WPP Media is testing synthetic audiences based on charity donations for programmatic buys

According to Tara Sadlak, group director of media delivery at WPP Media, early tests carried out across four beauty and fashion campaigns show a 2% lift in video completion rates, suggesting that more accurate targeting had led to higher engagement rates among audiences. Sadlak declined to name the brands in the test.

“We’re looking for discoverability and increased reach,” she told Digiday. 

Related Insights

WTF Series

WTF are synthetic audiences? Read More

Synthetic data — sometimes called digital twinning — refers to the use of AI to create audience profiles or consumer cohorts from survey data.

It’s been used by agencies such as Code & Theory and even smaller agencies like Marcus Thomas to create custom profiles, which can then be assembled into “focus groups” of AI personas, for example. Dentsu struck a partnership with startup Evidenza last year to use its synthetic audiences to inform media planning. And publishers like The Times and News UK have begun to offer advertisers access to synthetic personas built atop their own first-party data.

Agencies have shied away from using synthetic audiences as the basis for targeting bids, however. The risk that an LLM-enabled solution might hallucinate, or that thin underlying data might lead buyers astray, has kept them leery. WPP’s partnership with Givsly shows they’re shedding that caution.

Givsly’s system, built using the Claude LLM, uses a map of U.S. zip codes overlaid with third-party data drawn from sources like national census and election results. It also uses data gleaned from non-profit organizations revealing how many donations were made in a given zip code, how frequently they were made, and to what type of charity or non-profit. In basic terms, it maps a proxy for disposable income and social values against a demographic map of the U.S. 

According to Givsly CEO and founder Chad Hickey, the system uses that database to create audience segments containing demographic, behavioral and ethical reference points (for example, Californian women interested in wellness products), in response to advertiser briefs.

WPP Media added first- and second-party sources it possessed into the mix, said Sadlak. Those synthetic audience profiles were then used in PMP deals for four brands on the WPP Media roster. The test campaigns ran on CTV and online video inventory. 

While the video completion rate test suggested the solution had enabled a greater level of targeting accuracy, Sadlak said it also granted WPP Media buyers the ability to pursue audiences based on social values.

Forrester analyst Jay Pattisall noted that the benefits of the system likely still rested on the quality and depth of data that WPP and its clients were able to bring to the Givsly lens. “Synthetic audiences are only as good as the data that compiles them,” he said in an email. “For WPP, the InfoSum capabilities are fundamental to matching first-party client data with third-party signals to enrich and model synthetic audiences to plan against.”

According to GWI data, 86% of consumers say shared values are important when choosing between brands. “Ethical drivers aren’t always present in a standard brand brief,” Sadlak noted. 

“[But] targeting someone based on open web interactions doesn’t seem very pointed… We’re attempting to move beyond the behavioral, the demographic, and truly start to connect with what our consumers value and care about,” she added.

Following the initial tests, she said WPP Media planned to make the solution available to clients across the U.S.

Though Givsly’s system doesn’t incorporate political donations data, Hickey said the company had experienced a rise in inquiries from political advertising clients. 

“We have talked to a lot of political agencies who are looking for, say, women in Wisconsin who may go to church but also support a woman’s right to choose,” he said. “[Those are] things that you don’t necessarily think of in the same context, and we can spit out an audience in 30 seconds.”