Performance Max creative just got a real experiment spine.
On October 2, 2026, Search Engine Land reported that Google Ads is rolling out A/B testing for Performance Max asset groups, so advertisers can test headlines, descriptions, images, and videos without immediately replacing the live set (Anu Adegbola, Search Engine Land, October 2, 2026). Google's own Accelerate announcement frames the same rollout as new asset experiments in Performance Max: measure a different asset set, the impact of adding assets, or seasonal creative against evergreen, including assets from Asset studio (Google Business / Accelerate, accessed October 4, 2026). Help Center pages for A/B testing assets (Beta) and broader asset testing spell out the mechanics: control versus treatment inside one asset group, a traffic split you define, and a hard lock on edits while the test runs (Google Ads Help, answers 16807329 and 16030588, accessed October 4, 2026).
If your Q4 creative plan is still "swap the hero image and hope the blended ROAS moves," you are buying last year's workflow. The scarce seat this week is one locked experiment with a clean counterfactual, not another gut refresh that contaminates the learning period right before Black Friday.
What actually shipped
Keep the claim exact.
Google's A/B testing assets beta lets you compare two asset sets inside the same Performance Max asset group. Assets fall into three buckets: control (A), treatment (B), and common assets you leave out of either arm. Common assets keep serving to 100% of traffic alongside the split, so the test does not strip the campaign bare (Google Ads Help, A/B testing assets Beta, answer 16807329). You set the traffic split between control and treatment. The start date defaults to the next day. The Experiment Guidance System proposes an end date for significance. Google recommends running at least 4 to 6 weeks (Google Ads Help, answer 16807329).
The lock is the product, not a footnote. Once the experiment starts, you cannot edit, add, or remove assets in that asset group until it ends. Asset editing pages go view-only so mid-test swaps cannot poison the read (Google Ads Help, answer 16807329; Search Engine Land, October 2, 2026). You can only test assets within a single asset group per experiment. Control and treatment assets both count toward asset group limits. Newly uploaded treatment assets still pass standard policy review; disapproved assets will not serve (Google Ads Help, answer 16807329).
When the test finishes, Apply experiment updates the original asset group with treatment assets. You can add the B set and, by default, keep the A set as well, or uncheck keep-control to remove A. End experiment stops the test with no changes; treatment uploads are discarded and the group reverts to its original state (Google Ads Help, answer 16807329).
A parallel Help article covers asset testing subtypes that answer different questions: adding text, image, and video assets to a product-feed-only campaign, or measuring incremental lift from video by suppressing video (including auto-generated) in the control arm while text and images continue for both sides (Google Ads Help, Asset testing, answer 16030588). Those tests also run as a traffic split inside one campaign, which Google says can shorten learning versus duplicating whole campaigns (Google Ads Help, answer 16030588). After you create the experiment, do not change text customization, Final URL expansion, or video enhancement settings, and do not change assets once the test has started (Google Ads Help, answer 16030588).
Google's Accelerate post adds the roadmap around the core: Asset studio creatives are fair game in tests; a second success metric for experiments is coming soon; conversion lift studies and experiments sit on one Experiments page; expanded Google Ads API and manager account (MCC) support starts rolling out in a few weeks; the features themselves are currently rolling out (Google Business / Accelerate, accessed October 4, 2026). Treat availability as rolling, not universal on day one. PostMe Live's October 3 operator guide and Common Thread Collective's October 2 Coach's Corner post both describe the same control/treatment/lock/4-to-6-week shape advertisers are seeing as accounts light up (PostMe Live, October 3, 2026; Common Thread Collective, October 2, 2026).
Why before-and-after was never enough
PMax has always assembled ads from your asset pool plus Google's automation. Asset reporting and strength labels helped, but they were not a controlled creative counterfactual. The common workaround was duplicate campaigns or duplicate asset groups. That changes inventory access, auction dynamics, budget pacing, and bidding behavior. You were not isolating creative. You were comparing two algorithmic systems (Search Engine Land, October 2, 2026; PostMe Live, October 3, 2026).
In-asset-group A/B testing keeps the campaign architecture, budget, and bidding frame fixed. The variable is the creative set under test. That is the difference between a marketing learning loop and a story you tell after the fact.
Agency blogs will push urgency and category share numbers. Common Thread Collective writes that PMax "now drives 45% of all Google Ads conversions" and that campaigns with video see "25 to 40% performance improvement" (Common Thread Collective, October 2, 2026). Those figures are CTC's claims, not statements we verified in Google Help or the Accelerate announcement for this piece. Use them as agency framing if you want. Do not brief leadership as if Google published them in the experiment docs. The load-bearing product claim does not need those percentages: Google finally gave you a control arm inside the asset group.
What to test first (and what not to)
Google and the trade coverage point at concrete experiment shapes: new creative concepts versus existing assets; UGC-style versus polished brand; different messaging or image styles; adding or changing video; seasonal versus evergreen; Asset studio outputs versus production assets; feed-only plus incremental creative; video on versus video suppressed (Search Engine Land, October 2, 2026; Google Ads Help, answers 16807329 and 16030588; Google Business / Accelerate).
