Synthetic user research is the practice of answering a research question by simulating a population of users with AI agents instead of recruiting people. You describe the audience, put a real decision in front of it, and read how that population responds, usually within an hour rather than over several weeks.
It goes by several names. Synthetic users, synthetic respondents, simulated users, AI personas, and digital twins for research all describe versions of the same idea, and the differences between those terms matter less than one property they do not share: whether the agents are grounded in evidence about your actual users, or generated from a prompt.
That distinction is the whole subject. Everything below is downstream of it.
How a synthetic study actually runs
Four steps, in order. Teams that skip the first and last are the ones who conclude the category does not work.
1. Define the population. Not "our users" but a specification: which segments, what sizes, which attributes matter for this decision. Good tools make you say what you do not know here rather than filling it in silently.
2. Ground the agents. This is where fidelity is won or lost. Agents built from demographics alone are guessing from the average of the internet. Agents built from your interview transcripts, support tickets, survey responses, and event trails are modelling your market. We wrote the practical version of this in grounding synthetic panels.
3. Run a forced choice. "Do you like this?" is not a decision, and a panel asked that question will be agreeable. "This plan or that one, upgrade now or stay, variant A or the status quo" is a decision. Most weak studies were weak before the first agent ran.
4. Score it against reality. A simulation nobody checks is a story with charts. When the change ships, the outcome should grade the run that predicted it.
What the evidence actually supports
The category deserves neither the dismissal nor the hype it usually gets, so here is the state of the published record.
In Nature, GPT-4 simulations predicted the results of 70 preregistered experiments at r = 0.85, and reached r = 0.90 on unpublished studies the model could not have memorised. That is a real result and we wrote about what it does and does not establish. It measures models against nationally representative survey experiments. It does not measure them against your customers deciding whether to renew.
The grounding effect is measured too. In Park et al. (2024), agents built from interviews reached roughly 83 to 86 percent of the human test-retest ceiling on core survey instruments. Agents built from demographic attributes alone reached roughly 74 percent. That gap is not a model-quality gap that a better model closes. It is an evidence gap.
Read together, those two findings say something specific: the ceiling is high enough to be useful, and where you land under it is mostly determined by what you fed the agents.
What synthetic research is not
Four things it gets confused with, and the actual difference in each case.
| It is not | Because |
|---|---|
| A survey tool | Nobody is asked anything. There are no respondents to recruit, screen, or incentivise, and no fielding window. |
| A research panel | A panel is a roster of real humans. A synthetic population is generated per study and can model segments you have no access to. |
| A persona document | A persona is a cached summary of research that already happened. An agent is a view over evidence you still hold, so new questions hit the evidence rather than the cache. |
| A replacement for talking to users | It is the fast first pass. The hybrid workflow is how the two fit together. |
Where it works, and where it does not
The honest boundary matters more than the pitch, because reaching for the instrument on a question it cannot answer is the fastest way to lose trust in it internally.
It works well for screening many options cheaply, pressure-testing pricing and packaging, comparing design variants before build, finding the objection you did not anticipate, and reaching segments that are slow or expensive to recruit.
It works badly for genuinely novel interaction patterns with no behavioural prior to calibrate against, emotionally loaded decisions involving health or money under stress where affective fidelity is least validated, and any irreversible commitment where you need a human in the room before you sign.
There is also a failure mode that has nothing to do with the technology. If you have no transcripts, no tickets, and no behavioural data for a population, you do not have a synthetic panel. You have a prompt with a costume on. Collect evidence first.
How we approach it
We build Sentia around the two steps most implementations skip, because they are the two that decide whether the output is worth acting on.
Grounding comes from sources you already own, with provenance attached. Every attribute an agent carries traces back to a citation, and a field with no evidence behind it is surfaced as a gap rather than filled with a plausible default. Population structure is sampled from published census microdata and WorldPop demographics, so a panel reflects the real shape of a market instead of one flattened persona.
Scoring closes the loop. Our SDK emits an exposure event when a change reaches a user and an outcome event when they act, which ties the shipped result back to the simulation that predicted it. That produces a hit rate you can look at rather than an accuracy claim you have to believe. It is the difference between a vendor asserting fidelity and a system measuring it.
The third thing is newer and matters more every month: the whole platform is reachable by an agent. The same capability plane backs the web app, a REST API, the sentia CLI, and an MCP server, so a coding agent in your editor can build a population and run a study without a human opening a dashboard. If your team already works that way, research should not be the one workflow that forces a context switch.
Start here
If you are evaluating the category, the fastest way to form a real opinion is to read one report end to end rather than a feature list. We walk through a complete one in anatomy of a simulation report, and there is a day-by-day path from empty workspace to a panel you have personally backtested in your first week with a synthetic panel.
If you are further along and comparing tools, we wrote the questions we would ask in how to evaluate a synthetic user research platform.