The Trade Desk Says Its AI Cut CPA 32% Across 62 Campaigns

Count the AI performance claims that reached your inbox this month, then count how many named a sample size, a comparison, and a time period. Usually none of the three, and you’re expected to renew on it anyway.

The Trade Desk just published one that answers two. That makes it the most useful vendor number of the month whether or not you ever buy programmatic, because it shows what a checkable claim looks like and what one still hides.

The company released Zuma, the latest version of its Kokai buying platform, on August 27, and says a global platform analysis showed a 32% CPA improvement on average with Koa Optimizations enabled. CPA is cost per acquisition, the price of one conversion.

The footnote is the interesting part:

“The Trade Desk platform data, campaigns comparing the upgraded model with the previous model (n=62); results are statistically significant (p < 0.05).”

62 campaigns. A stated significance test. Read a month of vendor AI claims and count how many give you either one.

That footnote is a gift, and it should be treated as one. It’s also, read closely, a much narrower claim than the number on the slide.

What the sample size does and doesn’t buy you

The footnote establishes three things: 62 campaigns, a new-model-against-old-model comparison, and a result clearing a standard significance bar. That’s checkable in a way “our AI improves performance” never is.

Four things it doesn’t establish, and each changes what the 32% means.

No time period. The page doesn’t say whether these ran a week or a quarter. Optimization results move with campaign maturity, so the same 32% over short flights and long ones describes two different products.

The comparison is the old model, not the alternative. Koa against the previous Koa. Not against a competitor’s optimizer, and not against a competent human buyer with the same budget. Those are the two comparisons you face at renewal, and neither was run.

It’s The Trade Desk’s own platform data. The company holds the numbers, defined the test, chose the 62 and reports the outcome. Nothing there is improper, and none of it is independent. That’s the same gap we found when companies couldn’t tie their AI spend to a result they hadn’t measured themselves.

“Average” is doing quiet work. A 32% average is compatible with most campaigns improving slightly and a few improving enormously. Without a distribution, you can’t tell whether the median advertiser saw anything close to it.

Ask who benefits from you believing the number, then look at what it was measured against. The Trade Desk sells the platform, and the platform’s own prior version is the easiest opponent available.

None of that is specific to programmatic. Swap in your email platform, your content tool, your call-scoring vendor: every one of them will show you a lift number this quarter, and the four questions above are the same four questions. The Trade Desk is just the rare case where two of them already have answers printed on the page.

The buyers using it are calmer than the slide

Digiday talked to media buyers who run on Kokai, and the temperature gap is worth noting. Haley Feazell, VP Global Media at Mindgruve, on the pace of improvement:

“I wouldn’t say it’s been ‘leaps and bounds’. It’s been more of a steady evolution, with a lot of smaller improvements and refinements”

That’s someone with hands on the platform describing incremental gains, which is a different picture than a 32% headline suggests. Both can be true if the lift is concentrated in campaigns unlike hers.

Dominic Johnson, Associate Director of Brand Media at Collective Measures, went further:

“The Trade Desk seems to still be playing catch up in terms of their AI features”

Catch-up is a competitive judgment rather than a performance one, and it’s the judgment that moves budget. A 32% internal lift lands in a market where the question is which platform to be on, not whether this year’s model beats last year’s.

Tom Swierczewski, VP of Media Investment and Partnerships at Goodway Group, named what buyers actually wanted: more control over the buy, “in some ways returning to the control we had in Solimar”. The most-requested feature in an AI release was the manual levers coming back, which tells you what three years of automation cost these teams in practice.

Why the context matters

Kokai launched three years ago, buyers said it didn’t match how they worked, spend moved elsewhere, and the revenue miss took about a third off the stock in the fourth quarter of 2024. So this is a repair job with an AI number attached.

One detail from that repair travels further than the rest. AdExchanger reports the AI features are all opt-in, the same direction as the five social platforms that now let an agent buy ads directly. A published result from an opt-in feature comes from the advertisers who chose to switch it on, and that is never a random group. Check that on any vendor number you’re handed.

The counter-case: don’t punish the vendor who showed the math

There’s an obvious way this piece goes wrong.

The Trade Desk published a sample size, named its comparison, and stated a significance level. Almost nobody does. If a disclosed denominator draws a thousand words of scrutiny while an undisclosed one gets a free pass, every vendor learns to disclose less.

So the standard cuts the other way. A claim with a sample size deserves more weight, not less. The scrutiny here is possible because the disclosure exists, which is the argument for making every other vendor match it.

The verdict: 32% is the best-documented AI performance number a buying platform has published this year, and it still only says that this year’s Koa beat last year’s Koa on 62 campaigns of unknown length, chosen and measured by the company selling it. Take it as evidence the model improved. Don’t take it as a forecast for your account. And when the next vendor shows you a percentage with nothing under it, you have the three questions ready: how many campaigns, measured against what, over how long.