The operator rule is one variable per experiment. If you swap images, headlines, and video in the same treatment arm, you will get a blended winner and no causal story. Pick the question that changes a Q4 budget or creative production decision:
- Do we have enough video lift to justify production (or Asset studio) spend on this asset group?
- Does lifestyle or UGC beat flat product for this assortment?
- Does seasonal BFCM creative beat evergreen on the metric we actually buy to, or only on vanity CTR?
- Does a tighter headline system beat the current kitchen-sink text set?
Video incremental tests are structurally clean when you already have video in the group: control suppresses video, treatment keeps it, other assets serve both sides (Google Ads Help, answer 16030588). Feed-only retail tests answer whether incremental text/image/video assets beat feed-only serving (Google Ads Help, answer 16030588). Full A/B asset sets answer concept battles inside an already-rich group (Google Ads Help, answer 16807329).
What not to do: launch five experiments across five groups with five half-finished treatments; start a test you know you will break with a mid-November creative dump; treat Asset studio AI video as "proven" without a control; assume a win in one asset group generalizes to every other group without a local re-read (PostMe Live, October 3, 2026).
The lock is the staffing problem
The view-only lock is where media, creative, and merchandising collide.
If your team refreshes PMax assets weekly, a 4 to 6 week lock means you schedule creative before Schedule, not after. Search Engine Land's catch line matches Help: once the experiment begins, you cannot make additional changes to the asset group while the test runs (Search Engine Land, October 2, 2026; Google Ads Help, answer 16807329). That is intentional contamination control. It is also a calendar problem for Q4.
Translation for this week: pick the asset group that matters most for holiday volume or margin, freeze a treatment set you can live with for a month-plus, brief creative that the next refresh waits until Apply or End, and put a named owner on "no emergency swaps" unless you are willing to kill the experiment.
Shared budgets, portfolio bidding, Smart Bidding Exploration, overlapping experiment dates, unsupported features, and asset-limit maxes can block creation. Help lists those failure modes and the fixes (remove shared budget, pick non-overlapping dates, trim assets if the group is maxed, choose another campaign) (Google Ads Help, answer 16807329). Agencies waiting on MCC and API should plan UI setup in individual accounts now; Google says that expanded support is weeks away, not day-zero universal (Google Business / Accelerate).
The marketing consequence
This is a media and measurement buy, not a new PMax "AI creative" SKU.
The scarce work is choosing one hypothesis worth locking, defining control and treatment so only one thing moves, accepting a 4 to 6 week quiet period on that group, and deciding in advance what Apply means for brand and merchandising. A second success metric is coming on Google's roadmap (Google Business / Accelerate). Until it ships, write your primary metric and your "do not break" guardrail metric into the brief yourself so a CTR win that tanks revenue does not get auto-celebrated.
PMax will still be automated. Asset experiments will not make the black box transparent. They do give advertisers a cleaner answer to the question Search Engine Land put on the table: did the creative change actually make a difference (Search Engine Land, October 2, 2026)?
What to do this week
Open Experiments under Campaigns. If Assets / Performance Max / assets-provided-by-you is not visible yet, check daily through the rollout window (Google Ads Help, answer 16807329; Google Business / Accelerate; Common Thread Collective, October 2, 2026).
Pick ONE high-stakes asset group tied to Q4 revenue or a decision you will fund differently if lift is real. Write a one-page experiment brief before you click Schedule: hypothesis in one sentence, control assets, treatment assets (or video-on vs suppressed, or feed-only vs add assets), traffic split, primary metric, kill criteria, start date, and the 4 to 6 week end window Google recommends (Google Ads Help, answers 16807329 and 16030588).
Freeze edits. Put the lock on the calendar for creative, media, and merchandising. No mid-test hero swaps. No "quick" headline dump. If you need a seasonal launch inside that window, use a different asset group or accept ending the experiment (Google Ads Help, answer 16807329).
If you lack video on a priority retail or demand group, make the first test an incremental video or feed-plus-assets experiment rather than a vague full creative restyle (Google Ads Help, answer 16030588). If you already have video and a strong concept debate, run A versus B asset sets with common assets held out of the fight (Google Ads Help, answer 16807329).
Decide Apply policy before results land: add treatment and keep control, replace control, or End with no change. Default keep-control is a product setting, not a strategy (Google Ads Help, answer 16807329).
Do not brief leadership with unverified share-of-conversion or video-lift percentages from agency blogs. Brief them with Google's mechanics: same asset group, traffic split, lock, 4 to 6 weeks, Apply or End (Google Ads Help; Google Business / Accelerate; Search Engine Land, October 2, 2026).
Lock the asset group before you trust the lift. Before-and-after swaps inside a moving PMax pool were always a story. Control versus treatment with a frozen group is a test. This week's job for media and creative leads is one experiment worth freezing, not another gut refresh that burns the only clean Q4 learning window you just got.
