Performance Max Asset Experiments in 2026: A/B Testing Creative Without Guesswork
Performance Max has never given a straight answer to the simplest creative question: which asset actually earns the money. The Low / Good / Best performance labels in the asset report are an observation, not a test — they describe what the algorithm chose to serve, which is circular by design. Performance Max asset experiments close that gap. The campaign splits into two arms, each with its own creative set, traffic is divided by a share you choose, and you get a comparison with a confidence read instead of a hunch.
Here is how the mechanic works, how to launch one, how long to run it, which campaign settings quietly block it, and how to read the result without mistaking noise for a win.
What Performance Max asset experiments are, and why the asset report is not a test
Asset experiments are a native A/B test for creative inside a single Performance Max asset group. The feature started as a retail-only capability, expanded across Performance Max campaign objectives during 2026, and by autumn it was showing up in accounts without vertical restrictions.
The difference from the familiar asset report is the nature of the data. The report tells you how the assets the algorithm decided to serve performed. That is self-confirming logic: an asset got few impressions because the system did not trust it, and the system did not trust it because it had little data. An experiment breaks the loop by splitting the audience and giving both arms comparable conditions.
The performance label answers “what did the algorithm serve?” The experiment answers “what works better?” Those are different questions, and they justify different decisions.
The second difference is statistics. The experiment interface estimates whether you have gathered enough data for the gap between arms to be more than chance. The asset report offers no such check at all.
Control, treatment and common assets
The mechanic sorts the creative in your asset group into three buckets:
| Bucket | What goes in | Who sees it |
|---|---|---|
| Control | Your current assets — what is already running | Only the control share of traffic |
| Treatment | New assets — the hypothesis under test | Only the treatment share of traffic |
| Common | Elements that must appear in both arms: logos, business name, baseline headlines | 100% of traffic |
The common bucket is not a formality. Performance Max assembles ads dynamically, and if the treatment arm is missing a required format, the system simply will not build some placements. You would then be comparing reach, not creative. Everything that should stay identical belongs in common assets, so that exactly one thing differs between arms: the hypothesis.
The traffic split is yours to set:
- 50/50 when the campaign can absorb a weak treatment arm and you want an answer fast. Statistically the most efficient split, because both arms accumulate data at the same rate.
- 80/20 or 70/30 toward control when the campaign feeds the business and a bad treatment arm is expensive. You pay for that caution in time — often a lot of it.
A common mistake is setting 90/10 “to be safe.” At that share the treatment arm in most accounts will not accumulate enough conversions before the quarter ends, and the test closes without a verdict. If the budget cannot support 50/50, narrow the hypothesis rather than shrinking the split.
Where to find it and how to launch in eight steps
- Open Campaigns → Experiments in the left-hand menu.
- Under “All experiments,” click the plus button.
- Choose Performance Max as the campaign type and Assets as the test variable.
- Choose “Assets provided by you” as the variable type — as opposed to testing automatically generated assets.
- Select the campaign and one asset group. Multiple groups in a single experiment are not supported; run them sequentially.
- Sort your assets into control, treatment and common.
- Set the traffic split and dates. The system estimates how long it needs for a statistically meaningful result.
- Review the preview for both arms, launch, and leave it alone.
That last point is enforced, not advisory: while the test runs, assets in the participating group are locked for editing. It is a guardrail against the most popular way of destroying an experiment — “just tweaking a headline” in week two.
What is worth testing, and what is a waste of six weeks
Formally you can test images, text and video. Practically, the list of useful hypotheses is shorter than people expect.
Worth testing:
- A change of visual code. Product on white versus product in use. The gap is usually large and shows up quickly.
- A change of offer in the headlines. “Two-day delivery” versus “20% off your first order.” You are testing a motive, not a turn of phrase.
- Video versus no video, or your own video versus an auto-generated one. One of the few tests where Performance Max almost always shows a difference, because video unlocks different inventory.
