Key terms and concepts used in synthetic user research, participant modeling, and research-backed product validation.
CORE CONCEPTS
One AI-generated individual from a target audience, built from real evidence rather than imagination. A synthetic user has personality traits, cognitive biases, memory, and behavioral patterns drawn from market data and research, and you interview them one at a time the way you would a real person.
The same thing as a synthetic user. Candor uses both names, and prefers them over older wording because they make it clear you are interviewing an individual, not reviewing a summary of a whole group.
An AI-generated profile of a customer that you can question. The term covers two different things: a sketch that summarizes a whole group, and an individual with their own personality, memory and history. Candor builds the individual kind, and now calls them synthetic users or synthetic participants, because an individual is what you actually interview.
An AI-generated person who answers research questions in place of, or before, a real respondent. It is the survey and panel industry's word for a synthetic user. Synthetic individual, simulated people and AI persona mean much the same thing. A digital twin is different, because it copies one specific real person.
In research, an AI agent built to answer the way one specific real person would, made from that person's own data, usually a long interview plus a record of their choices. A synthetic participant is different: it is built from evidence about a type of customer and matches no real person.
A research methodology that uses synthetic users to simulate customer interviews, concept tests, and market validation studies. Designed to complement traditional research by providing fast, research-backed signal before committing to building.
The practice of building a synthetic user's attributes from verifiable sources (uploaded research documents, web evidence, and validated behavioral distributions) rather than generating them from a prompt alone. Each field in a participant's profile and memory is labelled with what is behind it.
The chip on each synthetic participant's card and profile, counted from the provenance badges on that participant's fields, with three bars under it for the Grounded, Derived and Assumed shares; the profile's section headers add up to them. Well sourced (green) means at most 15% of the fields read Assumed and at least half read Grounded. Thinly sourced (red) means more than 40% read Assumed, or fewer than 35% read Grounded. Partly sourced (yellow) is everything in between, and the chip's hover names the Assumed fields. A Derived field counts as sourced: it was built from your evidence or your setup even though no single source states it. A participant whose citations were never checked shows a grey 'Not checked' chip. Demographic fields count, so naming an age or region in your study setup gives more fields reading Grounded. Scores such as personality and biases carry no badge and are not counted. Groups on the audience screen carry a separate evidence-depth label.
A system for tracking where a synthetic participant's attributes came from. Each field in the participant's profile and memory carries one of three tags: grounded (traceable to evidence found for your audience, to what you told us when you set the study up, or to public population data), derived (follows from that evidence without a source that fits that field), or assumed (nothing pointed at it, and the model says so). Personality and bias scores, and the demographics summary at the top of the profile, carry no tag.
AUDIENCE & SEGMENTATION
The process of analyzing a target audience description and evidence to produce behavioral groups, signals, and calibration data. Candor divides an audience by psychographic and behavioral dimensions rather than demographics alone.
A discrete research finding extracted from evidence: a behavior, pain point, attitude, constraint, goal, belief, preference, or decision rule. Signals are tagged with type, sentiment, provenance, and intensity.
A meaningful part of an audience, defined by the behavioral or psychographic criterion that explains why its members make different choices from everyone else. Each group carries OCEAN ranges and primary cognitive biases, and the participants you interview are sampled from inside it.
PARTICIPANT MODELING
The five-factor personality model: Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism. Candor samples each trait from peer-reviewed population distributions calibrated by region and occupation, not randomly.
A systematic pattern in how people process information and make decisions. In Candor, biases are first-class traits with intensity values from 0 to 100, assigned based on group context and research-backed prevalence data.
The narrative knowledge that defines a synthetic participant's inner world. Four narrative types: identity memory (background, role), behavioral memory (goals, habits, constraints), belief memory (opinions, preferences, decision rules), and language memory (vocabulary, tone, communication style). Two more are kept separately and make up the six types the methodology essays describe: decision memory (the formative choices below) and conversation memory (what was said in earlier sessions).
Formative decisions pre-generated for each synthetic participant that establish their decision-making patterns. Each includes the choice scenario, factors weighed, outcome chosen, reasoning, and which cognitive biases were activated.
The process by which an AI model takes on a synthetic participant during an interview, drawing on their full profile (personality, memory, biases, prior decisions, and conversation history) to generate responses that stay in character.
INTERVIEW & RESEARCH
An interview mode where you type questions in real time and the synthetic participant responds conversationally. Each response is validated by a critic agent before delivery. Participants remember everything across multiple live interview sessions.
An automated interview mode where Candor conducts the interview based on a configured guide. It generates questions, tracks learning goal coverage, adapts probes based on responses, and detects when a synthetic participant is going off-profile.
