Trade Desk's Zuma: the 32 percent CPA claim rests on 62 campaigns

The Trade Desk released Zuma, an update to its Kokai media buying platform, and the number attached to it is a 32 percent average improvement in cost per acquisition. PPC Land reported on 27 August 2026 that the figure rests on a sample of 62 campaigns and compares one version of the optimisation model against the previous version, which is not the same claim as Koa beating no Koa.

That gap between the headline and the footnote is the reason this release is worth reading carefully rather than repeating. Zuma is built around interface changes, in-platform AI assistance and new measurement tools, and several of the capabilities that make the announcement interesting are restricted to closed beta. The reporting comes from PPC Land.

What The Trade Desk shipped, and when it said it would

According to PPC Land's 27 August 2026 report, the release is labelled the third-quarter 2026 highlights package from The Trade Desk, and the company groups the changes into three categories: interface, agentic AI and optimization, and enhanced measurement.

The timing carries context. The same report notes that Zuma arrives three weeks after chief executive Jeff Green named the Zuma upgrade first among five priorities for stabilising a business whose shares fell 24 percent on second-quarter results and soft third-quarter guidance. Green told analysts on 6 August that Zuma would ship later in the month and would concentrate on navigation, workflow and troubleshooting. It shipped on that schedule.

Eric Bosco, a product leader at The Trade Desk, described the release in a post on LinkedIn, saying the objective was not only new features but a cohesive experience that ties the platform's functionality together. He listed three objectives: faster decision-making through a revised interface and AI-assisted audience expansion, higher return on media spend through AI-driven optimizations, and better evidence of marketing impact through conversion lift management.

The 32 percent figure and the footnote underneath it

This is the part that will end up in a slide deck, so it deserves the most space. PPC Land reported on 27 August 2026 that a global platform analysis showed a 32 percent CPA improvement on average with Koa Optimizations enabled. The footnote to that claim states that the figure comes from The Trade Desk platform data, that it compares campaigns on the upgraded model against the previous model, that the sample is 62 campaigns, and that the result is statistically significant at p below 0.05.

Read those two sentences next to each other and the shape of the claim becomes clear. The comparison is between model versions, not between campaigns using Koa and campaigns not using it. PPC Land draws out that distinction explicitly, and it changes what the number can support. It is evidence that a newer optimisation model outperformed an older optimisation model on a set of campaigns the vendor selected from its own platform. It is not evidence about what happens when a buyer turns Koa Optimizations on for the first time, because no campaign in the comparison was running without it.

What a sample of 62 does and does not tell you

Statistical significance at p below 0.05 is a real result and should not be waved away. It means the difference observed is unlikely to be noise under the test that was run. What it does not mean is that the effect size generalises. A sample of 62 campaigns drawn from a global platform is a small slice of a very heterogeneous population: different verticals, different budgets, different countries, different creative, different conversion definitions, different starting CPAs. An average across 62 campaigns can be driven by a handful of large movers, and the distribution behind the average is not published.

The honest reading is that the direction is supported and the magnitude is not transferable. Nobody should tell a client that switching model versions will cut their CPA by roughly a third, because that is an average over campaigns that are not theirs, with a spread that has not been disclosed.

Why a vendor before-and-after is weaker evidence than a holdout

There is a structural issue with any before-and-after comparison run by the vendor on the vendor's own platform, and it has nothing to do with anyone's honesty. Campaigns that ran on an older model ran at a different time, against a different auction, with different seasonality, different inventory supply and different creative fatigue. If the newer model shipped later in the year, then the comparison is also a comparison of two calendar periods. None of that is separable from the model change without a design that holds the period constant.

A holdout test does hold it constant. In a holdout, comparable campaigns run in the same window, and a randomly assigned share of the eligible traffic or the eligible campaigns is deliberately kept on the old treatment. Both arms face the same auction, the same seasonality and the same supply, so the difference between them is attributable to the treatment rather than to the calendar. The report does not describe a holdout, and there is no indication in it that one was run. Absence of a holdout is not evidence that the effect is absent; it is a reason to hold the specific number loosely.