- People in frame versus pure product photography. Still produces multiples of difference in CTR on feed placements.
Not worth testing:
- Reordering words in one headline. The effect is smaller than the noise.
- Swapping one image out of twenty. Both arms will assemble almost identically.
- “New creative versus old” where the new set has five assets and the old has twenty. You are comparing library size, not ideas.
A simple filter: if you cannot state the hypothesis in one sentence — “showing people instead of studio shots will lift conversion rate on feed placements” — it is too early to run the test. The underlying discipline is the same one described in our guide to creative testing without burning budget.
How long to run it, and when a result is real
Plan for four to six weeks. That is not an arbitrary number. First, Performance Max needs time to redistribute impressions inside each arm and settle into a stable placement mix. Second, the decision is made on conversions, and conversions accumulate slowly.
The system estimates the required duration from your current campaign volume. If it tells you eleven weeks, that is a signal not to launch this test in this form — not permission to ignore the estimate.
Rough thresholds worth holding yourself to (directional, not a standard):
| Parameter | Bare minimum | Comfortable |
|---|---|---|
| Conversions per arm | 50 | 100+ |
| Duration | 4 weeks | 6 weeks |
| Full weekly cycles | 3 | 4–6 |
| Gap between arms | 10% on the decision metric | 15–20% |
A 3–5% gap on fifty conversions is not a result, it is a wobble. Such a test closes with “hypothesis not confirmed,” which is a perfectly useful outcome: you did not roll out a change that does nothing.
One detail on the decision metric. Use the metric your campaign optimises toward — conversion value and ROAS on a value strategy, cost per action on a target CPA. Comparing arms on CTR is tempting and misleading: in Performance Max a higher CTR often just means a different placement mix. If you want the full picture of where your budget actually went, pair the experiment read with Performance Max channel reporting.
Settings that block the experiment before it starts
The most common reason the create button is greyed out, or the test refuses to save, is an incompatible campaign setting. Check these first:
- Shared budgets. Campaigns on a shared budget are not eligible: splitting traffic conflicts with a pooled spend ceiling.
- Portfolio bid strategies. Same logic — a portfolio optimises across campaigns, while an experiment needs isolated economics. If you rely on portfolios, their behaviour is covered in our guide to portfolio bid strategies and shared budgets.
- Smart Bidding Exploration. If the campaign runs Smart Bidding Exploration, asset experiments are unavailable — two systems reshuffling traffic at once make the result unreadable.
- Disapproved assets or policy issues. An asset under disapproval will not be accepted into the treatment set and the configuration will not save.
- Asset group limits. Control plus treatment plus common must still fit inside the per-type asset group limits. If you hit the ceiling, move elements into common.
Separately: the editing lock covers the whole asset group taking part. Schedule the launch so that no seasonal offer change or promo falls inside those four to six weeks. If a peak period is approaching, wait it out — the preparation sequence is in our Black Friday promo peak playbook.
Mistakes that turn a test into noise
Changing two variables at once. New images and new headlines. If the treatment wins, you cannot say what won, and you cannot reproduce it anywhere else.
Launching during structural edits. Raising the budget 60%, changing the target ROAS or adding a new feed in parallel. Any of these restarts learning and contaminates the arms unevenly. Which edits reset learning is covered in our piece on the Smart Bidding learning period.
Reading results on day five. Week one in Performance Max almost always skews toward one arm simply because of the order in which combinations get tried. Early reads are the main source of false wins.
Stopping the moment the numbers look good. Classic peeking: check daily, stop at the first favourable slice, and your false-positive rate climbs sharply. Fix the end date before you launch.
Testing on a campaign that is already broken. If conversion tracking is misfiring or value is passed incorrectly, the experiment will faithfully compare two arms on bad data. Run the conversion tracking diagnosis first, then test.
What to do with the result
There are three outcomes, each with its own action.
