To test copy in Performance Max, change the headlines and descriptions and keep everything else identical: the same images, videos, audience signals and search themes. Then any difference in results comes from the words.
The usual way to do it is to duplicate the asset group and edit the text in each copy. That gives you a comparison, but not an even split, because Google decides which asset group serves. When you need a true split, Google's A/B testing assets experiment (in beta) divides traffic between two sets of assets inside one asset group.
The test from the call
On a mentoring call, a freelancer running PMax for an online art store showed us an asset group whose headlines described a community of artists. Nothing in them said buy. We wrote three new sets of copy: one built on best sellers and collections, one on the first order offer, one on trust, quality and service. Then the plan:
- Make three copies of the asset group, named A, B and C.
- Keep the same 20 images and the same videos in all three, since no new creative had arrived.
- Keep the same audience signal and search themes. Only the headlines and descriptions change.
- Pause the original once the three are live, and review after about a week.
On the follow-up call about two weeks later, asset group A had spent $250 with no conversions. Asset group C had spent $160 and brought 2 conversions at $79 each. C was the trust and service copy, with reviews, order tracking, live chat and the discount. We paused A and B and kept C, and the next test was images: same text, different pictures.
Two conversions is a direction, not proof. The run also had noise: the bid strategy had been switched on the day the test started, and a payment problem stopped the ads for about a day and a half. Those hit all three groups, but they're the kind of change we'd rather keep out of a test window.
Why duplicates don't split evenly
Google prioritizes asset groups with search themes the same way it prioritizes ad groups with keywords. If several are equally relevant to a search, the one that generates the ad with the highest Ad Rank enters the auction. Asset groups without search themes are picked by Ad Rank as well. Identical search themes in three copies means they compete, and one can take most of the spend early.
So read rates, not totals. Compare cost per conversion and conversion rate per asset group, and expect uneven spend.
The clean version: A/B testing assets
In Experiments, choose Assets, then Assets provided by you, then Performance Max. You pick one asset group, keep its current assets as the control arm and select or upload the treatment assets. Anything you don't assign to either arm keeps serving to all traffic. You set the traffic split. You can also start one from the asset group editor with the Save asset group changes as an A/B test toggle.
- Give it time. Google recommends at least 4 to 6 weeks.
- Hands off. Assets in the tested group are locked until the experiment ends, and only one asset group can be tested per experiment.
- Check eligibility. Campaigns with a shared budget or a portfolio bid strategy can't be used as is.
- Watch generated text. Text customization is on by default and can add headlines and descriptions from your landing page. They show as Automatically created in asset reporting.
The takeaways
- Change only headlines and descriptions. Freeze images, videos, audience signals and search themes.
- Duplicated asset groups compete by relevance and Ad Rank, so spend won't split evenly.
- Compare cost per conversion, not total conversions, and treat a handful of conversions as a direction.
- For a true split, use the A/B testing assets beta and let it run 4 to 6 weeks.
- Don't change bidding or budget during the test window.
Common questions
How do I A/B test ad copy in Performance Max?
Keep everything except the text identical: images, videos, audience signals and search themes. You can do that with duplicated asset groups and compare results, or, for a true split, use Google's A/B testing assets experiment (beta), which serves a control set and a treatment set of assets to separate shares of traffic inside one asset group.
Do duplicated asset groups in the same campaign get equal traffic?
No. Google uses the same prioritization for asset groups with search themes as for ad groups with keywords: when several are equally relevant to a search, the one generating the ad with the highest Ad Rank enters the auction. Asset groups without search themes are chosen by Ad Rank too, so spend can land very unevenly.
How long should a Performance Max asset test run?
For the A/B testing assets experiment, Google recommends a minimum of 4 to 6 weeks, and its Experiment Guidance System suggests an end date. For assets in general, Google suggests waiting about 2 to 3 weeks before checking Ad strength and replacing low performing assets.
Can I edit assets while a Performance Max A/B test is running?
No. Google locks the asset group being tested, so you can't edit, add or remove assets in it until the experiment ends. You can only test assets within a single asset group per experiment.
Can Google generated headlines affect a PMax copy test?
Yes. Text customization is on by default in Performance Max and can add headlines and descriptions based on your landing page when Google predicts they'll do better. They appear in asset reporting with Automatically created in the Source column, so check them when you read the result.