WPP and Dentsu Are Rebuilding Ad Targeting Around a Number No One Can Audit Yet

WPP has spent two years rebuilding how it aims a media buy. The target is no longer an audience anyone can name. It’s a string of numbers.

Alex Steer, chief data officer at WPP’s Data & Technology Solutions, described the change to Digiday. Where a planner used to hand a platform one named segment, WPP now hands it a compressed version of everything it knows about a customer.

“Instead of me giving you one data point, I’m giving you a recipe that’s based on all of my data.”

That recipe is good at finding people. It’s also unreadable, and the agencies testing it say so out loud.

Strip the jargon and a “vector” is the audience rewritten as a long list of numbers, one slot per signal, from geography to viewing history to how likely someone is to buy. A machine can then measure how close two people sit in that space without ever being told what a single slot means. Bundle that list into a file an ad-tech system can read and you’ve got an “embedding.” Vector-based targeting means the buy chases whoever lands near your best customer in that number-space, rather than whoever matches a written label like “women 25-34, in-market for a car.”

The pull is that the written labels are running out. Third-party cookies keep dying and the clean deterministic signals keep thinning, so agencies are encoding what’s left into coordinates instead. This is the same squeeze pushing money around everywhere else in the buy: global ad spend cleared a trillion dollars while budgets stayed flat, which means every efficiency claim now gets a hard look. Vector targeting is being sold as one of those efficiencies. It rides up the same pillar of the martech stack that’s been folding standalone tools into platform features for two years.

It isn’t only agencies. Amazon’s Brand+ tool has been available globally since January and, by Amazon’s own account, drove a 71% rise in product detail-page visits for advertisers using it. Dentsu’s planners have been testing vector methods for about a year. FreeWheel keeps a version running at an experimental stage. David Dworin, chief product officer at FreeWheel, told Digiday the direction looks settled even if the plumbing doesn’t:

“This is very likely the future of where targeting could go.”

Here’s the part the future talk skips. The buyer can no longer explain the buy. Tylynn Pettrey, SVP of analytics and AI at the ad-tech firm Chalice, named the tradeoff plainly:

“The terrible thing about embeddings is it can be a bit of a black box.”

A black box is a problem when someone has to sign off. A named segment is auditable: a brand-safety lead can look at “in-market for a car” and approve or reject it. A vector is a coordinate. It can quietly pull in audiences a brand would never have bought on purpose, and no label sits there to inspect on the way in. Reach expands, and so does the gap between what the campaign did and what anyone can account for.

The mechanism problem underneath is that the industry has no shared language for these files. One agency’s embedding isn’t readable by another agency’s model, because the two were built differently, so a vector is only as portable as the exact system that made it. Pettrey put the fix in five words:

“We all need a standard language.”

That standard is being drafted right now. The IAB Tech Lab is building one under the banner “Agentic Audiences,” seeded by a protocol LiveRamp developed and donated to the group. It’s early. The lab’s own reference materials note that the vectors in its demo are hand-crafted for teaching, not outputs of a real production model. A shared vocabulary that would let a client audit a vector the way they audit a segment does not exist yet.

Which leaves the only question a marketing operator actually has to answer this quarter: is it worth budget now? Jason Hartley, head of media innovation at the independent agency PMG, gave the honest read for most teams:

“There are probably other more pressing things to do in the near term.”

He’s right, and the reason isn’t that the technology is weak. It’s that “the future of targeting” and “worth your Q3 media budget” are two different sentences. Most streaming and CTV offers are still prototypes. The audiences aren’t portable. And the one control a CFO or a brand lead will ask for first, the ability to see why the model bought what it bought, is the exact thing a black box can’t provide.

So the verdict is a boundary, not a no. Vector targeting will probably win, because aiming at cookies that are already gone is the losing alternative. But a technique you can’t audit belongs in the experimentation column, not the production line, until someone can answer the sign-off question in plain words. Fund the test. Give it a named owner and a small number. Don’t yet move it, or the budget it would eat, into the part of the plan you’re expected to defend.