When you submit a project, Candor runs a multi-stage pipeline that researches your audience, extracts signals from evidence, and organises everything into segments. Here’s what happens at each stage and what it means for the audience you end up with.
Audience generation is one background job that runs through around twenty steps in sequence, and its participants are generated straight after it in the same run, with nothing for you to do in between. The setup wizard tells you the typical duration when you submit (around 25-35 minutes for both halves on most projects). The progress screen shows which step is running and updates in real time. You can close the tab.
One thing to understand up front, because it explains most of what follows: Candor proposes the segments early and then goes looking for evidence about them. It does not gather everything first and sort it into groups at the end. That ordering is why your audience description carries so much weight. It is the input to the first real decision the pipeline makes.
If you uploaded research documents, Candor waits until they’re parsed and embedded before starting. This is fast (under a minute per document) but bigger files take longer. If a document fails to parse, you’ll see a warning and the pipeline continues without it.
It then normalises your setup inputs into a working audience description, reading the list you confirmed on the last step of setup in place of your two longest answers, and pulls the relevant evidence out of your documents. The facts and quotes your documents state are weighted more heavily than anything found on the web, throughout the run. Where a document’s author interprets their own evidence, that sentence is kept as the author’s reading rather than as a fact: it is left out of the evidence participants are written from, and it cannot on its own define a group or a dimension. Your own two longest answers at setup reach participants and the searches only through the list you confirmed on its last step, Every participant will match this list: a line you kept, reworded or added builds every participant, and nothing else in those two answers does.
That audience description is what everything after it builds from, so anything you corrected on the last step of setup is put in front of this step as settled rather than read again from your original wording. Correct “Gen Z” to 22-30 before you start and 22-30 is the number this step is working from. The same goes for who you limited the study to: confirm “only Black women” on that step and the audience description is written for Black women, not for everyone.
Candor now proposes a set of candidate segments, grouped by behavioural and psychographic dimensions rather than demographics alone. Each one gets a primary differentiating dimension: the criterion that explains why people in this segment make different choices from everyone else.
A validation step immediately checks those proposals. If they don’t hold up, the pipeline regenerates them rather than continuing with a weak set, and you’ll see a retry step on the progress screen. That is the pipeline working, not failing.
With segments in hand, Candor plans what evidence it needs and then runs the main search pass on the web. The main pass is written from your brief alone: 40 searches across ten topics, half about your audience’s inner life (what they want, worry about and value) and half about their circumstances (money, routine, market and region). What you said you’re unsure about is left out, because the interviews answer it. Queries are region-aware, tuned to the region you pinned during setup.
When a gap-fill search asks why your audience does not do something, such as why occasional donors don’t give monthly, Candor adds one search about the people who already do it. Those findings are there as contrast. Each one is meant to name that neighbouring group, not to describe your audience.
Every source, whether a web page or a chunk of a document you uploaded, is read and broken into discrete signals: behaviours, pain points, beliefs, constraints, goals, preferences, and decision rules. Each signal is tagged with type, sentiment, intensity, and a link back to the source it came from.
A search steered only by your audience description finds the obvious, engaged, complaining middle of an audience. Four extra passes exist to correct for that, and they are a large part of why your synthetic participants come out varied rather than interchangeable.
Candor now audits what it has and identifies gaps: signal types that are under-represented, claims lacking supporting evidence, or segments that still aren’t well covered. It runs targeted searches to fill those gaps, one query at a time, until the audience is balanced or the search budget is spent. A topic your document’s author only reasons about, without a stated fact or quote behind it, still counts as a gap here, so the search goes looking for independent evidence for or against that reading. A rebalancing step then corrects a signal mix that skewed too far toward one type.
With the evidence settled, Candor seeds the attribute dimensions that participants will later be built along, and generates calibration hints: the realistic ranges and distributions for this specific audience, so sampled values land where real people land rather than at a generic average. Both steps have their own retry loops if the first attempt doesn’t validate.
Most signals are lifted from one page, with the original wording kept beside them. Calibration also writes signals of its own, pulling two or three findings together into a single statement about the audience. That pulling-together is where a claim is most likely to bend, so each one is re-read against the wording of the sources it cites before anything is saved.
Bending means a number that measured one thing described as another, a study’s country or care setting dropped, or two figures with different bases merged into one range. On an ordinary run the page shows nothing about this check. A line above the signal list appears only when the check didn’t run, or found no synthesized signal it could trace back to its sources, and it says those signals are unverified. That matters because a run that checked nothing and a run where everything held up produce the same quiet list. Nothing is removed either way.
