WPP Media launches programmatic CTV targeting using AI-generated synthetic audiences

This move signals a pivot in how agencies handle signal loss from cookie deprecation by using generative AI to create reliable bidding proxies. By integrating social-value data into programmatic layers like CTV, WPP is attempting to quantify 'ethical alignment' as a performance metric. For the streaming ecosystem, this shifts the focus from simple demographic targeting to high-intent, value-based segmentation. The 2% performance lift provides a concrete benchmark for other holding companies currently benchmarking AI datasets. Watch for whether DSPs begin offering standard 'values-based' synthetic tiers or if these remain proprietary agency-vendor integrations.

WPP’s push into synthetic targeting aligns with broader internal shifts toward agentic AI workflows. In February 2026, WPP Production launched a 'shaping performance' model, using synthetic personas to run focus-group simulations and catch cultural sensitivities before campaigns go live. Per internal reporting, this model was designed to combat 'creative fatigue' on digital platforms by shortening the loop between strategy and execution. The agency's midyear 2026 forecast, authored by Global President of Business Intelligence Kate Scott-Dawkins, notably projected that AI search ad revenue will reach $100 billion by the end of the decade, representing nearly 40% of the total search market.

Concurrent with WPP's moves, the wider programmatic landscape is moving away from identity-first logic. Per WPP Media's May 2026 insights, the retirement of legacy systems like Microsoft Invest highlights a transition toward conversational and agentic buy-side tools. In this environment, synthetic audiences derived from non-sensitive proxies like donor behavior help sidestep privacy regulations while maintaining reach. Other holding companies are following similar paths; for instance, Dentsu established a deal with Evidenza in late 2025 to scale synthetic modeling across its planning departments. However, recent academic preprints from the Slovak University of Technology have cautioned that while LLM-generated personas are grammatically rich and efficient, they still risk institutionalizing biases or repeating stereotypes if the underlying data lacks sufficient demographic variability.