Google Ads can now A/B test a budget increase across campaigns. Here's how to run one that answers the question

Every advertiser has had this argument. Someone wants to raise the budget, someone else says CPA will go up, and neither position is testable because the only way to find out has always been to raise the budget and see what happens. In September 2026 Google started rolling out the missing piece: an A/B experiment that can test a budget or ROI target change across several Search campaigns at once, rather than one campaign at a time.

What Google actually shipped

Three things were announced together on 20 August 2026, and they get conflated easily:

1. Cross-campaign budget and ROI target experiments. The headline change. Previously an experiment tested one campaign against a variant of itself, which made testing a budget increase close to useless: money reallocates between campaigns, so measuring one in isolation measures the wrong thing. From September you can put a group of Search campaigns into one experiment and change budgets or ROI targets across the group.
2. Brand and location controls inside AI Max experiments. You can now run AI Max tests with brand controls and location settings applied, so testing AI Max no longer requires dropping the guardrails you put there for a reason.
3. Performance Planner forecasts with one-click apply. Projections of what a bidding or budget change would do, applied directly to campaigns.

Only the first is an experiment. The third is a forecast, and the difference between the two is the whole point of this article: a forecast is a model's opinion about your account, while an experiment is a measurement taken from it.

Why testing a budget increase was broken before

Scaling spend changes performance through mechanisms that show up across an account rather than inside one campaign. Higher budgets buy lower-intent auctions, spend spills between campaigns that share keyword space, a smart bidding strategy re-learns, and the marginal customer costs more than the average one. Raise the budget on one campaign and watch that campaign, and you see some of that. You miss the part where the neighbouring campaign's volume fell by nearly the amount yours gained.

That is why so many scaling decisions get made on a before-and-after comparison, which is the weakest evidence available: it confounds your change with seasonality, competitor activity, promotions and everything else that happened that month. A cross-campaign experiment replaces the before-and-after with a same-period comparison, which removes most of that confounding in one move. That alone makes it worth using, and it makes the scaling playbook a question you can answer with data rather than argument.

Size the change to your conversion volume, not your ambition

The most common way these tests fail is that they cannot possibly succeed. An experiment can only detect a difference large enough to stand out from random variation, and the volume required grows sharply as the difference you care about shrinks.

As a rough planning approximation, the conversions you need per arm to detect a relative difference d at the usual 95% confidence and 80% power is about 16 ÷ d². Conversion counts dominate the variance in a CPA or ROAS comparison, so this is a reasonable back-of-envelope figure for test planning rather than a substitute for the significance calculation Google reports.

Read that table before you design anything. If your account produces 60 conversions a month per arm, you can detect a catastrophe and you cannot detect a 10% CPA drift, and no amount of patience within a quarter will change that. Two consequences follow. First, test big changes: a 50% or 100% budget increase is measurable where a 15% one is not. Second, if the decision you face is about a small difference, an experiment is the wrong instrument and you should decide it on margin economics instead.

A test design that produces a decision

1. Write the decision first. "If blended CPA in the test arm stays within 20% of control at 60% more spend, we roll the budget out permanently." A test without a pre-committed decision rule turns into a debate about the result.
2. Split 50/50. Equal arms reach significance fastest. A 90/10 split starves the variant and guarantees an inconclusive test.
3. Change one thing. Budget or target, not both, and nothing else in the account that touches those campaigns: no new ad copy, no landing page changes, no bid strategy switches.
4. Run it for at least two to three full weeks, longer for lower volume, and always in whole weeks so both arms see the same weekday pattern. Smart bidding also needs a learning period, so the first week of data is not representative.
5. Match the window to your sales cycle. If leads close in 30 days, a 14-day test measures form fills, not revenue. Either extend the window or accept that you are testing a proxy, and say which in the decision rule.
6. Check the conversion mix, not just the count. Scaling often buys more of a cheaper, lower-quality conversion. A stable CPA hiding a shift from demo requests to newsletter signups is a worse result than a CPA rise, which is the case for tracking cost per qualified lead rather than cost per lead.

Experiment, forecast or lift test?

These three tools answer different questions, and using the wrong one is how measurement programmes lose credibility.

