Prompt an LLM
The AI improvises a character from general knowledge, or from documents you upload. No personality model, no memory, no way to separate your evidence from what it filled in.
There are many ways to get customer signal. This page covers the two alternatives people weigh Candor against most often: prompting a large language model yourself, and other dedicated synthetic research platforms. For research with real people, the head-to-head pages at the bottom cover panels and moderated testing.
A chatbot gives you a plausible answer. A research instrument gives you an auditable one.
Prompt an LLM
The AI improvises a character from general knowledge, or from documents you upload. No personality model, no memory, no way to separate your evidence from what it filled in.
Candor
Builds participants from real evidence about your audience. Full personality. Consistency checks. Every important attribute labeled by where it came from.
You can, and for a quick gut check it’s fine. The difference is structural, and it shows up in six places.
| Dimension | Prompting an LLM | Candor |
|---|---|---|
| Evidence | The AI reads your prompt (and any documents you paste in) as background context, then improvises a character from it. It doesn't extract specific claims or organize what it found. It just has more material to wing it from. | Candor pulls specific signals from your documents and from published research, organizes them into a structured evidence base, and builds participants from that. You review the evidence before any interview runs. |
| Personality | You describe the character in a sentence: “You are a pragmatic, budget-conscious nurse.” The AI improvises the rest. | Each participant gets a full personality profile drawn from published population data for their region and occupation, plus decision-making biases modeled as traits that shape how they answer. Two people in the same group can hold different views. |
| Memory | Within a single conversation, the AI remembers what was said. Across sessions, it doesn't. There's no persistent identity, so you can't return to the same participant next week and pick up where you left off. | Participants carry their earlier interviews with them: the stories they told, the opinions they shared, the decisions they described. Come back later and pick up where you left off. |
| Consistency | Nothing checks whether answers are consistent. If the participant says one thing in question 3 and the opposite in question 7, you catch it yourself or you don't. | A separate system validates every response against the participant's profile and prior statements before it reaches you. |
| Interview | You run the interview yourself. The AI answers, but it doesn't follow up, doesn't notice what's interesting, and doesn't track what's already been covered. | The interviewer probes, adapts to what each participant says, follows threads that matter, and knows when a topic has been explored enough to move on. |
| Provenance | When the participant responds, you can't tell which parts came from your evidence and which parts the AI filled in on its own. The participant's traits, opinions, and behaviors aren't tagged with sources. | Every important attribute is labeled grounded, derived, or assumed, so you can check what came from your evidence, what was inferred from it, and what the system filled in. |
Both comparisons come down to what the tool actually produces. Below: two participants from the same study, same audience, same question. Different people, because they were built differently.
Relationships and emotions
Calls her advisor friend to vent. Her mom talks her through it on the phone. Feels discouraged, wonders if she did something wrong. Wants a clean number and gets a shifting range instead.
Systems and efficiency
Tracks bottlenecks in a four-way comparison spreadsheet. Frustrated that external dependencies keep resetting his timeline. Finds it ironic that he built a system to move fast and can’t.
From the interviews
Tell me about the last time something in your home search took way longer or turned out way more complicated than you’d expected going in.
Oh, that would be the whole pre-approval piece, honestly, I thought that part would be quick and it ended up dragging out way more than I expected. So what happened was, after that first call with the loan officer, I figured I’d just send over a few documents and get a pre-approval letter within a few days, that’s kind of what I assumed going in, since everything online made it sound pretty streamlined. But it turned out my income situation was a little more complicated than I realized, because I get part of my pay through bonus structure at work, and apparently lenders want to see a longer history of that before they’ll count it fully toward what I qualify for.
So instead of a few days, it turned into almost three weeks of back and forth, she kept asking for additional documentation, an extra year of tax returns, a letter from my employer explaining the bonus structure, that kind of thing. I remember feeling a little discouraged honestly, like did I do something wrong, even though she kept reassuring me it was totally normal for anyone with variable income. I called my advisor friend in the middle of it just to vent a little, and he explained that lenders average bonus income over a couple years usually, which actually made me feel better once I understood the why behind it.
