Google Ads Cross-Campaign Budget Tests: Can Yours Be Conclusive?

Google is rolling out the ability to A/B test budgets and ROI targets across several Search campaigns in a single experiment. It is a genuinely useful addition, and it will produce a lot of inconclusive results, because the constraint on a budget test has never been the tooling. It is whether your account generates enough conversions to tell a real effect from noise.

The short answer

Google announced cross-campaign budget and ROI target A/B tests on 20 August 2026, rolling out through September. Before designing one, size it: at 80% power and 95% confidence you need roughly 16 ÷ (relative effect)² conversions per arm, so detecting a 10% change needs about 1,600 per arm. Accounts under roughly 2,300 conversions a month cannot resolve a 10% effect in six weeks and should test a bigger change or use a geographic holdout.

What shipped, and what it sits on top of

Google announced on 20 August 2026 that advertisers would be able to test different budgets and ROI targets across several Search campaigns inside a single A/B test, with the capability rolling out through September. The stated use case is the one every account runs into: you are considering a materially larger budget, and you want to know what it does to efficiency before you commit it across a group of campaigns rather than one at a time.

Two related changes landed alongside it. AI Max experiments can now run with Brand Controls and location settings enabled, so testing AI Max no longer requires switching off the guardrails you put there deliberately. And Performance Planner now projects how a bidding or budget target change would affect existing campaign performance, with the suggested change applicable in one click.

All of this builds on Campaign Mix Experiments, the cross-campaign framework Google opened in beta earlier in 2026. Its mechanics are worth knowing precisely, because they determine what a valid test looks like.

Mechanic	How it works	What it means for your test design
Arms	Up to five experiment arms, each with a different mix of campaigns	Five arms split your traffic five ways. Each additional arm costs you statistical power, so two or three is usually the honest choice
Campaign overlap	The same campaign can sit in multiple arms, with traffic split between them	Useful for holding a constant, and a source of interference if the arms then compete in the same auctions
Traffic splits	Custom splits down to a minimum of 1%	A 1% arm will almost never reach significance. Small splits are for risk containment, not for learning
Normalisation	Results are normalised to the lowest split for comparison	Your effective sample size is set by your smallest arm, not by your total spend
Campaign types	Search, Performance Max, Shopping, Demand Gen, Video and App, excluding Hotel campaigns	This is the only instrument in the platform that can compare across campaign types
Budgets	Shared budgets are not supported — each campaign needs its own	Accounts running shared budgets have setup work to do before they can test at all
Duration	The documented guidance is to run for at least six to eight weeks	That window has to contain your conversion lag and avoid a seasonal break. In Q4 it usually does not
Single features	The documentation directs you to the dedicated feature experiment instead, such as Broad Match experiments	Campaign mix is for strategy questions, not for feature toggles

The sizing question almost nobody asks first

An experiment can only detect an effect large enough for its sample size. Before you design the test, work out the smallest effect your account could detect, then compare it against the effect you actually expect. If the expected effect is smaller than the detectable one, the test will come back inconclusive no matter how carefully you set it up, and you will have spent six to eight weeks learning nothing.

For a conversion-rate style comparison between two equal arms, at 80% power and 95% two-sided confidence, the sample size needed per arm reduces to a usefully simple approximation:

The sizing rule

Conversions needed per arm ≈ 16 ÷ (relative effect)². To detect a 10% relative change you need roughly 16 ÷ 0.10² = 1,600 conversions in each arm. This is the standard two-proportion sample size formula simplified for low conversion rates, equal splits, 80% power and 95% two-sided confidence.

Relative effect you want to detect	Conversions per arm	Total conversions in the test	Monthly account conversions needed for a 6-week test
5%	6,400	12,800	~9,200
10%	1,600	3,200	~2,300
15%	710	1,420	~1,030
20%	400	800	~580
30%	180	360	~260
50%	65	130	~90

Read the last column as a threshold rather than a target. Treat these figures as a floor, for two reasons. First, budget tests are usually judged on total conversions, cost per acquisition or return on ad spend, and those metrics carry more variance than a simple conversion rate, so real requirements run higher. Second, the numbers assume the only difference between arms is the thing you changed, which in a live auction is never quite true.

The practical consequence is uncomfortable and worth saying plainly: an account doing a few hundred conversions a month cannot resolve a 10% efficiency change in a six-week experiment. It can resolve a 30% change. So either ask a bigger question or use a different instrument.