The boundary that matters most is the one between the first and third rows. A budget experiment tells you what happens to measured conversions when you spend more. It says nothing about whether those conversions were incremental, because both arms sit inside the same platform's attribution. If the brand spend you are scaling would have converted anyway, an experiment will happily show you an excellent CPA while your total revenue stays flat. That question needs an incrementality test, and the two are complements rather than alternatives.

Four things this test cannot tell you

1. Whether the conversions are incremental. As above. Platform-measured lift is not business lift.
2. What it did to your other channels. Paid search scaling often eats organic clicks on the same terms and shifts credit between channels. Watch blended acquisition cost across the whole account, not just the campaigns in the test, which is the argument for blended ROAS as your headline number.
3. Whether the result survives seasonality. A test run in one demand environment tells you about that environment. A November result is a November result.
4. What happens above the level you tested. Scaling curves bend. A clean result at +50% is not a licence for +200%, and the next step needs its own test.

The Q4 timing problem

This feature has arrived at an awkward moment. Auction prices rise through Q4, conversion rates move, and a test that runs across a demand shift measures the shift as much as the change. Black Friday falls on 27 November 2026 and Cyber Monday on 30 November.

If you want a scaling decision backed by evidence for this peak season, the test window is now, not November. That is the same planning logic as the rest of the Q4 performance marketing plan: the decisions that need data have to be made before the data gets noisy.

What this changes in practice

Not the strategy. Budget scaling still obeys the same economics it always did, and the constraint on most accounts is still margin rather than measurement. What changes is the quality of the evidence available for the argument. Until now, the case for or against a budget increase rested on a before-and-after comparison that any competent sceptic could dismantle, so the decision usually went to whoever was more senior. A same-period comparison across the campaigns that actually interact is a materially better instrument, and it is free.

Use it for the decisions it fits, keep a forecast in its lane as a planning aid, and keep a lift test for the question of whether the channel is producing anything at all. Those three answers together are a measurement setup worth having.

Frequently asked questions

What did Google change about Search campaign experiments in September 2026?

Google announced on 20 August 2026 that advertisers can test budget and ROI target changes across multiple Search campaigns within a single A/B experiment, rolling out from September 2026. Previously experiments were run per campaign, which made budget tests unreliable because spend reallocates between campaigns that compete for the same queries. Google also added brand and location controls to AI Max experiments and one-click application of Performance Planner recommendations.

How long should a Google Ads budget experiment run?

At least two to three full weeks, in whole weeks so both arms see the same weekday pattern, and longer for lower-volume accounts. Smart bidding needs a learning period, so the first week is not representative. If your sales cycle is longer than the test window, the experiment is measuring a proxy such as form fills rather than revenue, and your decision rule should say so.

How many conversions do I need for a Google Ads experiment to be conclusive?

As a planning approximation, the conversions needed per arm to detect a relative difference d at 95% confidence and 80% power is roughly 16 divided by d squared. That is about 400 conversions per arm to detect a 20% difference, about 175 for a 30% difference and about 65 for a 50% difference. Accounts with modest volume can only detect large changes, which is why budget tests should use large budget changes.

What split should I use for a Google Ads experiment?

50/50. Equal arms give equal statistical power and reach significance fastest. A 90/10 split starves the variant of data and is the most common reason an experiment ends inconclusive after weeks of runtime.

Can a budget experiment tell me if my ads are incremental?

No. Both arms of the experiment sit inside the same platform's attribution, so the test compares measured performance at two spend levels rather than performance against no advertising at all. To answer whether conversions would have happened anyway, you need a geo holdout or incrementality test, which withholds spend from a control group. The two methods answer different questions and are best used together.

Is Q4 a good time to run a budget scaling test?

Generally no. Auction prices and conversion rates move quickly through peak season, so a test running across that shift measures the shift alongside your change. Test now through mid-October to inform peak decisions, scale on judgement during late October to early December, then re-test from mid-January when a stable baseline returns.

The takeaway

Cross-campaign budget experiments close a real gap: until now, the case for scaling spend rested on a before-and-after comparison that measured your change and the month it happened in together. Use the new experiment for large, in-platform scaling questions, at a 50/50 split, with a decision rule written before it starts and a change big enough for your conversion volume to detect. Keep Performance Planner as a planning aid rather than evidence, and keep a geo holdout for the separate question of whether any of it is incremental. Then run the test that matters before mid-October, while your baseline still holds still.