When a client asks if ad automation is working

2 questions arrive together.

The platforms have taken controls away from media buyers. Google said it would remove the campaign-level language setting from Search campaigns from late September 2026. Meta removed placement exclusions from ad sets and left bid adjustments in their place. Account-level blocks still exist under Advertising Settings, which matters to any client with a brand-safety clause in its contract. Spend has followed. Meta said on its Q2 2026 earnings call that Advantage+ was running at over $75 billion a year.

So clients ask 2 things at once. Is ad automation working? And if the platform does the buying, what is the agency for? Both get the same answer. The agency's value is knowing what the campaign caused, which the platform's own report doesn't show.

Tell them who published the numbers.

Almost every number behind ad automation comes from the company selling it. Google's Performance Max figure rests on data from late 2022, and its AI Max benchmark leaves out retail, as its own footnote says.

The independent evidence is mixed.

The largest set of experiments comes from a firm that sells incrementality testing, so it isn't neutral, but it runs real tests. Across 640 geo-holdout experiments on Meta since 2024, Advantage+ beat manual campaigns in 42% of them. An analysis of more than 250 Search campaigns on AI Max found a median 13% lift in conversion value alongside a median 16% rise in cost per acquisition.

Meta's own words point the same way. On its Q4 2025 earnings call it said its incremental attribution feature drives a 24% increase in incremental conversions against its standard model. The standard model is the one most advertisers read as performance.

A line for the meeting: the platform's report shows the platform's model, and a holdout test shows what the campaign caused.

Agree who owns the test.

Settle before the test whether the budget will move on the result. If it won't, don't run it. Agree the measure up front too, as incremental orders or revenue from the client's own records rather than the platform's attributed conversions.

Keep the client's own numbers.

When the levers move inside a platform, its reporting moves with them. An advertiser whose only record lives there has nothing independent to argue from. Custom attribution, available through Meta's Conversions API, lets an advertiser send its own order-level conversions back, so delivery optimises to the client's definition of a sale, not the platform's.

Price what the platform can't do.

If the fee is tied to set-up work the platform now automates, it shrinks with that work. Price what's left instead. That's designing the test, running the holdout, reading it and making the call. A geo-holdout takes 3 to 5 weeks and enough spend in each cell to reach significance, so scope it per test, with the client's spend as an entry condition.

Smaller accounts can't fund a holdout. For them, scope a lighter monthly check that reconciles the client's own orders against what each platform claims and flags the gap. It won't prove cause, but it stops the vendor's report being the only record.

What to say in the room.

"Ad automation does the set-up work well. Its reported results are the vendor's, and the independent tests are mixed. We'll test what it caused, on your numbers, and move the budget on the result."

The evidence is in Dataflos's article.

The full review, with every vendor claim and the experiments behind it, is in Dataflos's article on the ad automation numbers. Dataflos is Twist Media Group's data team. For the same conversation about AI search, see when a client asks for an AI visibility report.

For agencies, our data and measurement service builds the testing plan for a client account. For brands, performance marketing runs paid media and measures it.