A worked example

Take an account with 900 conversions a month, considering a 40% budget increase across five Search campaigns. Run an eight-week test at a 50/50 split and each arm collects roughly 900 conversions. Rearranging the rule above, the smallest relative effect that sample can detect is the square root of 16 ÷ 900, which is about 13%.

Now ask what effect you actually expect. A 40% budget increase on campaigns that are already close to full impression share typically moves cost per acquisition by high single digits, because the incremental volume comes from progressively less qualified auctions. Call it 8%. That is below the 13% the test can see, so the experiment will report no significant difference — and "no significant difference" will be misread as "the budget increase was free", which is the most expensive possible misreading.

Three legitimate responses, in order of preference:

Make the change bigger. Test the budget increase you would actually be excited about, not the cautious one. A 100% increase has a larger expected effect and is therefore easier to measure, and you can contain the risk with a smaller split rather than a smaller change.
Change the instrument. A geographic holdout measures incrementality with a cleaner counterfactual than a split inside the same auctions, and it does not care about your conversion rate variance in the same way.
Accept directional evidence and pre-commit to the decision rule. Write down before launch what you will do at each outcome, including the inconclusive one. "If inconclusive, we keep the higher budget and review in a quarter" is a defensible decision. Discovering you have no rule after eight weeks is not.

Which question belongs in which test

Cross-campaign experiments are the right tool for a narrow class of question. Using them for anything else is how accounts end up with a library of inconclusive tests.

The question	Right instrument	Why
Should we raise budget across this campaign group?	Cross-campaign budget A/B test	This is precisely what the September 2026 release was built for, and the group-level effect is what you need
Should we loosen the ROAS or CPA target?	Cross-campaign ROI target test	Target changes interact across campaigns competing for the same conversions; testing one in isolation misattributes the shift
Does AI Max beat our current Search setup?	AI Max experiment, with Brand Controls and location settings left on	Since the September update you no longer have to disable guardrails to test, so the test now reflects how you would actually run it
Is broad match worth it here?	The dedicated Broad Match experiment	Google's documentation directs single-feature questions to feature experiments rather than campaign mix
Is Performance Max cannibalising Search?	Campaign Mix Experiment with arms that change the mix, or a geo holdout	Cannibalisation is a between-campaign effect, invisible to any single-campaign test
What is incremental about our brand search spend?	Geo holdout	An in-platform split cannot create a true no-exposure group for brand demand that exists anyway
What happens if we spend 30% more next quarter?	Performance Planner as a prior, then a real test	The planner projects from your history; it does not measure. Use it to size the question, not to answer it
Which landing page converts better?	On-site A/B test	An ads experiment attributes the difference to the campaign, not to the page. This belongs in conversion optimisation

Six things that contaminate a cross-campaign budget test

Shared budgets. They are not supported. Accounts that pooled budgets for convenience have to unpick that first, and doing so mid-test invalidates whatever you have collected.
Auction overlap between arms. When the same campaign appears in several arms, or when arms target the same queries, the arms bid against each other. The measured difference then includes interference you did not intend to test.
Normalisation to the smallest split. Results are compared after normalising to the lowest split, so a 5% containment arm sets your effective sample size. If you want an answer, do not pair a large arm with a tiny one.
Conversion lag longer than your reporting window. If your sales cycle runs four weeks, the last four weeks of an eight-week test contribute incomplete data. Judge the test on conversions that had time to land, not on the raw window.
Learning periods inside the test. A budget or target change puts Smart Bidding back into learning. Early days will look worse than steady state, so either exclude the first stretch or extend the test, and decide which before you launch.
Measurement gaps that predate the test. A budget test is only as good as the conversion signal underneath it. If offline conversions arrive late they are excluded from the attribution modelling that Smart Bidding learns from, which we covered in the seven-day upload window. Fix the signal first; an experiment cannot rescue a broken feed.

What we'd do about it

Before launching, write a one-page test charter: the single primary metric, the expected effect size, the detectable effect from the sizing rule, the run dates, the exclusion window for learning, and the action you will take at each of three outcomes — better, worse, inconclusive. If the detectable effect is larger than the expected effect, do not launch. Change the question instead.

