How evidence grounding works

In Candor, evidence grounding means every synthetic participant is built from real evidence gathered before anyone is generated (published research, your uploaded documents, validated population distributions), and each field in a participant's profile and memory is labelled with what is behind it: something you can check, something that follows from the evidence, or an assumption Candor flags. Those labels are visible to anyone reviewing the study. This piece walks through how the pipeline actually works. For the category-level introduction, see what is synthetic user research.

If you want the opinionated version of why this matters, read why synthetic research needs evidence grounding. This piece is the technical companion: what the system does, step by step, and what the methodology gives you.

What evidence grounding means in practice

Most "AI persona" tools work the same way. The user describes an audience. The AI invents a character that fits the description. The output sounds plausible because language models are pattern matchers, and the patterns of "a 35-year-old marketing director" exist in training data abundantly enough to produce a fluent character on demand. There is no link between that character's attributes and any specific evidence about the actual audience.

Evidence grounding flips that. Before any participant is generated, Candor retrieves real evidence about the target audience: published research, peer-reviewed studies, market reports, any first-party documents the customer uploaded, and validated behavioral and demographic distributions. The participants are constructed from that evidence, not from a prompt alone. Each field in a participant's profile and memory carries a tag saying whether there is something to check behind it. When a participant answers an interview question, the underlying reasoning traces back to evidence that exists in the system, not to the AI's pattern-matched best guess.

The practical difference shows up in three places. First, the participants reason about pain points, beliefs, and behaviors in ways that reflect what the actual audience has been documented to say, not the AI's compressed average of every "marketing director" in training data. Second, the methodology is auditable: a UXR lead or insights director can inspect the provenance and push back where the evidence is thin. Third, when the audience evolves (new market entrants, shifting attitudes, new behavioral data), re-running the evidence stage produces participants who reflect the updated reality, not the same generic pattern.

The rest of this piece walks through the five stages that produce that result.

Stage 1: Evidence retrieval. Where the data comes from

The first stage of every Candor study is evidence retrieval. At study setup, the user describes the target audience: B2C or B2B context, the learning goals they want to address, the industry, and the region. They can also upload any first-party research they already have. The specific interview type is selected later, when the research plan is built, and name, value proposition and messaging tests are kinds of concept test rather than separate interview types. From the initial audience input, Candor builds a search plan and runs two parallel retrieval workflows.

Uploaded documents come first. If the customer has uploaded research (PDF, DOCX, CSV, TXT, or MD files: customer interview transcripts, survey data, market reports, prior research deliverables), Candor parses and indexes those documents. The facts and quotes they state become the primary evidence source. Internal research about your specific audience is more relevant than general public evidence about the broader category, so uploaded documents get weighted higher in retrieval. One thing is kept apart: when a document's author interprets their own evidence (a survey memo's "likely because they trust the existing process", or an interview summary written in the analyst's voice), Candor records that sentence as the author's reading rather than as a fact. It can colour participants and shape hypotheses; it cannot on its own define a group or a dimension.

The documents are parsed into chunks, embedded into a vector space, and stored in a searchable index. When the pipeline needs evidence about a specific dimension of the audience (their pain points, their decision criteria, their attitudes toward category X), it retrieves the relevant chunks via vector search.

Web search runs alongside the documents. In parallel with the document layer, Candor runs structured web searches across published research, industry analyses, peer-reviewed studies, public consumer-health and market-behavior datasets, and any other source likely to contain evidence about the audience. The search strategy is iterative rather than batch: the pipeline runs a broad initial pass, then assesses what's been retrieved, identifies coverage gaps, and runs targeted follow-up searches to fill them.

The web layer is critical when the uploaded documents are thin or when the audience extends beyond what the customer has researched directly. For pre-launch products, healthcare populations that are hard to recruit, regulated industries, and any audience where first-party research is incomplete, the web layer is often the bulk of the evidence.

Both layers feed the same evidence pool, and a signal's source tells them apart. Where an audience signal has a source, it shows whether that source was the customer's own upload or a public web page retrieved by Candor, so the customer can see which parts of the evidence pool are theirs.

Stage 2: Signal extraction. Turning raw evidence into structured participant attributes

Retrieved evidence by itself is not yet usable. A PDF of a customer-interview transcript or a peer-reviewed paper about consumer behavior is unstructured text. To build participants from it, Candor first extracts structured signals from the evidence.

