Microsoft Ads optimization experiments are generally available. Which one should you run?

Microsoft Advertising's September 2026 product update, published 30 September, quietly closed a gap that has sat open since Google Ads introduced campaign-level experiments years ago: Microsoft Ads now lets you test a change against a live control before committing to it, across Search, Shopping, Audience and Performance Max, not just Search. The feature existing is not the hard part. The hard part is that Microsoft Advertising now offers three different experiment types that answer three different questions, and picking the wrong one wastes a month of split traffic on an answer you didn't need.

What actually went generally available on 30 September

Before this update, Microsoft's optimization experiments worked on Search campaigns only: you could test a bid strategy, a keyword set or an ad group structure against a treatment version of the same campaign. The September update extended that same mechanism to Shopping, Audience and Performance Max campaigns. The platform creates a treatment campaign cloned from an eligible control, you edit whatever you want to test in the treatment, and Microsoft Advertising splits budget and traffic between the two for the duration of the test.

Two details matter more than the headline. First, the split does not have to be 50/50 — the results page scales treatment metrics so an 80/20 or 70/30 split still produces a comparable read, which is useful when you don't want to risk half your budget on an untested idea. Second, because the original Search-only version of this feature already existed, Microsoft renamed it Search optimization experiments specifically to separate it from the new Performance Max testing, which is where most of the confusion in early coverage of this update comes from.

Three experiment types, three different questions

"Optimization experiments" is the umbrella name for a Search, Shopping or Audience test. Performance Max has its own pair of experiment types that were introduced separately, earlier in 2026, and answer questions a Search-style test cannot.

The distinction is the first thing to get right, because the three types are not interchangeable. Running an Uplift experiment when what you actually want to know is whether a new bid strategy helps your existing Search campaign will answer a question you didn't ask, and tell you nothing about the one you did.

Which one applies to your situation

Match your actual decision to the row below before opening Campaigns > Experiments. This is the step the official rollout coverage skips, because it reports the feature rather than the decision in front of you.

Choosing a confidence level without guessing

The 80–95% confidence selector is where most accounts will make an avoidable mistake, because the natural instinct is to pick the highest number available. That's wrong whenever the test needs an answer faster than 95% confidence can deliver one on your traffic.

The trade is always speed against certainty, and the right side of that trade depends on how expensive being wrong is, not on how important the test feels. A Search optimization experiment on ad copy is cheap to reverse, so 80–85% is often enough. An Upgrade experiment that ends with you deleting a Search campaign is not reversible in the same way, which is the case for paying the extra weeks 95% confidence costs.

Check your volume before you start

Microsoft's own guidance for Performance Max Uplift and Upgrade experiments suggests at least 30 conversions in the trailing 30 days before you begin. Below that, a 4-to-8-week test frequently runs its full duration without crossing the confidence threshold you picked, which means the budget split cost you a month and produced no readable answer either way. The same logic applies, with lower stakes, to Search optimization experiments: a campaign converting a handful of times a week will need the long end of the duration range and the lowest confidence level you're willing to accept, or it should wait until it has more volume to split.

This is the same sample-size problem that shows up in every form of split testing, from website A/B tests to Google Ads budgets too small to generate a usable conversion rate — the test design is sound, but there isn't enough traffic flowing through it to reach a real answer in a reasonable time.

Setting one up

1. Open Campaigns, then Experiments in the left navigation.
2. Choose an eligible control campaign. Confirm it does not use a shared or lifetime budget and is not Dynamic Search Ads.
3. Create the treatment and make the one change you're testing — a bid strategy, a new PMax campaign, or a migration build, depending on which of the three types applies.
4. Set your budget split. It does not need to be 50/50; the results page scales the comparison for you.
5. Pick a confidence level and duration using the trade-off above, based on how reversible the change is and how much traffic the campaign gets.
6. Let it run the full term and check the side-by-side results page rather than ending it early on a lead.
7. Apply the winner to the original campaign, or use it to build a new one, once the result crosses your chosen confidence level.

Frequently asked questions

What changed in Microsoft Advertising on 30 September 2026?

Microsoft Advertising's September 2026 monthly product update made optimization experiments generally available for Search, Shopping, Audience and Performance Max campaigns. The feature previously covered Search campaigns only. It creates a treatment campaign from an eligible control, splits budget and traffic between them, and gives you a unified side-by-side results page to compare the two and apply a winner.

What is the difference between a Search optimization experiment and a Performance Max experiment in Microsoft Advertising?

A Search optimization experiment tests a change within an existing Search, Shopping or Audience campaign, such as a bid strategy or keyword change, against a treatment version of that same campaign. The two Performance Max experiment types ask a different question: an Uplift experiment tests whether adding a new PMax campaign creates incremental conversions on top of what your existing campaigns already deliver, and an Upgrade experiment tests whether migrating an existing Search or Shopping campaign to Performance Max would outperform keeping it as is.

What confidence level should I choose for a Microsoft Ads experiment?

Microsoft Advertising lets you choose 80%, 85%, 90% or 95% confidence, with a 4 to 8 week duration. Lower confidence levels reach a readable result faster on less traffic but accept a higher chance of calling a false winner; 95% needs more conversions and a longer run but is the level to use before a change you cannot easily reverse, such as migrating a Search campaign to Performance Max.

Which Microsoft Ads campaigns are not eligible for optimization experiments?

Campaigns using shared or lifetime budgets are not eligible, and neither are Dynamic Search Ads campaigns. A campaign has to be running on its own standard daily budget before you can create a treatment version of it.

How many conversions do I need before running a Performance Max Uplift or Upgrade experiment?

Microsoft's own guidance for Performance Max Uplift and Upgrade experiments suggests at least 30 conversions in the trailing 30 days before you start. Below that volume, the test usually runs the full 4 to 8 weeks without reaching a readable result at any confidence level, which wastes the budget split rather than answering the question.

The takeaway

Microsoft Advertising went from one testing tool to three in the space of a few months, and the rollout coverage has mostly reported the feature list rather than which type fits which decision. Match the question you're actually asking — tune a campaign you keep, add PMax on top, or replace a campaign with PMax — to the right experiment type, check you have the conversion volume to finish within 4 to 8 weeks, and only reach for 95% confidence when the change is one you can't casually undo. Get those three choices right and the new tooling does what Google Ads experiments have done for years: replace a guess with a number.