Are synthetic audiences a reliable source of audience insight? They are reliable for speed, hypotheses and scenario testing. They are not a substitute for real people. The agencies getting value from them treat every synthetic finding as a question to check, not an answer to present.
Adoption is moving fast. Digiday reports that agencies and brands are taking synthetic audiences out of the lab and into everyday planning. Aveeno Body cut strategy-brief writing from about 30 days to three using Vurvey's synthetic research, and Goodway Group added Arima's "synthetic population" of US, UK and German consumers to its media planning. In France, Stratégies reports that Patrick van Bloeme, after 30 years running research firms including Harris Interactive and Toluna, has launched Limpid, which runs qualitative studies on synthetic personas with no human respondents, in a few hours. Tellingly, Limpid "ne revendique pas de représentativité statistique": it makes no claim to statistical representativeness.
What are synthetic audiences, and why are agencies adopting them now?
A synthetic audience is an AI-generated set of personas, or a simulated population, designed to predict how real groups of people might think or respond. Vendors build them from different raw materials. Digiday describes Arima drawing on census data, US Bureau of Labor Statistics figures, Nielsen survey data and bespoke surveys, while Vurvey combines public data with qualitative video surveys of real consumers.
The use cases are broad: simulated incrementality tests, strategy briefs, product development and virtual focus groups. The big groups are in. According to Digiday, Havas backed Vurvey in an $8.5 million funding round, and Dentsu has worked with Evidenza to add synthetic data to media planning for more than a year.
The reason is time. Fieldwork takes weeks. A synthetic panel answers before lunch. For an agency with a pitch on Friday, that is hard to resist.
So are synthetic audiences reliable?
It depends on the question you ask them, and even the people building them are candid about the limits. Vurvey's Chad Reynolds told Digiday that "LLMs give you the top of the bell curve, the most common answer", which is why his company adds real-consumer inputs. TransUnion's Julie Clark warned against personas built on weak data: "You have to start with the right foundation." Code & Theory's Stef Hoffman put it most simply: "It informs judgment, but it doesn't replace judgment."
The bell-curve point should worry agencies most. Clients do not pay for the most common answer. They pay for the non-obvious tension, the minority behaviour that becomes a trend, the thing the whole category has missed. A tool that gravitates to the average is useful for ruling ideas out. It is risky for finding the idea.
Notice, too, who is doing the reporting. Most of the success stories come from vendors and their clients. That is not a reason to dismiss them. It is a reason to check.
Why it keeps happening: the data illusion effect
Synthetic research is persuasive for a reason that has little to do with accuracy. Call it the data illusion effect: the more an answer looks like data, the more we trust it, regardless of how the data was made. It is a close cousin of what psychologists call the illusion of validity, our habit of feeling confident in judgements that have little predictive power.
Synthetic outputs arrive fluent, quantified and neatly labelled. Imagine a simulated panel of "eco-conscious Gen Z shoppers" reporting that most would pay more for refillable packaging. It reads like a finding. It is actually a model's estimate of what such people might say. Put it on a slide with a percentage and it becomes a fact by Tuesday.
The insight underneath is one every planner should pin above their desk: data reveals patterns, but it cannot define humanity. Patterns are useful. They are not people.
Which audience data source should I use for what?
No single source is reliable on its own. Reliability comes from combining sources whose weaknesses do not overlap.
| Source | Best for | Weak at | How to check it |
|---|---|---|---|
| Synthetic audiences | Fast hypotheses, scenario and concept reactions, brief drafts | Surprises, minority views, statistical claims | Trace to observed behaviour; run a small live test |
| Surveys and panels | Stated attitudes at scale | What people actually do; speed | Compare with behavioural data |
| Published content and social conversation | What people say unprompted, in their own words; competitor activity | People who rarely post | Read the posts behind each conclusion |
| First-party and campaign data | What your customers actually did | People who are not customers yet | Incrementality tests |
| Qualitative interviews | Motivation, depth, the "why" | Scale and speed | Triangulate with the sources above |
How do agencies keep synthetic audiences honest?
- Ask what it is built on. Which sources, how recent, how often refreshed. A vendor that cannot answer is asking you to trust a black box with your client's money.
- Use synthetics to generate hypotheses, not conclusions. Let the panel narrow ten territories to three. Do not let it pick the winner.
- Ground each hypothesis in observed behaviour. Check whether real people are actually saying, sharing or doing what the panel predicts. If you cannot find the evidence, treat the finding as unproven.
- Validate the decisions that matter in market. Goodway, Digiday reports, will use Arima's tool to run incrementality tests in a simulated environment and then check the results with small-scale media activations. That is the right instinct.
- Label every insight in the deck. Mark each one as synthetic, observed or measured. Clients deserve to know which is which, and so does your own strategy team.
- Appoint a professional sceptic. Digiday notes that teams "had to maintain a professional distrust of their findings". Make it someone's job to ask: how would we know if this was wrong?
Where does real published content fit?
Published content is the anchor. What audiences and competitors actually post is unprompted, specific and dated, so it can be checked. It has its own blind spot, because it over-represents people who post, which is why it works best alongside the other sources in the table.
This is the principle behind Defy Monitor: every AI conclusion traces back to the real published posts behind it, so strategy is reverse-engineered from what audiences and competitors actually publish rather than from surveys. Personas are clustered from real conversations, which gives synthetic findings something solid to be checked against. You can see how that sits within the Defy platform.
Synthetic audiences are here to stay, and they will make agencies faster. The agencies that keep their credibility will be the ones that can show, for every claim, the real people behind it.
