‘Future of targeting’: Media agencies tentatively explore vector-based planning

Media agencies like WPP and Dentsu are exploring vector-based targeting, which uses numeric embeddings of audience data to improve ad targeting without relying on IDs. The technique is still nascent but growing, with early tests in the U.S. and Europe, particularly for streaming and CTV campaigns.

Connect with execs from Axios, The New York Times, Paramount and more. VIEW PASSES https://digiday.com/events/digiday-publishing-summit-sep-26/?source=display&utm-source=display ‘Future of targeting’: Media agencies tentatively explore vector-based planning For some media agency executives, it’s the “future of targeting.” Others say it’s a theory that will only get you so far. Though the case studies remain unwritten, agencies like WPP are already running early-stage vector-based targeting solutions in key markets, including the U.S., while peers at Dentsu see it as the natural next step for contextual targeting. “It’s becoming more widespread,” said Alex Steer, chief data officer, WPP data & technology solutions. “We expect to see it growing.” Vector-based targeting uses similar ideas to those underpinning lookalike https://digiday.com/media-buying/media-buying-briefing-how-generative-ai-is-being-applied-in-big-and-small-settings/ or contextual targeting https://digiday.com/media-buying/how-contextual-targeting-providers-pitch-to-brand-clients-and-agencies-has-changed/ . It refers to the practice of finding an audience based on a combination of several different pieces of information, like the different “dimensions” to a coordinate, rather than targeting using keywords. The relevant data could include anything from geographic information, viewing history or propensity to purchase. It’s translated into numeric form by an agency and packaged together in a “vector embedding,” a file that can be read by a computer model operated by an SSP or ad tech partner on the supply-side of the programmatic pipeline. “Instead of me giving you one data point, I’m giving you a recipe that’s based on all of my data,” explained Steer. The technique can be used as a form of advanced lookalike targeting, or to identify niche audiences an advertiser might want to exclude. When paired with some of the early-stage agentic tools marketing services groups are developing for media planning, it can allow buyers to fire ads toward dozens — or potentially hundreds — of audience profiles based on very rich data, relatively quickly. Furthermore, this method of identifying and acting on audiences means agencies skip the use of IDs https://digiday.com/media-buying/without-googles-cookie-cutoff-hanging-over-them-buyers-are-lukewarm-on-alternative-ids/ — and their associated “identity tax”, as Steer put it. That’s the theory. In practice, vector-based targeting is “nascent,” according to Tara Kilcoyne, managing partner, addressable product enablement at Dentsu U.K. Planners using Dentsu Media Exchange DMX have been testing solutions for about a year, she said, across online video, streaming and digital display inventory. Kilcoyne sees it as a means of accessing efficiencies through accuracy. “Basically, we’re trying to buy more efficiently and buy the right audiences for our clients,” she explained. Steer said WPP Media began developing vector-based solutions two years ago. “It’s becoming more widespread in the U.S. and Western Europe. We’ve got activity live in Germany and Spain,” he said. Steer and Kilcoyne both declined to name clients using the technique. ‘Tons of interest‘ In particular, Steer indicated that the technique works particularly well when used for streaming or CTV campaigns. Amazon, for example, forgoes the vector-based terminology but offers a targeting tool called “Brand+” that provides similar capabilities. The tool, available globally since January, uses Amazon’s AI systems to identify audiences likely to act upon seeing an ad, based on the company’s first-party data. Advertisers using the tool typically experience a 71% rise in visits to detail pages within its e-commerce platform, according to a spokesperson for the company. “We’ve had conversations with quite a lot of the streamers around it,” Steer said. “I wouldn’t say that I have seen a hugely mature offer yet … most are prototyping.” One pull factor? Because they act as a means of translation, vector embeddings could allow advertisers to target more consistently across multiple streaming platforms and CTV providers, noted Gregory Cornuz, Equativ’s chief product officer. “ Vector-based methodologies apply very well to such an environment where legacy deterministic decision-making criteria cannot apply because of the scarcity and the fragmentation of