test excludes campaigns with shared control surfaces
Google does not allow this experiment format when another system can reallocate budget, coordinate bids or run a competing experiment. Those exclusions are not administrative trivia; they describe the population represented by the result.
A positive result in an individually budgeted campaign does not automatically answer what will happen after the campaign returns to a shared budget or portfolio strategy. The observed lift belongs first to the tested control system. If the production setup uses a wider resource pool, treat rollout as a second operational change rather than a clerical switch.
The result should be read through search journeys, not one summary number
AI Max changes how queries, assets and landing pages can participate. Google’s reporting surfaces therefore matter to the interpretation: search terms, keywords, assets and landing pages show where the treatment actually expressed itself, while the search-term and landing-page journey view connects a query to its matched destination.
Locate the mechanism
Identify which queries, match sources, assets and landing pages created the difference. A favorable aggregate can conceal a narrow mechanism that will not scale.
Preserve the comparison window
Record the experiment period and rollout date. Enabling AI Max changes the campaign state, so the clean experimental contrast does not continue indefinitely.
Apply only after the mechanism is visible, conversion data is mature and the production budget and bidding system match the conditions that produced the evidence.
Questions about AI Max experiments
Does an AI Max experiment create a copy of the campaign?
No. Google says the experiment diverts traffic and budget within the existing campaign. A custom experiment instead creates separate control and trial campaign configurations.
Can a campaign with a shared budget run this experiment?
No. Shared budgets are among the documented exclusions, together with portfolio bidding, bidding exploration, Display Network targeting and active experiments.
Can Google apply AI Max automatically after the test?
The experiment can be configured to apply AI Max automatically when results are favorable. It can also be applied manually after the experiment or enabled later in campaign settings.
