FOUNDATIONS
Synthetic user research is a research method where AI-generated personas, grounded in real-world evidence about a target audience, are interviewed to surface how that audience would respond to concepts, problems, prices, and product decisions. It complements traditional research rather than replacing it, and works best for early-stage discovery, concept testing, and assumption validation where speed and breadth matter.
The category has emerged over the past three years as research demand has outpaced what traditional methods can deliver. Product, brand, and innovation teams ask more questions every quarter. Recruitment timelines, panel costs, and researcher capacity haven’t kept up. Synthetic research closes that gap by letting teams run dozens of studies in the time a single traditional study takes, at a fraction of the cost, with research-grade rigor where it’s been done well.
This page covers what synthetic research actually is, how it works mechanically, where it fits in a research operation, where it doesn’t, and how to tell rigorous synthetic research apart from AI improvisation dressed up to look like research.
Synthetic user research follows a consistent pipeline. The details vary by platform, but the structure is the same: evidence retrieval, persona generation, interview, synthesis. The quality of the output depends on how each stage is implemented.
Evidence retrieval is the first and most important stage. A rigorous synthetic research platform starts by gathering real evidence about the target audience: published research, customer interview transcripts, survey data, market reports, public web sources, and any first-party data the team provides. This evidence is parsed, indexed, and tagged with provenance so every downstream claim can be traced back to its source. Platforms that skip this step and generate personas from a prompt alone produce plausible-sounding but evidence-free output. The difference matters, and it shows up in the answers.
Persona generationis where the platform synthesizes the retrieved evidence into individuals. A good platform doesn’t generate one generic synthetic user. It generates a population, segmented by the dimensions that matter for the research question (industry, role, behavior pattern, attitude cluster), with personality traits and cognitive biases assigned from validated distributions rather than guessed at. Candor uses the Big Five (OCEAN) personality model sampled from peer-reviewed population data by region and occupation, and assigns cognitive biases at research-backed intensities rather than as binary labels.
Interview is where the platform earns or loses credibility. The interview can be a live conversation (you type a question, the persona responds) or an automated session where an AI interviewer drives the conversation against a research guide. Either way, the persona has to respond consistently with its established profile, history, and the evidence that grounds it. Inconsistency is the failure mode that ruins synthetic research, so rigorous platforms run a critic agent that validates every response against the persona’s established beliefs before delivery.
Synthesisturns dozens of interviews into structured findings: signals extracted, themes clustered, archetypes broken out, tensions identified, opportunities framed. The synthesis stage is where speed compounds. A team that’s just spent three weeks recruiting and interviewing 15 real customers still has to analyze the data. A team running synthetic research with Candor gets a structured report alongside the interviews, ready for the same kind of pattern-finding work, faster.
Throughout the pipeline, provenance tagging records where every claim came from. Was a behavior grounded in your uploaded research? Inferred from a validated behavioral pattern? Calibrated from a peer-reviewed distribution? Sampled at random because no evidence was available? The provenance is visible on every attribute. That visibility is what makes synthetic research auditable in a way that traditional research often isn’t.
The terms are used interchangeably in marketing copy, but the underlying methods can be very different. The distinction matters for buyers.
An AI personain the loosest sense is any AI-generated character meant to represent a customer segment. A prompt to a general-purpose AI model asking it to roleplay a 35-year-old marketing manager at a mid-market SaaS company produces an AI persona. So does a persona-generator tool that asks a few questions about your business and outputs a one-page profile. These are useful for brainstorming and stakeholder alignment. They are not research instruments. The AI improvises the persona’s attributes from training data, without any explicit grounding in real-world evidence about your audience, without a personality model, without bias calibration, without consistency enforcement, and without memory of prior interactions.
A synthetic userin the rigorous sense is a research-grade construct: an evidence-grounded persona with a calibrated personality profile, bias assignments at research-backed intensities, persistent memory across interview sessions within a study, and consistency validation on every interaction. The persona’s attributes have provenance. The personality and bias parameters come from peer-reviewed distributions. The interview behavior is checked against an established profile before each response. This is what makes synthetic research auditable. It’s also what separates synthetic research from AI improvisation.
The practical test is this: ask a vendor where their personas get their evidence, what personality and bias model they use, and how they enforce consistency across interview sessions. If the answers are “the AI figures it out“ or ”we use the latest models“ or ”we generate them from your description,“ it’s an AI persona. If the answers are specific (data sources, citation methods, personality framework, bias library, validation method), it’s synthetic user research.