the signal,” he explained. While streamers’ support for vector-based targeting might on the whole be at an early stage, media execs are gung ho for the practice. “This is very likely the future of where targeting could go,” said David Dworin, chief product officer at Comcast-owned FreeWheel. The company’s vector-based solutions currently remain at an experimental stage. “This is something we’re thinking about now, and probably planning a bigger implementation in the future,” said Dworin, adding that tests so far used direct buys but could be deployed in programmatic transactions in the future. Activating campaigns that separately target several dozen audiences, noted Steer, would ordinarily be labor-intensive for buyers. This is where vector-based targeting gets wrapped into parallel efforts to bring agentic AI into media planning and buying. “Instead of building one or two big audiences, I want to build 100 or 200 precise ones. Vectors make it easy and fast to do that,” he said. “ But what you’ve then got to be able to do is actually set up and buy the campaigns and produce the content. That’s where we’re starting to look at agents for.” Kilcoyne, too, said vector-based targeting was best thought of as a facet of “applied AI” in media planning. She said that Dentsu is working with several partners, among them Chalice and Equativ to develop a vector-based practice. “There’s tons of interest in it, but it’s still early stages for development across all parties,” said Tylynn Pettrey, svp of analytics and AI at Chalice. The AI company is working with several agencies to create the vector embeddings that compress audience targeting data into a usable packet, and then deliver it to an SSP or publisher. “The agencies are the drivers,” she said. Much of that work involves creating more efficient processes for encoding and decoding audience data. Though embedding provides a means of compressing information, there are limits to how much information can be included. “That’s where a lot of experimental work is being done right now,” said Pettrey, but she didn’t name the other companies it’s working with. ‘More pressing things to do‘ Some buyers see vector-based targeting as a logical progression from contextual targeting or lookalike techniques, and as part of a larger movement away from relying totally on deterministic-based targeting https://digiday.com/marketing/more-brands-are-blending-deterministic-and-probabilistic-data-for-hybrid-targeting-approaches/ . Others are leery of its practicality. “Vector-based targeting is definitely intriguing, and you can see the usefulness in the precision it provides,” said Jason Hartley, head of media innovation and trust at PMG. “It’s a promising thing for the future, but it’s probably going to end up just embedded in the targeting that the major platforms already do. For most brands, there are probably other more pressing things to do in the near term.” There are also drawbacks to the technique, many shared with the agentic systems some companies are using to implement it. Just as it’s difficult to parse why an LLM makes a specific hallucination, it’s hard to deduce why an AI model might have drawn connections between the data used to create an audience. “The terrible thing about embeddings is it can be a bit of a black box,” said Pettrey. Furthermore, the ad industry currently lacks agreed-upon technical language for the computer models that encode and decode vector embeddings, meaning agencies engaging in the practice need to understand multiple ways of working. “We all need a standard language,” she added. “It’s going to take some time for us to move both the technology and the behavior behind it” forward, said Dworin. “But once we do, it presents a lot of opportunity for us to make streaming much more valuable to advertisers across the funnel.” More in Media Buying Ad Tech Briefing: Private equity’s ad tech playbook returns https://digiday.com/media-buying/ad-tech-briefing-private-equitys-ad-tech-playbook-returns/ Plus how the IAB is seeking to modernize how digital video inventory is defined, bought and measured. Media Buying Briefing: How Mediasense and other consultancies are girding for a busy second-half https://digiday.com/media-buying/media-buying-briefing-how-mediasense-and-other-consultancies-are-girding-for-a-busy-second-half/ If the agency world is on the verge of another mediapalooza, then advisory firms need to be at the ready, including Mediasense which has a new CEO. Future of Marketing Briefing: The ad industry’s next mediapalooza is loading https://digiday.com/media-buying/future-of-marketing-briefing-the-ad-industrys-next-mediapalooza-is-loading/ Mediapalooza is coming. This time it might actually mean something.