A signal is a structured representation of one piece of behavioral or attitudinal information about the audience. Candor extracts hundreds of signals per study across several categories:

• Behaviors: what the audience actually does (purchase patterns, app usage, communication preferences, decision-making sequences)
• Pain points: what frustrates them, what they avoid, what they actively complain about
• Attitudes: what they believe, what they're skeptical of, what they value
• Constraints: what limits their choices (budget, time, regulatory, organizational, capability)
• Goals: what they're trying to accomplish, both stated and implied
• Beliefs: what they assume to be true about the category, the alternatives, the world
• Preferences: what they prefer when given choices, including the reasoning behind the preference
• Decision rules: how they actually decide, including shortcuts and heuristics

Each signal is extracted from a specific piece of evidence and tagged with metadata: which source it came from, which category it belongs to, what confidence level the extraction has, and what segment(s) of the audience it applies to.

The extraction is run by language models with structured-output validation, so the signals come out in a consistent schema rather than as free text. This matters because the next stages (segmentation, participant generation, interview behavior) all operate on the structured signal representation, not on raw evidence.

Signal extraction is also where the pipeline starts to see the shape of the audience: which segments cluster around similar pain points, which signals are universal versus segment-specific, where the evidence is dense versus thin. That shape feeds the next stage.

Stage 3: Provenance tagging. The audit trail at the attribute level

This is the stage where Candor's approach to rigor becomes visible in the product surface. Every signal extracted from evidence (and every profile and memory field on the participants later built from those signals) carries a provenance tag. There are three, and they answer one question: can the customer check this?

1. Grounded: there is something to check behind the value. On a signal, that means it traces to a specific source that states it, either lifted from a single web page or from a fact or quote in an uploaded document, or built from several named sources without being flagged by Candor's accuracy check; web sources link out and uploads name the file. A document author's own reading of their evidence is not grounded, even though the file is named: it reads Derived. On a participant field it means one of three things: the field cites evidence found for this audience and, where Candor checked those citations against the field's own wording, at least one matches, a value the researcher fixed themselves when setting the study up, or public population data such as the US Census for education and household on US consumer audiences. Hovering the label says which of the three it was; the profile does not name the document or signal behind a field.
2. Derived: it follows from that evidence, but no source states it for this field. Either the value was fitted to the type of participant the research described, or it was reasoned a step past what the research said outright. Weaker than grounded because the connection is Candor's rather than a source's. Stronger than a guess because there is evidence underneath it.
3. Assumed: no specific evidence for this field. Candor filled the gap from everything it knows about the participant, and says so rather than dressing it up. Treat these as things to test, not things to trust.

Candor stored five tags until September 2026 and still does internally. Five words became three: the ones that stopped being shown named the mechanism that produced a value, whether it came from a distribution or from the pattern of a group, which described how the pipeline works rather than what the customer can do about it. Nothing was retagged, so a participant generated a year ago reads the same way as one generated today.

The provenance tag is visible on every audience signal and on every participant memory field and attribute. Synthesis reports are traceable a different way. When a finding reads "Participants in Segment A consistently described claims-process friction as their top frustration," a discovery report shows how many participants and groups raised it and whether they raised it unprompted, and the full interview transcripts come with the report, so the customer can read exactly what was said.

This is the auditability that separates synthetic research from AI improvisation. A UXR lead reviewing a study can challenge the methodology at the attribute level, not just at the conclusion level. An insights director defending a launch decision to leadership can cite the provenance chain rather than relying on the team's word that the AI was "trained on the right data." A skeptical stakeholder asking "where did this finding come from" gets a real answer (in a discovery report, how many participants raised it; in a new validation report, how many agreed with each tested hypothesis and who disagreed; in every report, the full interview transcripts to check it against), not a wave at the model.

Stage 4: Iterative gap-fill. Closing the holes

Evidence retrieval doesn't return uniform coverage. Some audiences have rich published research and dense first-party data; others have sparse coverage in one segment or across one signal category. Candor's pipeline assesses coverage after the first retrieval pass and runs targeted follow-up searches to fill identified gaps.

The gap-fill iteration looks at:

• Signal-category balance: are some categories (behaviors, pain points, attitudes, constraints, etc.) under-represented relative to others? If so, run searches biased toward the under-represented categories.
• Segment coverage: are all the defined audience segments equally represented in the evidence pool? If one segment is thin, run searches targeted at that segment.
• Diversity within signals: are the retrieved signals reflecting a diverse range of perspectives, or are they clustering around a single source's framing? If the latter, run searches to broaden the perspective base.
• Recency: is recent evidence present, or is the retrieved material older than is useful for the question? If older, run searches biased toward recent sources.

