What changed
Google started rolling out asset experiments for Performance Max on June 11, 2026. Three test types are available: a head-to-head comparison of two full asset groups, an incrementality test for a single asset (does adding this specific video actually lift conversions?), and a seasonal-vs-evergreen variant test. Results are measured using conversion lift methodology (a holdout-based approach that estimates how many conversions the creative change caused, not just correlated with) rather than last-click attribution. Google's experiments and conversion lift study UIs were merged onto a single page. MCC support is live at launch; API access is coming in the following weeks.
Assets created in Asset Studio, Google's Gemini-powered creative generation tool, are eligible for experiments, which means AI-generated creative can be formally tested against human-made work inside the same framework.
Who it affects
Every PMax advertiser is affected, but the accounts that gain the most immediately are mid-to-large e-commerce and lead-gen operations that already have enough creative volume and conversion throughput to reach statistical significance in a reasonable window. Agencies running multiple PMax accounts stand to gain a systematic testing infrastructure once API access ships. If you're running PMax on a tight creative budget with one or two asset groups, you'll benefit eventually, but you need the volume first.
Why it matters
PMax has always been a trade: hand Google the keys to inventory, bidding, and creative rotation, and it will find conversions you'd miss with manual campaign structures. The cost was opacity. You could see that an asset group performed, but not why, and you couldn't isolate a single creative decision from the algorithm's inventory and bid choices happening simultaneously.
That's not a small problem. When you can't attribute results to a creative decision, you can't build a creative learning agenda. You end up cycling assets on gut instinct and hoping Google's asset ratings ("Best", "Good", "Low") tell you something actionable. They don't, reliably. Asset ratings reflect how often an asset gets served, which is a function of Google's predicted performance, not a controlled measurement of its actual incremental impact.
Conversion lift methodology changes the frame entirely. By holding out a portion of eligible users from seeing the test variant and comparing conversion rates between the exposed and unexposed groups, you get an estimate of what the creative change caused, isolated from the algorithm's other decisions. It's not perfect (holdout-based measurement never is at small scale), but it is a fundamentally different quality of signal than anything PMax has offered advertisers on the creative side.
The timing matters too. Asset Studio is now inside this testing framework, so the question "do AI-generated assets actually perform?" has a rigorous answer available for the first time, at least inside your account. That's a real research question worth running.
The play
-
Run your first full asset group test within 30 days. Set up your current best asset group as the control and build a challenger. The challenger should represent a genuine creative hypothesis, not just different images of the same concept. Watch the conversion lift estimate and its confidence interval. A statistically significant lift of even 8 to 10% justifies shifting creative investment.
-
Pick one video you've been unable to evaluate and run the single-asset incrementality test. The specific metric is the incremental conversion rate lift attributed to that asset. If the confidence interval is wide after three weeks, the account doesn't have enough conversion volume to isolate this signal cleanly; that's useful diagnostic information.
-
Test Asset Studio creative against your human-made control. Set the comparison up as an asset group test. The bar: if AI-generated creative comes within 10% of control CPA, it earns a permanent place in your rotation. If it's more than 20% worse, document that and use it to calibrate how much Asset Studio output you trust in this account without testing first.
-
Build an MCC-level test template now, before API access ships. When the API arrives, you want a repeatable structure: same test type, same success metrics, same minimum conversion threshold for a valid read. Agencies that have this ready will be able to run a systematic creative testing operation at scale. Those that don't will treat each experiment as a one-off.
Watch-outs
Conversion lift tests require volume. If an asset group isn't generating enough conversions to power a holdout comparison, your results will be noisy and the confidence intervals will be too wide to act on. Don't let an inconclusive result convince you an asset is neutral; it may just mean the account is too small for this methodology.
Don't over-test. Running multiple simultaneous experiments in the same PMax campaign creates interaction effects between holdout groups. Run one experiment per campaign at a time until you understand how the traffic splits behave.
The "seasonal vs. evergreen" test type sounds appealing around major sale periods, but be careful. If you run a seasonal experiment during a window where external demand is spiking (Black Friday, a category-specific event), the lift you measure is partly the creative and partly the moment. Don't generalize from a high-demand window to baseline performance.
Finally, asset ratings in the interface haven't gone away. They'll still push you toward swapping out "Low" assets. Remember that a low rating means low serving frequency in Google's model, not low incremental impact. The experiment result is the authoritative answer. The rating is a heuristic.
The WhyItWon angle
Asset experiments solve the attribution problem inside PMax, but they can't tell you what to test next. You still have to form the hypothesis: which creative concept, which hook, which format is actually worth a four-week experiment and the budget it consumes. Run a bad hypothesis and you've spent the test window learning nothing. That's where knowing what's already winning in your account and your category, before you commit spend, does real work. WhyItWon reads your existing ads, your competitors' creative, and customer signals to score the next concept before it runs. Pair that with PMax asset experiments and you're not just testing more rigorously, you're testing the right things.
Google Ads
