Campaign-level lift is not account-level lift

What the number actually says

Google’s figure is specific and, read carefully, modest. AI Max for Search campaigns see an average of 7 percent more conversions or conversion value at a similar cost per acquisition or return on ad spend, when using the full feature suite, compared with using search term matching alone. Google’s footnote limits it to non-retail advertisers and 2026 internal data. Every clause narrows it.

It compares AI Max with more features turned on against AI Max with fewer. It is an average across an unstated distribution. Google states it as a result for AI Max campaigns, which Digital Traction reads as a campaign-level measurement. And it is the platform reporting on its own product, which is normal and worth labeling rather than treating as either gospel or dishonesty.

None of that makes it wrong. It makes it a narrower claim than the way it gets quoted, which is usually as though AI Max makes accounts 7 percent better.

Google’s AI Max Help page publishes a second figure: advertisers that turn on AI Max typically see 14 percent more conversions or conversion value at a similar CPA or ROAS, from 2025 internal data for non-retail advertisers. Google states no baseline for it. Its wording, advertisers that activate AI Max in Search campaigns, reads as AI Max against no AI Max rather than the full suite against part of it, so the two numbers are not a before and after.

Why the level of measurement decides the answer

AI Max works by widening which queries a campaign can serve, beyond the keyword list, using assets and landing page content. If a campaign expands into queries another campaign in the same account was already winning, the conversions move. At campaign level that reads as a gain. At account level it reads as nothing.

This is not a criticism of the feature. It is a property of any change whose mechanism is expanded matching inside an account that already covers the space. The more complete your existing coverage, the more of the apparent lift is reallocation.

It also explains a pattern operators report and find confusing: automation tests that look strong on the test campaign and invisible on the account. Both readings can be accurate at once, because they are measuring different things.

The corollary is uncomfortable for well-run accounts. An account with thin coverage and an underused long tail has real room for expanded matching to find genuinely new demand. A mature account with dense coverage has less, and is therefore the account most likely to see campaign-level lift that does not survive to the bottom line.

What an honest test looks like

Measure the account, not the campaign, over a period rather than against a sibling. Fix total budget, hold the conversion definition still, and compare qualified outcomes for the whole account before and after, allowing enough time for the bidding to settle.

Fixing budget is the part most tests skip, and it is the part that makes the result readable. If total spend can drift, a lift in conversions may simply be a lift in spend, and cost per acquisition will tell you that only if it was stable to begin with.

Holding the conversion definition still matters just as much. Changing what counts during a bidding test destroys the comparison, and the temptation to fix measurement and test automation in the same month is strong precisely because both are overdue.

A campaign experiment is still worth running, but read it as a signal about that campaign, not a forecast for the account. The two numbers answer different questions and only one of them is the one paying for anything.

The general lesson, which outlives this feature

Platform-reported performance figures are almost always measured at the level where the feature operates. That is the honest place for a platform to measure. It is not the level a business is run at, and the gap between the two is where most disappointment with automation lives.

This will recur with every automation feature that follows, because the structure is the same each time: the platform reports the effect on the surface it changed, and the advertiser experiences the effect on the whole account.

The defense is a habit rather than a tool. Whenever a performance figure appears, ask what was compared, at what level, over what period, and against what alternative. For the AI Max figure the answers are: full feature suite against search term matching alone, non-retail advertisers, 2026 internal data with no stated test period, a campaign-level reading, and reported by the vendor. That is enough to decide whether it applies to you, which is all a number needs to do.

Before adopting an automation feature on a platform’s reported lift, ask what level it was measured at. Then run your own read at account level with budget fixed and the conversion definition unchanged, because that is the only version of the number that pays for anything.