Inside the New York Times newsroom last year, editors started running story ideas past a panel of 50 audience members. The panel was on call 24/7, returned reactions in seconds, and never asked to be paid. None of them were real.
The Times has been quietly using AI-generated audiences — synthetic focus groups built from clusters of real reader behavior — to pressure-test headlines, story angles, and product features. They aren't the only ones. By early 2026, most of the largest agencies and a growing number of in-house insight teams are treating synthetic audiences as a standard early-stage tool.
The economics are hard to argue with. A traditional in-person focus group runs $15,000 to $30,000 per session. A synthetic equivalent costs 80–90% less and returns insights in minutes instead of weeks. For research teams under pressure to do more with less, the math is settled.
But the research isn't.
What the Math Misses
Synthetic focus groups are good at a specific thing: surfacing the obvious objections to a concept fast and cheap, before a brand commits real money. They're a useful filter. They're not a stand-in for understanding the people the brand is actually trying to reach.
The reason is structural. A synthetic respondent is a probability distribution dressed in a persona. It generates answers that statistically resemble what someone in a demographic cohort might say — based on the data the model was trained on. That data is rich for some segments and thin for others. It captures explicit attitudes well, and tacit signals — the pause, the wince, the inflection — almost not at all.
Recent studies of synthetic focus groups consistently flag the same gap: outputs reflect the assumptions baked into the inputs. Bias enters through proxies — ZIP codes, device types, browsing patterns — and through the model's own learned associations. Run the same prompt across demographic variants and the differences don't always reflect reality. Sometimes they reflect the stereotype the model has already absorbed.
The result is research that feels rigorous because it's structured, but encodes the same blind spots that produced the campaigns the research was meant to catch.
The Hybrid Story Is the Real Story
The category leaders in research aren't framing this as replacement. They're framing it as triage.
A 2026 trends report from Rival Group found that 42.75% of researchers describe themselves as "not excited" about synthetic respondents — and that's coming from a research industry that, by other measures, is rapidly adopting AI. 64.1% of researchers increased the number of AI tools they used in 2025. 53% now use AI regularly. The skepticism isn't anti-AI. It's specifically skeptical of synthetic respondents as a complete answer.
The pattern that's emerging looks like this: synthetic audiences for early-stage hypothesis testing, real customers for the decisions that actually matter. The synthetic panel narrows the field. The human research validates what the model can't see — the parts of audience experience that don't compress neatly into training data.
Newer entrants in the space are explicitly building toward this hybrid model. Strella, an AI-powered qualitative platform backed by Bessemer, is positioning around scaling human research rather than replacing it. HBR's April 2026 analysis on AI in qualitative research called out the same pattern — the gains are coming from AI augmenting human-led research, not standing in for it.
Our Take
Synthetic focus groups didn't kill traditional research. They made the limitations of traditional research harder to ignore — and then proposed a faster, cheaper version of those same limitations. The brands that handle this well will use synthetic audiences for what they're good at: narrowing the option space, pressure-testing assumptions, generating cheap input for early-stage decisions. They'll reserve real audience evaluation for the decisions where the cost of being wrong actually matters. The mistake isn't using AI in research. The mistake is letting structured-looking output replace the evaluation step that catches what the model can't see.
How to Use Synthetic Audiences Without Inheriting Their Blind Spots
For marketing and insights teams, the question isn't whether to use synthetic audiences. It's how to use them without inheriting the assumptions baked into the model.
A few patterns separate teams getting real value from teams generating false confidence.
The first is treating synthetic outputs as hypothesis, not evidence. A synthetic audience that says urban Gen Z women would respond positively to a particular ad concept is generating a testable claim — not a verified one. The teams using this well take that claim and validate it with real audience signal before scaling spend.
The second is bias-testing the synthetic audience itself. Run the same prompt across demographic permutations. Look for stereotype-shaped outputs — answers that conform to what a model has been trained on rather than what audiences actually believe. If the synthetic women all want to talk about wellness and the synthetic men all want to talk about performance, the model is confessing where its training data ends and its assumptions begin.
The third — and most overlooked — is asking what the synthetic audience can't tell you. Models trained on aggregate behavior data are weak at predicting what's emotionally activating, what's culturally specific, what's about to shift. The richest signal is often the one that shows up nowhere in the training set, because nothing like it has happened yet.
The Decision That Matters
The brands that win the next two years of audience research won't be the ones with the best synthetic panels or the loudest skepticism of them. They'll be the ones who can tell the difference between a research method and a research answer.
Synthetic audiences are a method. They generate fast, cheap, structured input that's useful for narrowing options. They are not a replacement for evaluating creative against the actual audiences you serve, which still requires the kind of demographic-level resolution most research methods strip away.
According to XStereotype data, 84% of creative decisions still rely on intuition over data. The promise of synthetic focus groups is to close that gap. The risk is that they do the opposite — adding a layer of structured-looking output that reinforces decisions the team was already going to make, but with a synthetic respondent's name attached.
As we wrote about the average audience problem, aggregate scoring hides the segment-level signals that determine whether creative actually works. Synthetic focus groups, used carelessly, can make this worse — generating new aggregates from training data that itself has been aggregated and abstracted from the real audiences a brand needs to reach.
Used well, they're a useful tool in a hybrid research stack. Used as a substitute for actual audience evaluation, they generate confidence without resolution.
There's a reason XStereotype's models look different from a synthetic stack. The dataset was seeded by real people — actual audiences responding to actual creative — providing the kind of tacit signal that training data scraped from public behavior can't reproduce. That foundation doesn't replace human research; it means the model's outputs are traceable back to real audience response, not to assumptions about what audiences are likely to say. The methodology question isn't just whether the data is fast and cheap. It's where the signal came from.
XRay scores creative against those real-audience-grounded signals — not against synthetic personas, but against the demographic-level resolution that catches what a blended panel never could. See how it works.