A claim that doesn’t hold up also gets one attempt at a fix, during the same run, before anything is saved. Candor asks for a rewrite that stays true to the sources, re-runs the check on the rewrite, and keeps it only if the check now rates it better than the original. A claim it cannot improve is kept exactly as first written. Nothing is deleted either way, and whether a claim still counts as flagged is judged after that attempt, not before it. This happens inside the run, before the participants are built from these claims, so what a participant is generated from is the corrected wording.
A separate agent reviews the assembled signal set for weak evidence, contradictions, region drift, and over-reliance on single sources. If it finds fixable problems, the pipeline can loop back and regenerate dimensions or calibration before running a final quality check. This pass never blocks; it only feeds corrections back into the run.
Its findings aren’t shown on the review screen. What you read there instead is the evidence itself: how many signals sit behind each group, and how they break down by provenance. That is the same question answered in a form you can act on.
One more check runs later, while the participants are being built. Each participant’s profile and background is read for claims about how their field works, and a claim an expert would call wrong (a standard process described in the wrong order, a technical relationship reversed) is noted in the build’s records. Nothing is removed or rewritten: every participant is built exactly as written, and the check never stops a build.
The participant groups get their own review before anyone is built. When that review judges two groups to describe the same kind of person, the pipeline merges them into one, so a study can end with one group fewer than it planned. The participant total planned for your audience is kept; the merged group carries both groups’ share of it.
Early in the run, Candor also captures a snapshot of recent developments in your industry and region. It isn’t used to build the audience. It’s stored on the project and injected into interviews later, so participants can refer plausibly to what has been happening in their world instead of sounding frozen in time. You can refresh it from the project page when it goes stale. The interview panel and the auto-interview run page show its date read-only.
If a stage fails, the pipeline stops and shows a failure banner with a retry button. Common causes are transient web-search errors, rate limits, or evidence too sparse to segment cleanly. Retry usually works for the first two; for the third, broaden your audience description or add more evidence and create a new project.
That window covers both halves: the audience, then its participants, in one run. Most of the audience half is web search and signal extraction. Candor runs an initial broad pass, then iterates through gap-fill searches to balance the audience. Each iteration includes a search, a read of every returned page, and signal extraction from that content. There's no way to compress this without dropping quality. The pipeline also runs a critic pass at the end. The participant half then clusters groups and builds individual participants from the evidence gathered. If you want faster, narrower audiences, use more upfront evidence (uploaded documents) to reduce the search budget needed.
Four kinds of issue. Weak evidence (signals that lean on a single thin source). Contradictions (signals that conflict without explanation). Region drift (evidence from outside the region you pinned creeping in). Over-reliance on single sources (too many signals citing the same page). The critic doesn't block, it flags, and its findings are not shown on the review screen. Judge the same thing from the evidence that IS shown there: the signal count behind each group and its provenance breakdown. A separate accuracy check runs earlier, on synthesized signals only: those are built from several sources at once, and each claim is re-read against the wording of the sources it cites, and anything that doesn't hold up gets one automatic attempt at a rewrite during the run, kept only if re-checking rates it better, before the participants are built from those claims. On an ordinary run the page shows nothing about that check; a line above the signal list appears only when the check did not run or found nothing it could trace, and says those signals are unverified. Because participants generate in the same run, you're reading all of this after both halves finish: use it to decide whether to regenerate or add more evidence before you run interviews.
Because segments are proposed before the main search runs, sparse evidence usually shows up later as segments that can't be grounded rather than as a failure to form segments at all. You'll see it as thin evidence depth on the review screen, on the groups that could not be grounded. Occasionally an audience does come back with too few groups to build participants from. The review screen says so in place of the group picker, lists what it did find, and offers you a way to reach us, because no choice you make there can get past it. The fix is to broaden your audience description, upload more documents, or both. Reuse research you already have. Anything you didn't upload is something Candor has to search the web for, which produces shallower signal than your own primary research.
You can create a new project with the same audience description and adjusted evidence (more documents, a refined description, a different region). The original project doesn't get overwritten. Rerunning from scratch usually produces meaningfully different output, since the gap-fill iteration is non-deterministic in which sources it pulls. If the issue is one segment looking weak, you don't need to rerun everything. The segment selection gate lets you exclude weak segments without rebuilding the audience.
Yes, significantly. Region tunes the search queries (different countries produce different sources), affects which sources get prioritised, and propagates into each synthetic participant's world-context snapshot. A study pinned to Brazil will find Brazilian sources and produce participants operating in Brazilian context. Pinning region matters most when your audience is country-specific or regulatorily different across regions. For a global audience, "Global" is a valid choice and Candor will treat it as such.
Start with a free project, or walk through Candor with us first.