Digital Twins: The Synthetic Data That’s Actually Going to Stick in Market Research

Synthetic data is one of the most contested topics in market research right now. Some of what's being marketed under that banner deserves the skepticism it's getting. Boosting thin survey samples with model-generated responses, for example, is a shortcut that solves a sampling problem by quietly inventing the very signal you were trying to measure. Researchers are starting to push back on this, and with good reason.

But 'synthetic data' is a broad, complicated category, and not everything underneath it is the same thing. There's one application I think will outlast the hype cycle and become a permanent fixture in market research: digital twins of real customers. Not generic AI personas. Not LLM-generated archetypes. Digital twins built from real, identified humans who agreed to participate.

Read this post as a wakeup call: market researchers are ideally placed to lead this revolution. But if we don't, other disciplines will, eroding our standing.

Defining digital twins

The term originated in engineering. NASA used early versions in the 1960s to mirror physical spacecraft on the ground so engineers could troubleshoot Apollo 13 from a console in Houston. The modern industrial version is everywhere now: Formula One teams run thousands of pit-stop simulations on digital twins of their cars before a single tire hits the track. Manufacturers run digital twins of factories and refineries to test changes before they're built. The pattern is the same in every case: a high-fidelity virtual replica of something real, kept in sync with that real thing, that you can interrogate and experiment on without disturbing the original.

In market research, a digital twin is a virtual replica of an individual human respondent, built from that person's own words, behaviors, decision history, and preferences: accessible on demand, and capable of responding to new questions in a way that reflects how that specific person would actually think, not how an 'average' version of their demographic would.

“Typical AI personas represent an average, ironing out strong opinions and regressing to the mean, which removes the opportunity to truly understand variance, polarization, and niche appeal. Digital twins circumvent the tyranny of averages.”

Hadley Edwards, Head of Market Research & Insights, Google Cloud Marketing

Speed, not cost

Global market research spending is roughly $140 billion a year. But cost isn't the real problem. What is real is the time problem.

The Economist Intelligence Unit surveyed 1,135 executives: 58% said they went with their gut on their last big decision. And these aren't small calls: nearly one in three of those decisions was valued at $1 billion or more.

This is where Colin Powell's 40-70 rule comes in. Powell's formula was P = 40 to 70, where P is the probability of being right, and the numbers are the percentage of information acquired. Less than 40%, you're shooting from the hip. More than 70%, the moment has passed and someone else has made the call for you. The sweet spot is somewhere between informed enough to be useful and fast enough to still matter.

The relevant question isn't 'is a digital twin as good as a 4-week survey?' The relevant question is 'is a digital twin better than the gut-only decision the team was going to make on Friday?'

When the alternative is zero data, a 70-85% confidence answer delivered in 15 minutes is a breakthrough.

Building quality twins

We believe purpose-built digital twins, developed from scratch with the contributor and designed specifically to capture how that person makes decisions, are the highest-fidelity option available. Research from Stanford and Columbia points the same direction. When you build a twin from a substantive narrative interview, you get connective tissue: the stories, contradictions, decision logic, and category-specific behaviors that a model needs to respond as that person, not as a stereotype of that person.

By contrast, twins built from existing survey data are extremely limited, and in practice they're closer to pure LLM-generated personas than to actual digital twins. Loading your brand tracker into a platform doesn't give the model the deeper layer required: the needs, motivations, pain points, and category behavior that drive decisions. The data wasn't designed for that purpose, so the resulting 'twin' is mostly the LLM filling in gaps from its training data, which is exactly the failure mode researchers are right to worry about.

If the source data wasn't built to capture how a human makes decisions, the twin isn't going to be able to model how that human makes decisions.

Transparency and accountability

Researchers are rightly wary of black-box AI products. A lot of digital twin platforms today use a ChatGPT-style interface: you ask a question, you get an answer, and you have no idea what's under the hood or where the answer came from. That's not research-grade. That's a magic eight ball with a nicer UI.

A more responsible approach treats the digital twin session like a qualitative read: you see the transcript, you see the citations, you see which contributor said what, and you can audit the chain from contributor narrative to model output. This is the approach taken by the start-up Focus (currently in closed beta): a focus-group-style interface where the researcher reads the conversation, not just the summary.

Transparency and citations, on top of purpose-built twins, is the version of this technology that belongs in our toolkit. The black-box version is the version that gives the whole category a bad name.

The three-year outlook

Here's my prediction: within 3 years, every enterprise company will have an always-on layer of virtual customers they can query for on-demand insights. Not as a replacement for traditional research, but as a permanent operating layer on top of it. The way dashboards became permanent in the early 2010s, on-demand customer intelligence will become permanent in the late 2020s.

The question for us as market researchers isn't whether this happens. It's who gets to define what good looks like when it does. If we don't lead this, other parts of the organization will pre-empt us. Marketing, strategic planning, UX, and product all have their own AI initiatives now, and many of them touch customer simulation in some form. None of those functions have what we have: a discipline built around understanding when data is trustworthy, how to scope a research question, how to read variance, how to distinguish signal from artifact, and how to put findings in the context of established methods.

That's the unique value we bring. Not the methods themselves, but the judgment about when and how to apply them. If we sit out the digital twin moment and wait for it to come to us, we'll find that adjacent functions have built their own versions without that judgment baked in. Research teams should get involved now. Partner with sales, marketing, strategic planning, UX. Set the standards. Define what trustworthy customer simulation looks like, because that question is going to get answered with or without us.

This category will not wait for researchers to approve it. The useful question is how much of our discipline we build into it now.

First published on Vital Findings, 25 May 2026.