The treatment wins significantly. Apply the set to the asset group, log the hypothesis, and carry the principle — not the files — into other asset groups as a separate test. What worked for accessories is not obliged to work for electronics.
The difference sits inside the noise. Verdict: this variable does not move the needle at your current volume. That is knowledge too — it removes a hypothesis from the queue and saves you next month. Move up to a bigger swing: not “a different shade of background,” but a different type of visual altogether.
The treatment loses. The most underrated outcome. You just avoided rolling out a change that would have cost you money — which is exactly what would have happened without the test, because the new creative “looked fresher.”
Either way, keep a log: date, hypothesis, split, duration, conversions per arm, decision metric, decision. Six months in, that log becomes the only reliable source of truth about what works in your account, and it stops the team from re-running the same test every year.
Turning experiments into a process
Asset experiments die in a “we’ll run one when there’s time” culture. What works is a conveyor:
- A hypothesis backlog. One sentence each, plus the expected effect size. Prioritise by effect size divided by production cost.
- One active test per campaign. Parallel tests in the same campaign compete for traffic and stretch both timelines.
- Creative production runs one cycle ahead. While a test runs for four weeks, the next set is already in production — otherwise you get two-week gaps between tests.
- A quarterly review. What confirmed learning has been rolled out everywhere, which conclusions have aged out, which hypotheses should be retired.
If you run several accounts, keep a shared hypothesis register by vertical: a proven visual mechanic often transfers between similar offers and saves an entire test cycle. For the broader control surface you still have over the channel, see our guide to taming Performance Max, and for query-side steering, search themes in Performance Max.
The infrastructure behind continuous testing
Continuous testing is rarely limited by ideas. It is limited by account stability: you cannot run a six-week experiment on an account that may stop in week three. That is why teams operating at volume run their tests on accounts with impression history and a confirmed advertiser status — which is what Google Ads agency account rental from PPC Rebels is for. The wider toolset the team uses day to day is listed under PPC Rebels services.
Building the treatment set: a working method
The hard part of an experiment is not launching it — it is sorting assets into buckets. A mistake here makes every subsequent week meaningless. A four-step procedure that holds up:
Step one: write down what must appear in both arms. Logos, business name, the baseline headlines and descriptions without which the main formats will not assemble. All of it goes into common assets. When in doubt, make it common — better over-cautious than stuck with an arm that cannot serve on half your placements.
Step two: state the hypothesis so it can be checked with a number. Not “the new creative is better,” but “a visual with a person in frame will lift conversion rate over product-only photography by at least 15%.” That threshold is your decision boundary, and it must be named before launch rather than reverse-engineered from the result.
Step three: equalise count and type. The number of images, headlines and videos must match between control and treatment for each asset type. If control has eight images and treatment has four, Performance Max will assemble control combinations more often simply because it has more to choose from — and you get a skew that has nothing to do with creative quality.
Step four: review previews for both arms. Walk the main formats and confirm both arms assemble correctly. The classic catch at this stage: the treatment arm has no vertical video, so a chunk of inventory is unavailable to it. Five minutes here saves a month.
Sizing the test yourself
The interface gives its own estimate, but it helps to be able to do the arithmetic — especially when you have to explain to a director or client why a 4% gap means nothing.
A back-of-envelope check: to distinguish two variants reliably, the gap between them should be roughly double the random spread. Random spread in conversion counts is approximated by the square root of the count. At 50 conversions per arm the root is about 7, so a swing of ±7 conversions (±14%) is entirely normal even with no real difference. At 200 conversions the root is 14 — ±7%. That is where the “100+ conversions per arm” rule of thumb comes from: only then do effects smaller than 15% become visible.
| Conversions per arm | Natural spread | Smallest detectable gap |
|---|---|---|
| 30 | ±18% | Only very large effects, 40%+ |
| 50 | ±14% | 30%+ |
| 100 | ±10% | 20%+ |
| 300 | ±6% | 12%+ |
| 1000 | ±3% | 7%+ |
Treat these as rough orientation rather than a rigorous method; their job is to tell you, before launch, what size of effect you are even capable of seeing. If your hypothesis predicts an 8% lift and the campaign produces 60 conversions in six weeks, there is no point running the test — there will be no answer.