How much of Zuma is actually available this quarter

PPC Land's 27 August 2026 report is careful about availability, and a buyer planning a quarter should be equally careful. The Koa AI assistant, described as an in-platform companion that helps locate information, uncover insights and give on-demand support while coordinating a growing set of specialised agents across campaign creation, audience building, troubleshooting and performance analysis, is in closed beta. The Koa frequency features, which use AI to adjust frequency settings dynamically against campaign objectives, are also in closed beta.

Koa Optimizations, the existing bid and inventory prioritisation layer, is not marked beta. It is described as more effective through the combination of real-time performance signals with historical insights, and it is the component carrying the 32 percent figure. So the piece with the number attached is generally available, and the pieces that make the release feel like a change in how the platform is operated are not.

Closed beta matters for planning in a specific way. It means the capability cannot be assumed in a media plan, a staffing plan or a client commitment for the current quarter, because access depends on a decision the vendor has not published criteria for. A trading desk that has been offered beta access has something real to test. A trading desk that has read the release page has something to ask about. Those are different positions, and it is worth knowing which one you are in before a capability appears in a pitch.

The agent layer is a continuation, not a new departure

The architecture described in Zuma builds on work The Trade Desk had already shipped. PPC Land notes that the company shipped Koa Agents in April 2026 with Stagwell as pilot partner, and published the Open Agentic Kit on the same day. What Zuma adds on top is a conversational surface that routes requests to those agents. Reading it as an entirely new capability overstates it; reading it as a repackaging understates it. It is a shell over an existing agent set, currently gated.

The interface changes and the bulk editing history

The first block of changes concerns navigation and comprehension rather than new buying capability. Refreshed page designs are described as making it easier to identify trends and locate areas needing attention, and an Applied Settings view surfaces targeting parameters in a single place rather than leaving them spread across separate configuration screens. For anyone who has tried to reconstruct why a line item is pacing the way it is, a single settings view is a genuine time saving, even though it buys nothing new.

Bulk editing is the more interesting item, and it is marked coming soon rather than available. The updated version will allow the potential impact of changes to be previewed before they are applied, with no date attached. PPC Land supplies the history that makes this notable: bulk editing has been a recurring friction point on Kokai, and when The Trade Desk shipped campaign management enhancements in September 2025, industry feedback indicated that bulk upload functionality had been reduced during the migration, with certain placement settings no longer editable in bulk.

The distinction PPC Land draws is precise and worth carrying forward. A preview-before-apply model addresses the risk of large simultaneous edits. It does not by itself restore the scope of what can be edited. Those are two different complaints, and shipping a fix for one is not a fix for the other. A buyer who lost the ability to bulk edit certain placement settings still cannot edit them, whatever the preview shows.

Audience Unlimited and the pricing that did not change

Audience Unlimited receives simplified access to third-party segments paired with AI-powered expansion in this release. PPC Land's 27 August 2026 report supplies the background: that product launched in September 2025 with a tiered structure charging 3.3 percent and 4.4 percent of impression costs in Control Mode, and bundling data access at no incremental cost in Performance Mode.

Zuma does not change that pricing, and the release page does not mention pricing at all. That is a fact worth stating rather than glossing, because data fees are a real line in a programmatic budget and a release that expands data access without touching data pricing leaves the economics exactly where they were. Anyone building a programmatic budget on the back of this announcement should price segments the same way they did last quarter.

The evidence table

The claim and the qualifications travel together in the source, so they should travel together in any summary of it. This is what PPC Land reported on 27 August 2026 about the headline number.

How to check the claim in your own account

A platform-wide average is not a forecast for any individual advertiser, but it is testable in miniature. The work is unglamorous and it is the only thing that turns a vendor number into a decision you can defend.

Record your baseline before anything changes. Pull CPA by campaign for a stable window, with the conversion definition frozen, and note the spend, the flight dates and any creative or audience change inside the window. A comparison against a baseline you reconstructed after the fact is not a comparison. It also depends on the measurement layer underneath recording the same event the same way in month two as it did in month one, which is why an unresolved GA4 measurement migration quietly invalidates any before-and-after you try to run on top of it.

