Google Ads adds A/B testing to the Performance Max asset editor

Google Ads lets advertisers turn changes to a Performance Max asset group into an A/B test directly from the creative editor. Instead of immediately replacing the existing configuration, an advertiser can compare the original version with a modified alternative within the same group and campaign. Google documents the option as a beta feature, adding an experiment entry point to the interface where advertising materials are prepared.

The change concerns the test creation workflow, rather than the first introduction of Performance Max experiments. The separate Experiments section continues to support comparisons between asset sets. With the editor integration, changes to headlines, descriptions, images or other assets can form the treatment, while the previous configuration remains the reference for comparison.

The asset comparison belongs to optimisation experiments. Google distinguishes it from tests measuring the incremental contribution of adding a Performance Max campaign alongside other campaign types, and from experiments simulating a budget shift from Shopping or other campaigns into Performance Max. In the new editor workflow, the question being tested concerns the selection of advertising assets within the chosen group, rather than moving the entire strategy to another campaign type.

Asset changes can be saved as an experiment

On the asset editing page, Google describes a dedicated panel for saving and testing changes. Enabling the A/B test option changes the main action from a normal save to experiment creation. Before confirmation, the interface shows the control arm, treatment arm, experiment name, treatment traffic allocation, duration and success metrics that will be used to evaluate the result.

One setting directly affects the final configuration: automatically applying the changes if the treatment wins is enabled by default. According to Google, the changes take effect after the experiment ends and only if the results show that the treatment performed better than the control. Creating the test therefore includes a decision about what happens at the end, as well as how traffic is distributed during the comparison.

The advertiser can review that option alongside the prefilled settings before creating the experiment. The documented behaviour does not mean that every test automatically changes the campaign: application depends on the selected setting and the experiment's result. That distinction matters in accounts where approving creative materials and changing the production configuration are separate responsibilities.

Three asset categories within one group

The experiment compares two asset selections, but it does not necessarily divide the entire group into two completely separate sets. Version A contains existing assets chosen as the reference. Version B can use other existing materials or new files uploaded for the comparison. Assets that are not selected for either arm remain common and continue to be eligible across all the campaign's traffic.

Under this model, common assets can serve alongside the materials specific to A or B, depending on the traffic segment. The experiment is therefore interpreted through the differences between the tested selections, rather than as a comparison between two independent campaigns. Google limits each such experiment to a single asset group, keeping the subject of the comparison within the chosen configuration.

The traffic allocation and the budget split are also different concepts. Google's explanations of experiments state that allocating traffic does not promise identical spending between the arms. Cost differences have to be read alongside the metrics and the test outcome, rather than automatically equating an equal division of traffic with two equal amounts spent.

Duration, asset approval and editing restrictions

For setup through the Experiments section, the documentation sets the start date to the following day. The Experiment Guidance System estimates the end date from the selected campaign and the duration needed to obtain statistically significant results. Google recommends running the experiment for at least four to six weeks, while allowing the end date to be changed during configuration.

Once the experiment starts, the tested group becomes view-only: its assets cannot be edited, added or removed until the comparison ends. The restriction concerns the group included in the experiment, rather than every asset in the account. At the same time, new files in the treatment undergo the usual Google Ads policy review; disapproved materials cannot serve in the test.

Assets from both arms count towards the group's limits. The test configuration therefore has to respect the limits applicable to the assets concerned, even when some are reserved for the alternative version. The experiment does not create a separate environment in which campaign approval rules or asset limits are suspended.

Applying a result remains a separate action

Google describes applying and ending an experiment as separate commands. Applying prompts the advertiser to confirm an update to the original group using the tested changes. Ending without applying stops the comparison without changing the original configuration; new treatment assets are discarded from the experiment and the group returns to its original state. That workflow coexists with the option to apply a winning treatment automatically, selected when creating the test from the editor.

There are also conditions that can prevent an experiment from starting. The documentation lists removed campaigns, dates overlapping other experiments, certain budget or bidding configurations and unsupported features. Those situations are distinct from an advertising asset being disapproved: the first category concerns the configuration's eligibility for an experiment, while the second concerns whether a specific asset can serve.

For Google Ads campaign management, the change brings testing closer to the routine process of editing creative materials. The essential comparison settings remain explicit: which materials change, which assets are common, how traffic is divided and whether a winning treatment will be applied automatically. The documented feature remains in beta, with a workflow that separates preparing a change, running the experiment and updating the original asset group.