My mom was on the phone with me one night when I was getting kind of frustrated about it, and she said something like, well if it takes longer to get it right, that’s better than getting a number that falls apart later. Which, she’s right, but in the moment it was just annoying, I wanted a clean number to plug into my spreadsheet and instead I had this range that kept shifting.
Think of a recent step in your home search that ended up taking a different amount of time or effort than you’d first pictured. What happened?
The preapproval sprint itself is the clearest example, I pictured that taking maybe a week, like get four quotes, build the comparison, done, because I’ve done fast research sprints before in other parts of my life and I’m generally efficient at this stuff. But it’s dragged out closer to a month now, mostly because loan estimates don’t all arrive on the same timeline, one lender took four business days just to send the actual itemized form after I’d already had a rate quote from them verbally, and another one wanted extra documentation on my equity compensation before they’d even generate a real estimate, which I hadn’t budgeted time for at all.
I think I underestimated how much of this process is actually bottlenecked by other people’s internal systems rather than by my own analysis speed, which is a weird blind spot for me because in my actual job I’m usually the one who understands where the bottlenecks live. Every time I thought I was close to having a clean four-way comparison, one lender would send a revised estimate because rates moved, or I’d ask a clarifying question that took two days to get answered, and I’d have to redo a chunk of the spreadsheet.
It’s not that any single step took absurdly long, it’s that the whole thing kept resetting the clock every time new information came in, which is sort of ironic given that I built this whole system specifically to move fast once I had the data.
One participant across two interviews. The follow-up opens with a simple check: in your own words, what does it do? Ari recalls the concept accurately, then brings back the objection he raised the first time, unprompted.
I’m going to share something we’re exploring: a pre-approval tracker that shows you, for each lender, which documents are still outstanding and whose court they’re in. What is your first reaction?
Honestly, this is close to what I already built for myself in a spreadsheet, so the idea lands. All four lenders in one view with the outstanding documents next to each one, that’s the comparison I’m maintaining by hand right now. My problem is the data source. If the lenders are the ones updating their status, the whole thing inherits their timelines, and one of mine took four business days just to send an itemized form after I’d already had the rate verbally. A tracker that waits on the slowest lender isn’t a tracker, it’s a prettier version of me waiting. If it pulled status automatically, different conversation.
When we spoke before, you saw a concept: a pre-approval tracker for your lender comparison. In your own words, what does it do?
Right, the dashboard idea. One view across the lenders you’re working with, showing which documents each one still needs and who owes the next step. I think it claimed same-day status updates, or maybe it was within a day, something like that. Useful in principle, I’m the guy with the four-way spreadsheet, so I’m clearly the target. But like I said before, it lives or dies on whether the lenders actually update their side. My bottleneck was never my own tracking, it’s their timelines, and a dashboard doesn’t move those.
He recalls the concept, hedges on the detail he’s less sure of, and holds the position he took the first time. That’s what memory looks like in a transcript: not a database recall, a person picking up where they left off.
These exchanges are recreated for illustration. The behavior they show comes from real customer studies, whose transcripts stay private.
Purpose-built synthetic research tools exist, and the serious ones have real participant models. From a landing page, they all sound like Candor. The difference shows up in the output, and in what a vendor can show you when you ask.
Most platforms work like an open-book exam. The participant is a character description, and when you ask a question, the system searches your documents for useful passages and writes them into the answer. The participant doesn’t know your research. It looks things up.
Candor spends the first half of every study researching the audience before anyone exists to interview: evidence is gathered, checked, and clustered, personality is drawn from published population data, and each participant’s profile is built from that evidence, each field labelled with what is behind it. The participant doesn’t consult the evidence. They were shaped by it.
Both are grounding, and lookup has real strengths: it’s fast and stays current. They just produce different participants. One that references your research, and one whose opinions, habits, and blind spots came from it.
The tell in a synthetic report is that it could have been written without running the interviews: no names, no counts, nothing traceable to any participant. A Candor report shows its work: discovery findings carry participant counts and group coverage and flag whether a theme surfaced unprompted or only when asked, validation and concept-test reports name the participant behind each paraphrased piece of evidence, and every report comes with the full transcripts, so you can trace a claim instead of taking it.
