A realistic look at Google's Campaign Mix Experiments

A realistic look at Google's Campaign Mix Experiments

For as long as Google Ads has had multiple campaign types, advertisers have been guessing at one of the hardest questions in paid media: What is the real impact of adding a certain feature or setting to the whole account? Is PMax cannibalising Search, or is it adding net new conversions? Is my granular, super focused search structure better than broad, consolidated structures?

Mix Experiments are Google's first real attempt to let us stop guessing.

What Mix Experiments actually do

Available in beta as of early 2026, Mix Experiments let you build up to five "arms," each containing a portfolio of campaigns, and split traffic between them. Crucially, the same campaign can appear in multiple arms. So if you want to test the incremental value of adding Demand Gen on top of your existing Search + PMax setup, Arm A holds Search + PMax, Arm B holds Search + PMax + Demand Gen, and the only variable changing is the Demand Gen layer.

This is a big thing. Previously, Google Ads experiments were locked inside a single campaign. You could test a bid strategy, a creative, a keyword match type. You could not test whether your entire account architecture was wrong. Now you can. The right question isn't "what's the ROAS of this campaign?" It's "what does the portfolio look like, and what happens to the whole thing when I change the mix?"

The traffic split is cookie-based, which means once a user is assigned to an arm, they stay there for the duration. That matters because consumer journeys are messy: someone might see a YouTube ad on Monday, click a Demand Gen ad Wednesday, and convert via branded Search Friday. Cookie locking keeps the synergy effects clean…given that the cookie stays, of course.

Let's be honest about the 2026 privacy landscape: cookies often don't stay. If a user sees your YouTube campaign on Day 1, but doesn't perform that branded search until Day 20, their original cookie is likely wiped. They will enter the system as a brand new user, potentially get sorted into the opposing arm, and your experiment will completely miss the upper-funnel assist.

Because of this tracking fragility, the tool works best for direct response channels and tweaks. When testing variations of PMax against Standard Shopping, or evaluating bottom-of-funnel Search structures, the time-to-convert is typically much shorter. This fits safely within that shrinking cookie window.

Where many teams will get this wrong

The tool is powerful. The traps are also real. A few things to watch for:

Confidence intervals are not a tone setting. Google's interface lets you read results at 95%, 80%, or 70% confidence. A lot of the early commentary frames this as "pick your risk tolerance." Fine in theory, but 80% confidence means you're wrong one time in five, and 70% means three times in ten. For a directional creative test, that's acceptable. For a decision to shift 30% of your Search budget into Demand Gen across an account, it's not. Default to 95% for anything that moves real money.

Mix Experiments measure attribution, not incrementality. This is the most important caveat and the one most likely to be missed. The experiment tells you which arm performed better according to Google's conversion tracking. It does not tell you whether either arm actually drove conversions that wouldn't have happened anyway. To answer that question, you still need a Conversion Lift study or a geo lift.

The minimum runtime is longer than people think. Six to eight weeks is the working number. Algorithms need time to stabilize on a new traffic partition, conversion lag has to clear, and you need enough volume to actually reach the confidence interval you've selected. Calling the test at week three because the treatment arm "looks promising" - don’t do it.

Both campaigns have to already exist. You cannot create a brand new Demand Gen campaign inside the experiment. Both arms must be built from live campaigns, which means if you want to test "add Demand Gen," you have to launch the Demand Gen campaign first, let it warm up, then duplicate the setup into the experiment. Worth knowing before you scope a test plan with a client.

How we think about it at NoA Connect

Mix Experiments earn a place in the toolkit, but they don't replace anything we already do. They sit alongside Conversion Lift, geo lift, and ROAS/CPA split tests, each answering a different question. Mix Experiments answer "is this channel split better?" Conversion Lift answers "are these ads incremental at all?" Geo Lift answers "what would happen if we turned this off?" Confusing the three is how you get bad client conversations.

Pick one structural hypothesis, hold everything else constant, run for at least six weeks, evaluate at 95% confidence, and pair the result with whatever incrementality evidence we already have for the channels involved.

The wider point

The agencies that win won't be the ones running the most Mix Experiments. They'll be the ones reading them properly: picking the right hypothesis, holding their nerve through the learning phase, knowing what the data can and can't say. Google has handed us a much better question-asking tool, but the answers are still ours to interpret.