Groups vs participants

Why two layers and not one

A participant on their own is one specific person. A pile of participants without structure is hard to reason about. The group layer gives you the structure: it says here are five distinct ways people in this audience approach decisions, and then each group is realised as multiple participants with the same behavioral DNA but different identities, life context, memory, and language.

Groups are good for thinking. Participants are good for interviewing.

The groups view

Each group card shows you the shape, not the people. The key fields:

Name and description: the human label and a paragraph on what makes this group distinct.
Defining dimensions: the specific axes that make this group different from the others, shown as tags like risk_orientation: cautious or tooling_preference: minimal. Hover any tag to see why this dimension is distinctive for this group. There are 3 to 5 defining dimensions per group.
Personality range (OCEAN): the Big Five personality traits, scored 0 to 100 against real population data and shown as min-max ranges rather than single values. A group might run Openness 55-75, meaning every participant in this group falls somewhere in that band, with a short line beside the bar saying what that band looks like in conversation. The bands are how groups hold together as behavioral shapes while still producing variation across participants. The (i) beside the heading explains how the ranges are checked, and when a second pass sets one itself.
Primary biases: the cognitive biases this group carries, each shown as a 0 to 100 range on the same kind of bar, with a line saying how hard it pulls and what it does. Status quo 50-70 means every participant in this group carries that bias somewhere in that band. Each group lists three to six biases. The bands come from research on how each bias varies across people. When three or more groups carry the same bias, they must spread across that range rather than bunching in the middle: the model writes each band, and a second pass sets one itself when the model keeps missing that spread. A bias carried by only one or two groups is placed by each group's own description and evidence instead.

Group cards collapse by default to the name and description. Click Show details to expand the dimensions, OCEAN ranges, and biases. Worth doing for every group before you start interviewing. It tells you what differences to expect and which conversations are worth running.

The participants view

Each participant is one of those group shapes brought to life. Groups do not all get the same number of them. Candor works out how much of your audience each group stands for, from the share of the evidence its segment holds, and gives it roughly one participant for every 5% of the audience. Every group gets at least three, so even a small one can show real variation, and none gets more than five, which is the range the checks that keep participants in a group distinct from each other were built for. Your study’s total is whatever those add up to, up to a ceiling of 24. The ceiling only bites on a broad audience that breaks into seven or eight groups, where every group is held to three. If Candor merges two groups that turned out to describe the same kind of person, the one it keeps takes over its segment’s share rather than both groups’ participants, so people are not piled into one group and the total usually goes down.

The participant card in the list shows:

Name: a believable first + last name.
Short descriptor: a single line capturing who they are and the situation they’re in.
Demographics pills: quick-reference tags like age, occupation, location, household, business size. The exact axes shown depend on whether the audience is B2C or B2B.
Sourcing: a chip reading Well sourced, Partly sourced or Thinly sourced: how many of this participant’s fields trace back to your evidence and your setup, counted from the same badges the profile shows. A participant whose citations were never checked shows a grey Not checked chip instead. See the participant profile guide for what feeds it.
Quality checks: not on the card at all. A participant that failed one is named, with a sentence saying what was counted, in the Participant quality check panel at the top of the page. A near miss is not flagged there or anywhere else, and neither are the internal checks that watch our own pipeline rather than the participant. On a business study, two of those checks compare the participant with what you set up: an Organization size line if they work somewhere outside the company size you gave, and a Seniority line if their role is more junior than the titles you gave. Both are enforced when participants are built: for a participant who misses either one, the fields that miss are corrected, up to twice, and the participant is left out only if it still misses after that, so those two lines are rare. A community or gender your setup words limited the study to works the other way round: it is settled before anyone is written rather than checked afterwards, so nobody is left out for falling outside it.

The card is deliberately thin. Personality and biases are one click deeper: open the profile and you get all five Big Five traits as a single value per trait, each with a reading of how the participant comes across at that score (each one is sampled from somewhere inside their group’s range), and every bias with its intensity and what it does. Click any participant card to open that profile and start interviewing.

Two buttons sit at the top right of the participants page, beside the page title, under the heading Next steps. Review evidence takes you back to the audience screen to check what the participants were built from. Start research takes you into research: to your research plans if you already have one, or straight into building one if you don’t. Viewers see only Review evidence, since their role can’t run research. Under the title, a line gives the count, for example “20 participants across 5 groups, built from your inputs and sourced evidence.” How many of each depends on your audience: Candor builds one group per evidence segment it found, splitting a segment in two only where its evidence shows two genuinely different kinds of person, and fills each group in proportion to how much of the audience it stands for. So a group covering more of your audience gets more participants than a small one. That count can be lower than planned: each participant is checked against the requirements you confirmed at setup once it is written. For one that contradicts a requirement, the fields that contradict are corrected, up to twice, and the participant is left out only if it still contradicts after that, so the audience is never widened to fill a slot. One whose profile never mentions a requirement gets the same correction and is kept as long as the corrected profile does not contradict it. On a study with confirmed requirements, the other boxes and readings you set, such as region or industry, are checked too, and so is whether the participant’s own facts fit each other: a participant who contradicts one is corrected the same way and left out only if the contradiction remains. Groups get the same check before any participant is written: a group whose description, personality range or bias range rules out a requirement is rewritten once. The groups are then checked again, and any group that still rules one out is folded into its closest group, so you may see one group fewer. A study already down to two groups keeps it. When no one meeting a requirement could sit at one end of a personality trait or bias, no group is pushed to that end to widen the spread. Below the buttons is the participant quality check, collapsed to a single line. Expanding it shows only the participants that raised a warning or an issue, plus a breakdown of what shaped this set. Participants that passed everything are not listed, so an expanded panel with nothing in it means nothing needs your attention. At the foot of the page sits Audience and participant inputs, collapsed, holding what you typed during audience setup so you can check it without leaving the page.

How to use the two surfaces together

Read groups first. They tell you what you’re looking at. Without that frame, every participant feels equally important.
Pick participants to interview from across groups, not within one. Three participants from the same group share the same reasons, so you hear them three times. One participant from each of three groups gives you genuine difference.
When tensions show up in synthesis, look at the groups. Genuine tensions almost always trace to two groups pulling in opposite directions on a specific defining dimension. That’s the build-or-pivot signal.

The question at the bottom of the page

Below the groups you’ll be asked once, per audience, how well these participants reflect the people you’re actually trying to reach, with an optional box for what stands out or what’s off. It goes to the Candor team, not into the research, and it’s the fastest way to tell us when an audience missed. Answering or dismissing it retires it for you on that audience; a colleague opening the same page is still asked, because you may read the same group very differently.

A note on rigor

The group layer isn’t a stylistic flourish. It’s the difference between a set of participants who all sound the same and one whose members genuinely disagree with each other in ways grounded in measurable behavioral differences. Big Five ranges, bias intensities, and defining dimensions are all calibrated against research data; the group is what holds them together as a coherent shape. The same review that checks those ranges also judges whether two groups describe the same kind of person; when it names such a pair, Candor merges them into one group before building participants, so you may see one group fewer than usual. The participant total does not fall: the merged group carries both groups’ share of it.