DoReveal - Synthetic Data in Market Research: Promise, Pitfalls, and a Path Forward

There's a quiet revolution happening in how companies understand their customers, and it doesn't involve a single survey respondent. Synthetic data and digital twins are moving from academic curiosity to practical tool, landing on the desks of marketing strategists, insights professionals, and product teams worldwide. But with the excitement comes real questions: Is this just hype? Where does it actually work? And what happens when we trust it too much?

Here's a grounded look at what synthetic data means for market research today, and where it's headed.

What We're Actually Talking About

The term "synthetic data" gets thrown around loosely, so it's worth being precise. At its core, synthetic data is AI-generated information designed to mirror the statistical properties of real-world data, without exposing any actual individual's personal details. Its origins were largely about privacy: how do you extract useful patterns from sensitive data without the risk of exposure?

But the concept has evolved significantly. Today, practitioners draw a useful distinction between two related ideas:

Digital twins are algorithmic representations of a specific individual, designed to simulate how that person might respond to a given marketing message, product feature, or price point. Think of it as a behavioral model of you, built from everything known about your preferences and choices.
Digital personas take a broader view. Rather than modeling one person, they represent an entire market segment, a synthesized "typical buyer" whose responses can stand in for the group as a whole.

One helpful analogy: synthetic data is a bit like the synthesizer in music. It doesn't sound exactly like a piano, and it has its own quirks, but it made music more accessible, faster to produce, and opened creative possibilities that acoustic instruments couldn't offer. The same tradeoffs apply here.

The Real Value Proposition: Speed, Scale, and Access

The most obvious benefit is speed. Traditional market research, recruiting respondents, running surveys, analyzing results, takes time that fast-moving businesses often don't have. Synthetic data can compress that timeline dramatically, enabling teams to test concepts, explore scenarios, and gut-check decisions in hours rather than weeks.

But focusing only on efficiency misses the bigger picture. The more transformative opportunity is doing things that simply weren't possible before.

Consider a few examples:

Hard-to-reach audiences: In B2B markets, getting a statistically meaningful sample of, say, procurement directors is nearly impossible. You might cobble together five conversations over several weeks. Synthetic personas trained on relevant data can give you something far more robust to work with.
Legally restricted research: In regulated industries like pharmaceuticals, companies are sometimes prohibited from directly engaging certain populations. Synthetic approaches can provide a viable alternative where no alternative previously existed.
Sensitive topics: Some research questions are difficult to ask real people, whether due to emotional sensitivity, social desirability bias, or legal concern. Synthetic data can at least provide a starting point.
Scenario simulation: Perhaps most intriguingly, AI allows researchers to build simulated social systems, digital populations that can interact with each other and respond to policy changes, competitive moves, or product launches. It's the kind of controlled experiment that's simply impossible to run in the real world.

The key insight here: in many of these situations, the alternative to synthetic data isn't "better human data." It's no data at all. That fundamentally changes the calculus.

Where It Actually Works, and Where It Doesn't

The research community is starting to get clearer on when synthetic data is reliable and when it falls short.

On the positive side, studies have found that well-designed synthetic personas can replicate real survey results with impressive accuracy, particularly when estimating beliefs, attitudes, and stated preferences. In some controlled tests, synthetic responses have tracked closely to how real respondents answered the same questions days or weeks later.

But there are important limits.

Synthetic data is more rational than real people. AI-generated respondents tend to make consistent, logical choices. Real consumers don't. They're emotional, habitual, and situation-dependent in ways that AI struggles to capture. This matters a lot for categories where irrational or emotionally driven behavior is the norm, which, frankly, is most consumer categories.
AI is overly enthusiastic about novelty. When evaluating new products or concepts, synthetic data tends to overestimate interest and adoption intent, much like the classic problem with stated-preference surveys, where people say they'll buy something and then don't. If you're using synthetic data to forecast new product success, expect to need a correction factor for real-world inertia.
AI is AI-positive. This sounds obvious, but it has real implications: synthetic respondents tend to rate AI-assisted products and services more favorably than real consumers do. Research has consistently shown that human customers often downgrade their evaluation of something once they learn it was made with AI, even if it's objectively better. Synthetic data won't show you that effect.
Context transfer is fragile. In one well-documented example, a model trained on consumer data within one product category (say, smartphones) failed significantly when applied to a related but different category (tablets). The training data doesn't automatically generalize, you have to test it.

