A synthetic user is an AI-generated persona that answers research questions in place of a real participant. The useful part of the definition is not “AI-generated”. It is what the persona was generated from, whether it can show why it answered as it did, and whether anyone measured how it compares with the people it stands in for. Those three things separate a research instrument from a chat with a system prompt.
The term arrived in 2023 with a startup’s name attached, spread through UX teams, and by 2025 the largest survey platform in the world was selling synthetic respondents. Behind one phrase now sit several very different systems. This is the plain version.
The definition
Nielsen Norman Group, which has spent two years testing these systems against real research, defines a synthetic user as “an AI-generated profile that attempts to mimic a user group, providing artificial research findings produced without studying real users”. Three parts matter. A profile: the persona has attributes, not just a name. A user group: it stands for a segment, not one person. Without studying real users: the findings come from the model, not from fieldwork.
That last part is the whole debate. A synthetic user’s answer is only as good as what the model knows about the group it is imitating, and how faithfully it reproduces the variation inside that group rather than a smoothed average of it. The evidence on accuracy turns on exactly this.
The words the market uses
| Term | Who says it | What it usually means |
|---|---|---|
| Synthetic users | UX and product research; also a vendor’s brand name | Personas for interviews and usability-style questions |
| Synthetic respondents, synthetic panels | Survey platforms and agencies | Simulated survey answers at scale, often calibrated on panel data |
| Synthetic personas, AI personas | Marketing and persona tools | Segment profiles you can question in natural language |
| Digital twins | Some vendors and academics | A model of a specific individual, often built from an interview with that person |
| Generative agents, persona simulation | Academia | Language-model agents conditioned on a persona, studied for how well they reproduce human responses |
The systems behind the words differ more than the words do. A “digital twin” built from a two-hour interview with one real person and a “synthetic user” generated from a one-paragraph brief are different instruments with different evidence behind them, and both get called synthetic research.
Four kinds of system
- Prompted personas. A language model is told to act as a segment and answers from its general training. Fast, cheap, and the kind the negative research was done on: sycophantic, flat, and unstable across prompt wording.
- Calibrated panels. Answers are generated at scale and tuned against real survey data. Qualtrics describes its Edge Audiences this way: researchers “ask questions of synthetic respondents designed to represent their target customers”, with the company claiming research time cut “from weeks to minutes” and costs “by as much as 70%”. The claims are the vendor’s.
- Prediction engines. Thousands of agents simulate a population to forecast reactions. TechCrunch reported in December 2025, citing sources, that Aaru, one such company, had raised a Series A of more than US$50 million led by Redpoint Ventures, with some of the equity priced at a US$1 billion valuation and the blended valuation below that.
- Grounded personas. The persona is built from real material, interview transcripts, briefs, documents, and constrained to it: it cites what it used and declines what it does not know. The academic version is PersonaCite, published in January 2026, which reframes personas as “evidence-bounded research instruments”. This is the kind Perplica builds.
What is inside a persona
Vendors rarely say. It matters, because a persona is only the attributes the system can act on. At the thin end there is a name, an age and a job title in a prompt. At the thick end there is a psychological profile on named frameworks, a life story, memories that persist between sessions, and attached source material with a declared level of fluency. Perplica’s personas carry a 16-dimension profile in seven categories, a narrative identity with beliefs and wounds, six memory types and a knowledge layer; the method page lists what each part changes in a reply.
A persona you cannot inspect is a persona you cannot trust. The first question to a vendor is not “how accurate” but “built from what”.
What they are used for
The honest use cases were set out by the same NN/g article that delivered the verdict: preparing for research studies with real users, treating whatever synthetic users say as hypotheses that need testing, and drafting proto-personas and proto journey maps to be revisited after real research. Commercial practice has added concept and message tests, pricing rehearsals and early screening of a wide option set before fieldwork. The pattern that is winning is sequential: synthetic first to narrow, human second to decide. Our pricing playbook is written for that order.
Where the definition stops
NN/g’s conclusion is unambiguous: “synthetic users cannot replace the depth and empathy gained from studying and speaking with real people.” The reasons they found are the reasons the whole category argues about: models want to please, and their personas “seem to care about everything”, a flat approximation of the many people they stand in for. The sycophancy note explains why that happens, and what counters it.
Where Perplica sits
Perplica is the fourth kind: grounded personas for market research. Personas are built from your interview transcripts, briefs and documents; every reply stores the reasoning behind it; trust starts neutral and is earned, so personas push back; and a study compiles into a report where every finding links to the exact turn that supports it. It is a rehearsal, not a replacement, for talking to real customers, and it says so on the label.
What is a synthetic user?
An AI-generated persona that answers research questions in place of a real participant. Nielsen Norman Group’s 2024 definition is the cleanest: an AI-generated profile that attempts to mimic a user group, providing artificial research findings produced without studying real users.
Are synthetic users the same as synthetic respondents?
In practice yes. UX research says synthetic users, survey and market research says synthetic respondents or synthetic panels, and academics say persona simulation or generative agents. The systems behind the words differ more than the words do.
Who sells synthetic users?
Pure-play startups such as Synthetic Users, Aaru and Evidenza, established platforms that added synthetic panels such as Qualtrics, and persona and UX tools. The market maps of 2026 group them into pure-play, hybrid and AI-enhanced traditional platforms.
Can synthetic users replace user research?
No, and the honest vendors do not claim it. Nielsen Norman Group’s verdict is that synthetic users cannot replace the depth and empathy gained from studying and speaking with real people. They are a rehearsal for real research, not a replacement.
What is persona simulation?
The academic term for the same idea: conditioning a language model on a detailed persona and letting it answer as that person. Perplica uses the term because it describes what the system does, and because a persona built from real material behaves differently from a one-line prompt.