When you do not have the volume for an experiment
Most accounts fall short of the conversions a clean Performance Max experiment needs. That is not a reason to stop testing — it is a reason to change method.
- Move up to micro-conversions. If purchases are scarce, test against a higher-volume intermediate event: add-to-cart, checkout start, a qualified scroll depth. The decision transfers to the main metric with caveats, but it stops being guesswork. The approach is in micro-conversions and funnel optimisation.
- Make the hypothesis bigger. Small effects at small volume are undetectable in principle. Test a whole concept rather than a shade — a 40% gap shows up even on a modest sample.
- Test at a noisier but faster level. CTR and engaged-session share accumulate an order of magnitude faster than conversions. Weaker evidence, but it gives a signal within a week and lets you discard clearly bad sets without burning a full cycle.
- Pool accounts in the same vertical. If you run several similar projects, test the same hypothesis in three accounts in parallel and look at the direction of the result. That is not statistics, but a consistent result in three places beats one noisy test by a wide margin.
One more note for search-heavy accounts: if most of your volume runs through search campaigns, start with ad testing rather than Performance Max — the signal there is cleaner. The method is in A/B testing responsive search ads.
FAQ
How is an asset experiment different from a campaign experiment?
A campaign experiment duplicates the whole campaign and tests settings — bid strategy, goal, structure. An asset experiment works inside one asset group and changes only the creative. For creative hypotheses the second is cleaner: fewer confounding variables.
Can I test several asset groups at once?
Not within one experiment — you pick a single group. You can test groups sequentially, or in parallel across different campaigns, but parallel tests inside one campaign split the traffic and lengthen both.
What happens to the assets when the test ends?
If you do not apply the experiment, the group reverts to the control set. If you apply the winner, it becomes the group’s live set. The editing lock lifts once the test closes.
Why is the create-experiment option greyed out?
Usually a shared budget, a portfolio bid strategy, Smart Bidding Exploration enabled, or disapproved assets in the group. Check those four in that order.
Which traffic split should I pick?
50/50 if the campaign can absorb a weak treatment arm — it delivers an answer fastest. 70/30 or 80/20 if the campaign is business-critical. Anything below a 20% treatment share usually means the test never gathers enough data.
Can I stop the test early if the treatment arm is tanking?
Technically yes, and with a genuine collapse that is sensible. But “I am stopping because day seven is down 12%” is almost always premature — early Performance Max behaviour is unstable. A reasonable rule: stop only on a sustained decline lasting more than two weeks.
Does the experiment affect algorithm learning?
Splitting traffic inside a campaign does not reset learning the way a bid strategy change does, but it does reduce the data each arm receives. That is exactly why experiments are uninformative at low volume: neither arm gets enough signal.
Should I change the budget during the test?
No. Any budget change shifts both arms unpredictably. Fix the budget before launch and leave it.
What if the season changes mid-test, or a promo runs?
External shifts hit both arms roughly equally, so the relative comparison stays valid — though you cannot compare absolute numbers to a prior period. If the promo only touched products concentrated in one arm, restart the test.
Can I test automatically created assets against my own?
Yes, but that is a different variable type in the experiment wizard, not “assets provided by you.” It answers whether automation beats manual production in your vertical, and deserves its own test. Set up brand guidelines before running it, or you will be comparing your creative against automation’s worst version rather than its real one.
How many tests per year can one campaign realistically support?
At four to six weeks per test plus a week for preparation and analysis: eight or nine sequential tests a year per campaign. That is the practical ceiling to plan your backlog around.
