Synthetic Audiences Are Reshaping Media Planning. The Industry Still Hasn’t Figured Out How to Trust Them.

The erosion of traditional targeting signals such as third-party cookies, device identifiers, and behavioral tracking has pushed the media planning industry to wonder if AI-generated consumer personas can stand in for the real thing.

Agencies are increasingly incorporating synthetic data stacks into their planning and creative processes, modeling audiences without relying on user-level tracking data. The advantages are straightforward; synthetic data stacks are positioned as privacy-safe, scalable, and available without the latency of traditional research. 

But the challenges are equally clear. As with all machine learning systems, synthetic models are only as good as their training data, and the difference between a well-constructed persona and an actual buying consumer can be wide.

Paul Frampton-Calero, CEO of Goodway Group, described the case for synthetic modeling in a recent piece for Campaign, arguing that it allows brands to understand a market at finer granularity than first-party data can typically deliver, without personally identifiable information (PII) and without the financial cost of traditional incrementality testing, which typically requires a brand to go dark in a market for six to eight weeks to isolate a causal effect. 

Whether the industry is ready to commit budgets against that promise is a different matter.

‍

A Useful Starting Point, Not a Finishing Line

Many practitioners closest to this work share a consistent perspective: synthetic data earns its place as a planning input, not a planning answer. 

“Synthetic data and AI-generated audiences shouldn’t be viewed as a silver bullet, but rather as a powerful addition to the media planning toolkit,” Natasha Tibbetts, SVP of integrated media at True Media, says. “AI personas offer a solid baseline, but they are inherently backward-looking or static unless continuously contextualized.” 

Real consumer behavior, she says, is shaped by seasonality, macroeconomic shifts, and cultural nuance that static models don’t capture. “Synthetic data gets us halfway there by generating informed hypotheses, but true confidence is only earned when those models are paired with deliberate test-and-learn frameworks in the wild.” 

To be sure, the erosion of targeting signals doesn’t mean buyers are suddenly planning blind, notes Jasmine Presson, chief strategy officer at Mediaplus North America. “Media has always included channels where direct targeting and measurement signals were limited, and we’ve built models and research approaches to account for that,” Presson says. 

What makes synthetic data worth attention, in Presson’s analysis, is its potential to connect fragmented signals across paid, owned, and earned channels and different data environments. That said, she advises testing. “We run the risk of garbage in, garbage out, which is why at Mediaplus, we look closely at how it was built and validate it against syndicated, first-party, campaign and business data.”

Brian Lerche, co-founder and head of AI, data, and media at Charts+Darts, draws a similar picture, one where synthetic data can do a lot of work, but needs to be realistically evaluated for its suitability to various tasks.

For instance, Charts+Darts uses AI agents to represent samples of up to 400 actual consumers, Lerche notes. Their activity is weighted to a target audience and modeled on real-world public data, with each agent’s persona informed by 1,200 possible combinations of psychographic and demographic attributes. “We move faster and more confidently because our synthetic audience research paints a broader picture of our target audience,” Lerche says. 

But he is explicit about the ceiling: “Confidence decreases if we start expecting synthetic data to predict media response, sales, or ROI.” 

Jono De Costa, head of media and performance at RP3 Agency, is more skeptical about where that ceiling is. 

“I am very skeptical of using synthetic data to confirm audience behavior or decide elements in a media plan,” De Costa says. “Until synthetic data and AI can replace common sense, the human expertise and decision-making still matter in media planning.” 

He identifies a structural problem with how most synthetic models are presented to buyers: “Most synthetic audiences promise that they use thousands or even millions of signals to ensure accuracy, but few offer out-of-the-box validation mechanisms.”

‍

The Validation Gap

The consistent thread across practitioners is that synthetic audiences produce actionable data, but that data still requires real-world confirmation. 

Lerche puts it this way: “Synthetic audiences let us turn assumptions into hypotheses.”

The mechanisms for confirmation remain underdeveloped, Tibbetts says, adding, “To commit real working dollars, brands need to see synthetic audiences validated against live campaign performance, downstream measurement metrics, and evolving market context. Until synthetic models demonstrate an ongoing ability to learn from real-time sales and response data, smart practitioners will treat them as directional inputs rather than definitive media plans.”

“The real test is whether synthetic audiences can consistently capture the complexity of people, not just approximate an average consumer,” Presson says. “As we get into category-specific beliefs and behaviors, the nuances matter.” 

Her prescription for building that confidence is methodical: repeated testing against syndicated, first-party, and behavioral data, tracked over time to enable meta-analysis across categories and use cases. “That’s how we move from validating a single model to understanding where synthetic data is consistently reliable.”

Mansoor Basha, chief technology officer at The Marketing Cloud, describes his firm’s approach as grounded in deterministic data from the outset. The company’s synthetic audience tool is built on digital marketing provider Stagwell’s ID graph, which uses attitudinal and behavioral attributes as the foundation before personas are modeled on top. “That grounding gives us confidence in their usefulness for creative exploration,” Basha says, “but it doesn’t mean we can assume they fully reflect real-world behavior.” 

He is pointed about the sequencing: “Confidence should grow with the evidence, not ahead of it.”

The validation question becomes most concrete when testing whether synthetic personas match what consumers actually buy. De Costa notes that ecommerce is currently the most tractable environment for that test, where loyalty data can be compared directly against synthetic model outputs. “However, testing remains challenging for products purchased offline or those with long sales cycles,” he says. “Without significant research investment, confirming the synthetic audience model’s accuracy is almost impossible.”

Lerche approaches the validation question from the model architecture side. “The key question isn’t whether the synthetic audience is accurate,” he says. “It’s whether the LLM harness or framework is malleable enough to capture human nuance, yet structured enough in its data practices to reduce hallucinations.” 

In his telling, a poorly constructed LLM foundation produces validation problems that have nothing to do with the quality of the underlying data. That’s why his firm has moved away from consumer archetypes toward what he calls truly synthetic consumer audiences built on statistically weighted samples.

‍

Where Real Purchase Data Fits

The practitioners who express the most confidence in synthetic audiences share a common approach: they ground their models in real behavioral data before generating personas. No one is generating personas and hoping they align with behavior afterward. 

Basha describes an “always-on panel of real, relevant consumers” feeding respondent-level data back into the model as the mechanism that keeps synthetic audiences tethered to current reality. Lerche describes layering real client sales data onto market hypotheses generated by the AI model as the special sauce that turns interesting output into actionable insight.

According to Presson, Mediaplus applies a testing methodology that models the purchase patterns that might be expected from the synthetic population. Then, those findings are benchmarked against syndicated purchase data, first-party transactions, and campaign response. “Over repeated tests, that lets us measure where synthetic audiences reliably reflect real consumer behavior and where they introduce bias or distortion,” she says.

That architecture — synthetic modeling validated against deterministic purchase behavior — points to where the category is heading, Presson suggests. 

Synthetic models that can be continuously updated against real-world transaction data are far more effective and trustworthy than static personas built once and applied until the next research cycle. The difference matters most at budget commitment time.

“The biggest hurdle isn’t generating the personas,” Tibbetts says. “It’s proving that synthetic intent translates into real conversion behavior.”