An interview with a real person where an AI, not a human researcher, asks the questions, follows up on the answers and records the conversation. The respondent is real and only the moderator is automated. That makes it a different method from synthetic research, where the person answering is AI-generated too.
A structured configuration for auto-interviews that defines sections, learning goals, and question variants. Candor uses the guide to drive the conversation, ensure coverage of all research objectives, and decide when to move between topics.
A quality control step where a separate AI model reviews every response a synthetic participant gives, checking it for consistency. Hard contradictions (direct conflicts with established beliefs) are caught and regenerated. Soft tensions (mood shifts) are flagged but delivered.
The auto-interview system's method of monitoring which learning goals have been substantively addressed. It ensures every section of the interview guide receives adequate coverage before the interview concludes.
A mechanism in auto-interviews that identifies when a synthetic participant's responses are going off-profile or becoming repetitive. When detected, the system re-anchors the conversation or moves to the next topic to protect transcript quality.
ANALYSIS & OUTPUT
The structured output of an interview run, generated automatically once the last session finishes. Every report has four sections: what matters most, the findings, the next steps, and an evidence appendix. What sits inside the findings depends on the interview type, because a hypothesis verdict and a willingness-to-pay range are built differently. Today's reports paraphrase the evidence rather than quoting it, and the full transcripts sit in the evidence appendix.
Grouping related signals into named themes. In a problem-discovery report the themes are merged with the problems that surfaced on their own into a single findings list, ranked by how broadly each one showed up across synthetic participants. Other interview types order their findings differently: concept testing leads with the strongest verdicts.
A genuine disagreement surfaced in synthesis: different groups or synthetic participants expressing directly opposing views on the same topic. Tensions highlight real product decisions where you can't satisfy every group at once.
A mark on a problem discovery finding whose evidence is too slight to lean on yet. Reports do not label findings with a confidence level; where the evidence is weak enough to matter, the report says so on the finding itself. To judge weight elsewhere, read the amber banner at the top of the report if there is one, and the participant count that problem discovery and problem validation put on each finding. On a newer validation report, the count on a tested hypothesis is how many participants agreed out of those asked about it, with the ones who disagreed named beside it; an older one counts everyone who spoke to the topic. Newer discovery reports count strictly too, counting only people who said more than one distinct thing about a finding, and say so above the Findings list; older ones use the looser count and say only that the list is ordered by how broadly each finding showed up. Concept testing and price testing carry no such count, leading with a verdict per concept and a recommended price instead.
An explicit call attached to a tested item, shown as a coloured pill. Each interview type uses its own vocabulary: problem validation returns validated, partially validated, not supported, not applicable, or inconclusive; concept testing returns pursue, iterate, inconclusive, drop, or rewrite-and-retest.
An internal synthesis step that turns findings into product-relevant language. You will not see it as a heading in a report. Its output lands in the findings and next steps: what is working versus what needs rethinking for concept testing, value gaps and willingness-to-pay ranges for price testing, hypothesis verdicts for problem validation.
STUDY TYPES
A study type for exploring unknown problems in a market. You know who you want to serve but not what problems matter most. Synthesis highlights themes and tensions without opportunity framing. The goal is understanding, not solution validation.
A study type for testing specific hypotheses about customer problems. Write your hypotheses as problem statements, then interview synthetic participants to see whether the evidence validates each one. Synthesis delivers one of five verdicts per hypothesis (validated, partially validated, not supported, not applicable to audience, or inconclusive) with the evidence behind it.
A study type for evaluating a specific product concept. Synthesis frames what's working, what needs rethinking, and who the concept is actually for. Structured to drive build-or-pivot decisions, not just generate themes.
A study type for understanding how customers perceive value and what they're willing to pay. Synthesis extracts value perception gaps, price anchoring effects, and willingness-to-pay ranges segmented by group.
The second choice you make inside concept testing. A standard test shows whole concepts cold and compares them against each other. A name test, a value proposition test, and a messaging test each expose one anchor concept first, then present the candidate options inside that context, so nobody judges a name or a headline in a vacuum.
A concept test that evaluates candidate names for one concept. Synthetic participants meet the concept first, then react to each name: whether it fits, whether it resonates, and whether it connects to what the product actually does.
A concept test that evaluates candidate benefits for one concept. Synthetic participants see them one at a time in the concept's context and say which is most compelling and most believable.
A concept test that evaluates candidate messages or headlines for one concept: which is clearest, which is most persuasive, and which makes a synthetic participant want to learn more.
A methodology the guide switches to automatically when you give it enough items to compare. Synthetic participants sort the items into buckets and rank them, and the interviewer then deep-dives the top-ranked ones rather than treating all of them equally.
Start with a free project, or walk through Candor with us first.