WPP tests charity-driven synthetic audiences for programmatic, sees 2% engagement lift

For media buyers, the line between planning and bidding has long been drawn with synthetic audiences on the safe side. Now WPP's experiment suggests that AI-generated donor profiles can actually improve programmatic performance, opening a new chapter in cookieless targeting.

WPP Media, the media buying arm of British holding company WPP, has taken a notable step by quietly testing synthetic audiences—built in part on charity donation data—for live programmatic media buying in the United States. The move, executed through a partnership with 2019-founded tech firm Givsly, challenges a long-held industry taboo against using AI-generated personas in real-time bidding, a precaution fueled by fears of LLM hallucination and data thinness. Early experiments across four beauty and fashion campaigns delivered a 2% lift in video completion rates, a seemingly modest but significant signal that accurate targeting can be enhanced by synthetic profiles. This development comes as agencies grapple with the deprecation of third-party cookies and seek new, privacy-compliant methods to maintain reach and relevance.

“

The 2% lift in video completion rates, while early, hints that the synthetic audiences are indeed finding more receptive viewers.

Synthetic audiences—sometimes labeled digital twins—are not new to advertising. Agencies like Code & Theory, Marcus Thomas, and Dentsu (which partnered with Evidenza) have deployed AI-created cohorts for media planning and focus-group-like simulations. Publishers such as The Times and News UK have also built synthetic personas from first-party data. Yet, until now, none had ventured to use these constructs to inform actual bid decisions in programmatic pipes. The hesitation was rooted in the very nature of generative AI: models like the ones underpinning these personas can confabulate, producing cohorts that don’t map to real-world behaviors, potentially leading to wasted ad dollars. WPP’s test suggests that barrier is beginning to fall.

Givsly’s approach leverages Anthropic’s Claude large language model, overlaying zip-code-level maps with third-party sources including U.S. Census data, election results, and aggregated donation patterns from nonprofit organizations. The donation data is especially intriguing: instead of relying on individual donor records, Givsly uses frequency and volume metrics per geography to infer philanthropic propensity. This creates a behavioral signal that may correlate with brand affinity, enabling advertisers to reach audiences more likely to engage with cause-related or value-oriented messaging. For WPP, the goal is “discoverability and increased reach,” as group director of media delivery Tara Sadlak explained. The 2% lift in video completion rates, while early, hints that the synthetic audiences are indeed finding more receptive viewers.

The implications for programmatic advertising are profound. As identity-based targeting crumbles under regulatory and platform pressure, synthetic audiences crafted from public and aggregated data could become a mainstream substitute. They sidestep personally identifiable information, potentially aligning with ethical data use standards. However, the reliance on an LLM to assemble these profiles introduces clear risks. Hallucination is one; another is the thinness of the underlying signals. Charity donation frequency in a zip code may not robustly predict purchase intent for fast-moving consumer goods, for example. Moreover, the black-box nature of AI-generated segments makes it difficult for buyers to audit or explain why certain bids were placed, which could raise transparency concerns with clients. WPP’s disclosure of a 2% lift is a positive data point, but it comes from a narrow test in beauty and fashion verticals, and scaling across other categories or buying a broader inventory may yield different results.