In practice, you can already see this distinction across the vendor landscape. Platforms like Synthetic Users, an established player whose science hub cites peer-reviewed parity studies, and Candor sit at the rigorous end: evidence retrieval, calibrated personality models, bias libraries, critic validation, persistent memory. Many AI-persona generators sit at the lighter end, producing one-page profiles from a few inputs without any of those discipline checks. There’s a spectrum in between, and the gap between the two ends is wider than the category-level marketing suggests. Both have legitimate uses (brainstorming personas is not the same job as research-grade synthetic interviews), but they shouldn’t be confused.
The category is young enough that this distinction isn’t yet obvious in vendor marketing. Buyers who don’t know to ask end up with AI personas dressed up as research instruments, and the output reflects that.
Synthetic research is most valuable when speed, breadth, or access constraints make traditional methods impractical. Five situations where it’s the clearly right tool:
Early-stage discovery. Before you commit to a research budget and a recruitment cycle, you want to know what shape the problem is. Synthetic research lets you talk to a population of personas about their pains, current behaviors, workarounds, and constraints in hours. The output sharpens what to ask your real customers about, and which segments are worth recruiting. See problem discovery.
Concept testing at the early gates.Most concepts that go through traditional concept testing don’t make it to launch. The economics of running every concept through a $30K-plus panel round are bad. Synthetic concept testing screens concepts in hours, so you only spend traditional panel money on the survivors. See concept testing. A single synthetic concept test (audience definition through to synthesis report) typically completes in one to two hours of platform time. Test-and-iterate cycles can span days, but that’s intentional iteration, not waiting on the tool.
Price testing without anchoring panel respondents. Price is one of the most-asked questions and one of the most-distorted answers in research. Synthetic price testing across a variance of personas surfaces price-sensitivity patterns and acceptable-tier signal faster than real-panel price research, and without the contamination that comes from panelists who’ve seen too many price tests. See price testing.
Assumption validation before commitment. Every product plan and pitch is built on assumptions. Some are true, some are nice-to-believe, some are wrong. Synthetic research can stress-test each assumption against a population of personas and return per-assumption verdicts with evidence. The cost of being wrong is much higher than the cost of checking. See assumption validation.
Research in populations that are slow or impossible to recruit. Regulated industries (healthcare, financial services, defense) face recruitment friction that can stretch a study to three months. Synthetic respondents grounded in published patient or member research bypass the recruitment timing without bypassing the methodology rigor.
The connective tissue across these situations is the same: traditional research is the right tool when the question warrants it, but a lot of research questions don’t warrant a $30K, six-week panel cycle. Synthetic research fits the questions that need answering in hours.
Synthetic research is not appropriate for every research question. The failure mode is using it where it doesn’t belong and then concluding the method is broken when the answers are weak. Five situations where traditional research wins:
Real-behavior measurement over time. Synthetic research can modeldrift by re-grounding personas with updated evidence or mutating their memories, letting you study how attitudes and perceptions plausibly shift. What it can’t measure is what real customers actually did: who churned, who bought, which features got adopted or abandoned, how market share actually moved. That ground truth comes from real customer analytics, real product telemetry, and real longitudinal panels. The right disposition is to use synthetic research to hypothesize drift and direct what to look for in your real-behavior data, not to replace the real-behavior data.
In-market behavior tracking.Whether customers actually buy, return, retain, or churn is a question about real behavior in a real market. Synthetic research can predict directional response to messaging or pricing, but it can’t replace the ground truth of in-market data.
Regulated claim validation. If a claim is going on a label, in a clinical document, or in a regulatory submission, it needs to be substantiated against real respondent data with documented methodology. Synthetic research is appropriate for exploring concepts, not for substantiating regulated claims.
Cultural and emotional nuance that depends on lived experience. Synthetic personas are grounded in evidence about populations. They are not grounded in the unspoken, the felt, the culturally specific moment that only a real person from a specific community can articulate. When the question depends on that depth, real interviews are non-negotiable.
Anything where the cost of being wrong demands real validation. Bet-the-company strategy decisions, regulated launches, high-investment product pivots: validate against real customers before committing. Synthetic research is for the decisions where the value of fast directional signal exceeds the cost of being a little wrong.
The honest answer is that synthetic research is a complement to real research, not a replacement. The teams getting the most value from it use synthetic research to make their real research sharper and more focused, not to skip it.
Reliability in synthetic research breaks down into four discipline checks. A platform that does all four can produce research-grade output. A platform that does fewer is producing AI improvisation.