Read-out rules

One primary metric, chosen in advance. Cost per qualified lead or return on ad spend, not a dashboard of nine numbers you can pick from afterwards. Picking the metric after the fact is how every test becomes a win.
No daily peeking for decisions. Monitor for breakage daily, judge on the schedule you set. Repeatedly checking a running test inflates your chance of a false positive.
Same window, both arms, full weeks. Compare complete weeks so day-of-week composition matches, and hold the window fixed once the test starts.
Reconcile against your own numbers. Platform-reported conversions and your CRM or order data diverge, and a strategy decision should rest on the source that includes revenue quality. This is the blended versus platform ROAS problem in miniature.
Write down what you learned even when it is nothing. "This account cannot detect effects below 13% in eight weeks" is a durable, reusable finding that changes how you plan every future test.

The Q4 timing problem

A six-to-eight week test started now runs into the middle of peak trading. With Black Friday on 27 November 2026 and Cyber Monday on 30 November, an experiment spanning that period is measuring a budget change and a demand shock at the same time, and the two are not separable after the fact.

So the realistic calendar is short. Either start a test that completes before the first week of November, or wait until mid-January. If you start one now, keep it narrow and fast, and expect to throw it away if the seasonal curve arrives early. This is the same planning constraint that governs what you can safely change on the site in the same weeks, and it is worth setting both calendars together rather than discovering the conflict in November.

Frequently asked questions

What are Google Ads cross-campaign experiments?

They are experiments that test a change across a group of campaigns at once rather than one campaign at a time. Google announced on 20 August 2026 that advertisers can test different budgets and ROI targets across multiple Search campaigns inside a single A/B test, rolling out through September 2026. They build on Campaign Mix Experiments, Google's cross-campaign framework, which supports up to five arms across Search, Performance Max, Shopping, Demand Gen, Video and App campaigns, excluding Hotel campaigns.

How many conversions do I need to run a Google Ads budget experiment?

As a floor, roughly 16 divided by the square of the relative effect you want to detect, per arm, at 80% power and 95% two-sided confidence. That is about 1,600 conversions per arm to detect a 10% change, 400 for a 20% change and 180 for a 30% change. Over a six-week test with a 50/50 split, detecting 10% implies roughly 2,300 account conversions per month. Metrics like cost per acquisition and return on ad spend carry more variance than a plain conversion rate, so treat these figures as a minimum rather than a target.

How long should a cross-campaign experiment run?

The documented guidance for Campaign Mix Experiments is at least six to eight weeks. That window has to be long enough to cover your conversion lag and the Smart Bidding learning period after the change, and short enough to avoid a seasonal break. In practice this rules out tests that span Black Friday on 27 November 2026 and Cyber Monday on 30 November, because a demand shock and a budget change inside the same window cannot be separated afterwards.

Can I use shared budgets in a Campaign Mix Experiment?

No. Shared budgets are not supported, so each campaign in the experiment needs its own individual budget. Accounts that pooled budgets for convenience have to restructure before they can test, and restructuring partway through a running experiment invalidates the data collected so far.

Should I use a cross-campaign experiment or a geo holdout?

Use a cross-campaign experiment when you want to compare strategies inside the platform and you have the conversion volume to resolve the effect. Use a geographic holdout when you need a clean counterfactual, particularly for incrementality questions such as what your brand search spend actually adds, or when your conversion volume is too low to detect the effect you expect. A geo holdout creates a genuine no-exposure group, which an in-platform split cannot do for demand that exists anyway.

What does an inconclusive experiment result actually mean?

It means the test could not detect a difference at the size your sample allows, which is not the same as there being no difference. If your account can only resolve a 13% effect and the true effect is 8%, the result will read as no significant difference while the change is still costing or saving you real money. That is why the detectable effect should be calculated before launch and the action at each outcome, including inconclusive, written down in advance.

The takeaway

The new cross-campaign tests remove a real limitation: until now, group-level budget and target questions had to be answered campaign by campaign, which is the wrong unit of analysis. Use them for exactly that. But size the test before you build it, because the sizing rule will tell most accounts something unwelcome and useful, which is that the question they want to ask is smaller than the answer their data can support. When that happens, ask a bigger question, switch to a geographic holdout, or pre-commit to a decision rule you can live with. And if the six-to-eight week window would run through peak trading, it is better to wait until mid-January than to collect eight weeks of data you cannot interpret.

Google Ads can now A/B test budgets across campaigns. Most accounts can't resolve that test