Then hold everything else still. If the optimisation layer changes and you also refresh creative, widen the audience and raise the budget in the same fortnight, you have a result you cannot attribute. Sequencing changes is slower and it is the difference between knowing and guessing. The same rule governs a test on any auction-based channel, including the paid social campaigns that usually sit in the same plan, and it applies with more force when the variable being changed is a model you cannot inspect.

Finally, insist on a period-matched comparison where you can get one. If your trading desk can run a genuine holdout, that beats a before-and-after every time. If it cannot, at least compare like periods rather than a strong quarter against a weak one, and say plainly in the write-up that the design is weaker than a holdout.

What this means for Thai marketers

Start with what the source does not cover. PPC Land's 27 August 2026 report says nothing about availability in Thailand, nothing about Thai inventory, and nothing about whether the campaigns in the sample of 62 included any activity in this region. Anyone quoting the 32 percent number to a Thai client is quoting a global platform average with an undisclosed geographic mix.

The useful posture for a Thailand-based buyer running programmatic through a DSP is a short list of questions for the trading desk before the number gets into a plan. Ask whether the campaigns in the vendor's analysis resemble yours in vertical, budget size and conversion type, and accept that the honest answer may be that nobody knows because the composition was not published. Ask whether the optimisation model on your seat has actually changed, and on what date, because a model-versus-model result is meaningless to you until your seat has moved between the two. Ask what the desk's own before-and-after looks like on your account, and whether the period being compared contains a Thai seasonal event, since a comparison that straddles a major holiday period is measuring the calendar as much as the model.

Ask one more thing: what would count as disconfirming evidence. If the answer is that nothing in the reporting could show the change did not help, the test is not a test. This matters in a market where programmatic budgets sit next to search budgets in the same plan, and where a lift claimed for one channel is often a shift of credit from another. If you are running Google Ads alongside programmatic, watch total blended cost per acquisition across both, not the DSP's own CPA in isolation.

There is no public benchmark for a good programmatic CPA in Thailand, by vertical or otherwise, and there is unlikely to ever be one that survives contact with a real account. The mix of inventory, the conversion definition and the price point move the number too much: a retailer running ecommerce campaigns on a low average order value and a B2B advertiser counting a form fill are not measuring comparable things. The benchmark that works is your own trailing performance, recorded before the change, with the definition written down.

Frequently asked questions

Does Zuma cut CPA by 32 percent?

Not as a general promise. PPC Land reported on 27 August 2026 that a global platform analysis showed a 32 percent average CPA improvement with Koa Optimizations enabled, based on The Trade Desk platform data with a sample of 62 campaigns, significant at p below 0.05, and comparing campaigns on the upgraded model against the previous model. The number describes that comparison and does not forecast a result in any individual account.

Is the Koa AI assistant available to use now?

No, it is in closed beta according to PPC Land's 27 August 2026 report, as are the Koa frequency features that adjust frequency dynamically against campaign objectives. Koa Optimizations, the bid and inventory prioritisation layer that carries the 32 percent figure, is not marked beta.

Does the release change what Audience Unlimited costs?

No. PPC Land reported on 27 August 2026 that Zuma does not change Audience Unlimited pricing and that the release page does not mention pricing at all, leaving the tiered structure from that product's September 2025 launch, 3.3 percent and 4.4 percent of impression costs in Control Mode with data access bundled at no incremental cost in Performance Mode, as reported.

Is bulk editing fixed?

Not yet, and not in the way the earlier complaint described. The updated bulk editing is marked coming soon rather than available with no date attached, and it adds a preview of the potential impact of changes before they are applied, which PPC Land points out addresses a different problem from the reduced bulk upload scope reported after the September 2025 migration.

Is Zuma available in Thailand?

Not covered by the source. The 27 August 2026 report does not address regional availability, so a buyer here should confirm with their seat holder or trading desk rather than infer it from the announcement.

If programmatic sits in your media plan and a vendor performance claim is about to sit in your next deck, the useful next step is to write down your current baseline before anything is switched. Our team can help structure that test so the result means something when it lands.