A 182-study review of synthetic participants (Kuric et al., 2026, preprint) found typical output loses the contradictions and stray details that mark a mind at work. Candor’s discovery reports keep them: a finding whose participants contradicted themselves carries a “Contradictions they carry” block, because a participant who never contradicts themselves is a warning sign.
Typical synthetic output presents every claim at the same confidence. Candor’s participants flag the topics where they’re outside their experience, and those flags sit in the transcript next to the answer. Reports frame findings as a map to what to test next, not a verdict.
If the tool can’t show you, the participant is improvising from the model’s general knowledge, no matter what you uploaded. Candor’s answer: participants are built from an evidence base you can open and review before any interview runs, and every field in a participant’s profile and memory carries a tag telling you whether it’s grounded, derived, or assumed.
Without a real model underneath, five participants collapse into five versions of the same moderate opinion. Candor’s answer: personality sampled from published population data, biases modeled as traits, and on top of that a memory of their own. Background, habits, beliefs, ways of speaking, each labeled by where it came from. Patricia and Ari above are what that looks like.
If nothing checks answers against the participant’s profile and prior statements, the person in turn ten may not be the person you started with. Candor’s answer: a separate system validates every response before it reaches you. The memory exchange above shows what that looks like across two interviews.
Looking for a comparison against a specific platform? Read Candor vs Synthetic Users for an honest comparison of two rigorous synthetic research platforms, Candor vs UserTesting for how synthetic research fits alongside human-moderated UX testing, or Traditional vs synthetic user research for how synthetic screening fits before panel-based concept testing, brand tracking, and substantiated-claim research.
Weighing a tool that builds digital twins of real people? Read Candor vs Brox for how twins differ from participants built from evidence about a type of customer, and what each design can and can’t check.
Want a research team to run synthetic surveys and hand back a marketing plan? Read Candor vs Evidenza for how a managed service for marketing teams compares with interviews you run yourself.
Weighing a tool that surveys reusable synthetic audiences? Read Candor vs Minds for how survey-style audiences differ from participants built for one-to-one interviews, and which questions each answers.
Need survey-style numbers for a large population? Read Candor vs Aaru for how population simulation compares with one-to-one interviews, and where each one is the better fit.
Candor is secondary research that makes primary research sharper. For the four moments where it fits into a research process, see where Candor fits on the how it works page.
Synthetic research is a complement to real user research, not a replacement. It's best for early-stage validation, hypothesis generation, and pressure-testing assumptions before committing real research budget. Candor's evidence grounding and provenance tagging make its findings more transparent than typical AI outputs. But the gold standard is still talking to real customers. The best teams use synthetic research to make their real research sharper and more focused.
Not entirely, and it's not designed to. Candor is built for the step before traditional research, or for situations where traditional research isn't available (budget constraints, tight timelines, early-stage exploration). It generates evidence-backed signal fast, so you can decide where to invest real research effort. Think of it as a research accelerator, not a replacement.
When you prompt ChatGPT to roleplay a user, the model generates a plausible-sounding character from training data. No evidence grounding, no personality model, no cognitive bias calibration, no consistency validation, no memory across sessions. Candor runs a multi-stage pipeline: evidence retrieval, audience segmentation, OCEAN sampling from real distributions, bias assignment, critic validation on every response, and persistent memory. The difference is between a guess and a research instrument.
Three things. First, evidence grounding: each synthetic user is built from your research documents and real market data, not generated from a prompt. Second, calibrated psychology: OCEAN personality traits sampled from peer-reviewed population distributions by region and occupation, with cognitive biases assigned at research-backed intensities. Third, consistency enforcement: a critic agent validates every response against that participant's established beliefs and prior statements before you see it.
Synthetic users are grounded in real data, but they're not real people. They're synthetic individuals whose attributes are derived from your uploaded research documents, web evidence about your target market, and validated behavioral distributions. Each field in their profile and memory carries a label saying whether it rests on something you can check, follows from the evidence, or was assumed. The individuals are generated. Their foundations are real evidence, not imagination.
Start with a free project, or walk through Candor with us first.