Evidence grounding.Personas should be generated from real evidence, not from a prompt alone. The platform should retrieve and index documents, search relevant public sources, and explicitly cite which evidence informed which attribute. Without this, the platform’s output is the AI’s best guess about your audience, which is no better than asking a smart generalist who’s never met your customers.
Calibrated psychology.Personality traits and cognitive biases should be sampled from peer-reviewed distributions, not assigned by the AI’s intuition. The OCEAN model (also called the Big Five: Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism) is the established psychometric framework, and population distributions are available by region, occupation, and age in published research. Cognitive biases should be assigned at calibrated intensities, not as binary labels, and the biases included should be the ones with the strongest empirical support. Candor’s bias library is grounded in the cognitive bias research catalog.
Consistency enforcement.Personas should respond consistently with their established profile, history, and evidence. Every response should be validated against the persona’s prior statements and beliefs before delivery. Without this, a persona’s answer to question 7 can contradict its answer to question 2 and the team only notices when a stakeholder reads the transcript and asks why the person who said they hate subscriptions just signed up for one. Critic validation catches inconsistencies before they reach you.
Persistent memory. Personas should remember across sessions within a study. If you interview the same persona twice, the second interview should reference the first: stories told, opinions shared, decisions described. Without memory, personas reset to baseline every conversation, and you lose the longitudinal depth that makes synthetic research useful for ongoing exploration. See how persona memory works.
There’s also a broader rigor check at the level of the research industry. A growing number of peer-reviewed studies compare synthetic and real interviews across multiple methodologies and report convergence at certain thresholds. Synthetic Users calls this synthetic-organic parity, a useful piece of category vocabulary that captures the practical question of whether synthetic interviews converge with real ones for a given research method. The literature is young and the parity is conditional, not universal: it holds for some question types and breaks down for others. The right disposition is to treat synthetic research as evidence-generating, not evidence-replacing, and to validate findings against real-customer signal wherever the stakes warrant it.
The questions where synthetic research is strongest map cleanly to the use cases above. In rough order of demonstrated reliability:
The questions where synthetic research is weakest are equally clear: real-behavior measurement over time, in-market behavior tracking, regulated claim validation, anything depending on lived experience nuance, and anything where the cost of being a little wrong is unacceptable.
If you’re considering synthetic research, the question to ask is which of your current research questions fall into the strong-fit list and which don’t. The ones that fall into the strong-fit list are the ones to run synthetically. The ones that don’t stay on traditional research. Most research operations have plenty of both.
The audience for synthetic research has expanded fast. Three audiences are the most active:
Product discovery teams: product managers, designers, product ops, and the UX research leaders serving them. They use synthetic research to scale early-stage discovery. The pattern is consistent: PMs have more research questions than the research team can absorb. Synthetic research expands research throughput without expanding headcount, so researchers stay in deep work and PMs stop waiting.
Consumer insights teams at CPG, retail, and consumer-brand companies use synthetic research to screen concepts and test prices faster than panel research allows. The economic case is acute: 20 concepts a year, $23K median per traditional panel round, two concepts that make it to launch. Synthetic research screens the unpromising concepts before panel money is committed.
Healthcare and regulated CX teamsat payers, health systems, banks, and insurers use synthetic research to bypass the recruitment friction that defines patient and member research. Real patient panels can’t be recruited in two weeks, but a concept test that’s blocking a product launch needs to happen in two weeks. Synthetic respondents grounded in published patient research provide a usable answer without the recruitment timeline.
Beyond these three, consultants and agencies use synthetic research to add fast validation to their advisory work, and founders use it to validate before committing engineering resources. The common denominator across every audience: speed and access constraints that traditional methods don’t solve.
Synthetic user research is a research method, with strengths and weaknesses like every research method. When used in its strong-fit areas with rigorous methodology, it generates research-grade signal in hours that would take traditional research weeks. When used in its weak-fit areas or with weak methodology, it produces plausible-sounding AI improvisation that misleads more than it informs. The difference between the two is methodology rigor, evidence grounding, and honest scoping. Vendors that hide those, or that overclaim, are not worth your trust.
For the opinionated companion to this piece, read why synthetic research needs evidence grounding. To see synthetic research applied to specific decisions, explore Candor’s use cases: problem discovery, problem validation, concept testing, value-prop testing, price testing, and assumption validation. Or read the full how-it-works walkthrough. To see how Candor compares to other approaches, including traditional research and other synthetic platforms, read the comparison page.
Be the first to know when it launches.
No spam. Just a note when Candor is ready. Powered by Highline Beta.