The Bias Problem Nobody Talks About Enough

Here's a finding that should give pause: research suggests that the more senior someone's position in an organization, the more likely they are to over-rely on AI outputs. And separately, the less someone understands generative AI's limitations, the more they tend to use it uncritically.

In other words, the people making the biggest decisions are often the least equipped to spot when AI is leading them astray.

This isn't an argument against synthetic data. It's an argument for cultivating real AI literacy at leadership levels, and building organizations that treat AI outputs as inputs to human judgment, not substitutes for it.

A practical approach: use AI to steelman your own arguments, not just to confirm them. Ask the tool to generate counterarguments, identify weaknesses in your reasoning, or stress-test your assumptions. Used this way, synthetic data becomes a tool for sharper thinking, not a shortcut around it.

A Framework for Getting It Right

Given all of the above, how should a team actually approach synthetic data responsibly? A few principles have emerged from both research and practice:

Start with the decision, not the data. What are you trying to find out, and why? The goal shapes everything, which approach makes sense, how much data quality you actually need, and what "good enough" looks like. Analytics should be decision-driven, not the other way around.
Assess contextual fit. Is the context you're working in similar enough to what the synthetic model was trained on? The further you stray from the training data's domain, the less you should trust the output.
Define validation metrics upfront. What would it mean for the synthetic data to be "good enough"? Set benchmarks before you run the analysis, not after. This prevents unconscious post-hoc rationalization.
Augment, don't replace. Even where synthetic data is working well, pairing it with real human input, even a small number of actual customer conversations, tends to anchor the outputs to reality in important ways. Think of it as calibration, not redundancy.
Build feedback loops. Track how your synthetic-data-informed decisions actually play out. Over time, this creates a learning system that improves both your use of AI and your judgment about when to use it.

The Human Expert Isn't Going Anywhere

One of the more persistent fears in the research industry is that synthetic data will eventually replace the need for human expertise. The evidence so far doesn't support that, but it does require reframing what that expertise is for.

The value of a skilled insights professional isn't in doing the laborious work of data collection and processing. It's in asking the right questions, interpreting nuanced findings, understanding what the numbers don't capture, and connecting insight to action. Those capabilities don't become less valuable when AI handles more of the upstream work. If anything, they become more valuable, because someone has to know how to direct the technology, evaluate its outputs critically, and translate them into decisions.

There's an old idea from competitive chess: a human player combined with a powerful computer can outperform either one alone. The same logic applies here. The researchers who will thrive aren't those who treat AI as a threat to resist or a magic solution to deploy uncritically, they're the ones who learn to work with it well.

What Comes Next

The field is still early. Most teams are experimenting, not deploying at scale. The statistical frameworks for combining human-provided and synthetic data are still being developed. And the underlying models themselves will continue to evolve, potentially in architectures quite different from today's transformer-based systems, which may unlock capabilities we can't yet anticipate.

What that means for practitioners: now is the time to experiment, not wait. Test synthetic approaches on decisions where the stakes are manageable. Compare results to what you'd have gotten from traditional methods. Build institutional knowledge about where it helps and where it doesn't. Don't anchor to one vendor or one tool, the landscape is changing fast, and early choices shouldn't lock you in.

And resist the urge to dismiss it based on a few disappointing early results. The technology being used today is not the technology you'll be using in three years. The teams who learn the most from careful experimentation now will be best positioned to take advantage of what's coming.

Synthetic data is not a silver bullet. It is not a replacement for human judgment, emotional intelligence, or the irreplaceable insight that comes from actually listening to real customers. But used thoughtfully, it's a genuinely powerful addition to the researcher's toolkit, one that can unlock decisions that were previously too slow, too expensive, or simply impossible to make well.

That's not hype. That's a new tool. Learn how to use it.

Synthetic Data in Market Research: Promise, Pitfalls, and a Path Forward