Each gap-fill iteration runs targeted web searches against the identified gaps, extracts signals from the new evidence, and re-assesses coverage. The loop continues until either (a) the gaps are sufficiently closed, or (b) the search budget for the study is exhausted.

This iterative pattern is why audience generation in Candor takes 25 to 35 minutes (rather than seconds): the system is actually doing meaningful evidence work, not generating a character from a prompt. The minutes are spent retrieving and extracting evidence, not spent waiting on slow infrastructure.

Stage 5: Critic review. Quality control before participants are generated

Before the evidence pool is handed off to participant generation, a critic agent reviews the assembled evidence and signal set for quality and diversity. The critic is a separate AI agent with a structured rubric covering:

• Diversity of perspectives: are multiple viewpoints represented, or is the evidence biased toward one framing of the audience?
• Signal redundancy versus coverage: has the gap-fill iteration produced redundant signals (the same finding from too many sources, with diminishing marginal information) or genuine breadth?
• Provenance health: is the evidence pool dominated by signals Candor flagged as its own assumptions (which would produce hypothesis-grade participants rather than research-grade ones), or is there a healthy proportion that traces back to real sources?
• Segment differentiation: do the segments actually differentiate from each other on the signal evidence, or are they overlapping enough that the segmentation may not be meaningful for the study?

When the critic flags issues, the pipeline either runs additional retrieval to address them or returns the issues to the synthesis stage as caveats the customer should see in the final report. The critic isn't a binary pass/fail gate; it's a quality-signal layer that surfaces concerns the customer should know about before drawing conclusions.

The critic step also catches a class of failure modes that earlier synthetic research tools regularly shipped without noticing: when the evidence pool contains internal contradictions, when one source dominates the signal pool disproportionately, or when the segmentation doesn't actually correspond to differentiated signal patterns. These aren't catastrophic failures; they're quality issues that, surfaced honestly, let the customer decide whether to proceed, re-run with different inputs, or escalate the question to real-customer research.

What evidence grounding gives you (and what it still can't do)

The cumulative effect of these five stages is research with an audit trail. Specifically:

Participants who reason about the audience's real patterns. Pain points, beliefs, and decision rules in the synthetic interviews track to documented evidence about the audience, not to AI-pattern-matched stereotypes. This produces interview content that researchers familiar with the audience will recognize as authentic rather than dismiss as generic.

Findings that survive stakeholder scrutiny. When a synthesis report goes to a stage gate, the answer to "where did this come from" is the transcripts behind the finding and the tagged evidence behind each participant. Researchers can defend the methodology. Insights leads can challenge specific attributes rather than dismissing or accepting the whole study as a black box.

A methodology that adapts to new evidence. When the audience evolves, re-running the evidence stage produces participants grounded in the updated evidence, not in last year's snapshot. This is part of why synthetic research can model drift over time: the underlying evidence is reload-able.

A clear handoff to real-customer research. Synthetic research is at its best as a first-pass layer that sharpens what to study with real customers. The provenance trail makes this handoff concrete: the surviving hypotheses are documented, the evidence each one rests on is named, and the gaps the real research should address are explicit.

What evidence grounding still can't do:

It can't produce statistically-bounded point estimates. The signals are qualitative; the participants reason about the audience; the synthesis finds themes and tensions. None of this produces "27% of customers prefer X, plus-or-minus 3pp at 95% confidence." That number requires real-respondent panel research at sufficient N.

It can't substitute for real-customer ground truth. The evidence pool reflects what's been documented; real customers behave in ways that haven't been documented yet. Real-customer research surfaces the unexpected; synthetic research reflects the expected within evidence-supported bounds.

It can't make up for missing evidence. If the audience is genuinely under-researched (a brand-new category, a niche professional segment with no public coverage), evidence grounding only goes as far as the evidence does. The gap-fill iteration helps; it doesn't manufacture evidence that doesn't exist. In those cases, the critic surfaces the gap, and the responsible move is to lean harder on the customer's first-party data or treat the synthetic findings as hypothesis-grade.

It doesn't replace real human judgment about what the research means. Evidence grounding produces grounded findings. Strategic interpretation of those findings remains a human job. The teams getting the most value from Candor treat the platform as an evidence-acceleration layer and the strategic synthesis as their own work.

Related reading

For the downstream stages that build on evidence grounding, see how group clustering works (turning the evidence pool into group templates), the OCEAN model in synthetic participants (the personality calibration layered on top), and how participant memory actually works (the six memory types that make a participant behave as the same person across sessions). For the broader category context, see what is synthetic user research, the use cases Candor supports, and the comparison hub for how the methodology stacks up against alternatives.