Where synthetic research fits in your workflow

The question isn't synthetic or real. It's where synthetic fits in what you're already doing, and there's more than one right answer.

The question people ask first is usually "synthetic or real?" It's the wrong question. The useful one is "where does synthetic fit in what I'm already doing?" And there's more than one answer.

The story you hear most is synthetic first, then human. Screen a pile of ideas with synthetic, take the survivors to a panel. It's a good pattern. It's the one Bain and most of the industry lead with. But it's one pattern, not the whole map.

I've talked to teams using synthetic research three different ways, and the differences matter, because the right one depends on what you're trying to learn and what you already have in hand.

Pattern 1: Synthetic first, then human

Use this when you have more ideas than budget. You screen wide and cheap with synthetic, then spend your real-research money on the few things that survive.

This is the concept-testing and assumption-screening case. You've got 30 concepts, or a spec full of "the user will" claims, and no way to put all of them in front of real people. Synthetic gives every one of them a grounded reaction, you kill the obvious losers and sharpen the rest, and the panel round runs on a short list instead of a long one. Bain's framing is to "screen early concepts... so human research focuses on the highest-value questions."

The boundary: synthetic does the screening, humans make the call. The deeper version of this is in when concept testing with synthetic users works and validating product assumptions before you build.

Pattern 2: Human first, then synthetic

Flip the order when you already have a little real signal and you want more coverage than you can afford to collect.

This is the one I didn't expect to hear as often as I have. A team runs five or six real interviews, learns something, and then wants more data without booking another month of recruiting. So they take the real interviews they just ran and use them as the grounding evidence for a synthetic audience. Upload the transcripts, generate personas grounded in what real people actually said, then interview the synthetic ones for breadth. The handful of humans set the direction. The synthetic layer stretches it.

It works for pressure-testing, too. You get a surprising result from a small sample and you're not sure if it's a real pattern or just those five people. Synthetic can tell you fast whether the finding holds up across a wider, evidence-grounded audience before you commit to another recruit to find out. The mechanics of turning real research into grounding are in how evidence grounding works.

The boundary: synthetic grounded in your real interviews is only as good as the interviews. A few rich conversations make a strong base. One thin one doesn't.

Pattern 3: Synthetic on a loop

The third pattern isn't a sequence with human research at all. It's using synthetic for the steady stream of small questions you'd never stand up a full study for.

Think about what it takes to ask your audience one quick question today. A survey means writing it, fielding it, waiting. Interviews mean recruiting and scheduling. So most of those small questions just don't get asked. The bar is too high for the size of the question.

Once you've built an audience in Candor, that bar drops. You generate the personas once, and after that you can open a chat and ask them something whenever it comes up, no recruiting, no scheduling, no fielding. The same personas remember what they told you in earlier sessions, so you can come back a week later and keep the thread going rather than starting cold. That's what makes ongoing, low-overhead questioning practical, and it's covered in how persona memory actually works.

Two honest limits here. There's still an upfront step: generating the audience takes real evidence work, it isn't instant. And personas are scoped to a single study, so "over time" means across the life of a research project, not a permanent panel you query forever. Inside those limits, this is the pattern the teams describing "I just want to ask a few questions without the whole production" are reaching for.

Most teams end up mixing them

These aren't three boxes you pick one of. The teams getting real value tend to move between them. Synthetic to map an area, humans to interpret what they find, synthetic again to pressure-test a surprising result, humans to go deep on the decisive one. One industry writeup describes synthetic running "before, after, and sometimes between" human studies, and that matches what I see. The sequence bends to the question, not the other way around.

What doesn't change across any of them is the boundary. Synthetic gives you direction, breadth, and speed. It doesn't give you statistically-bounded numbers, it doesn't react to a visual mockup, and it doesn't replace a real person for the calls where being wrong is expensive and hard to undo. Those still go to humans, in every pattern. I made the full case for where that line sits in what synthetic research skeptics get right.

The mistake isn't picking the wrong pattern. It's assuming there's only one, and forcing every question through it. Figure out what you already have and what you're actually deciding, and the right shape usually picks itself. See how Candor works for the mechanics, and the use cases for where each pattern tends to land. Which of these three is closest to how your team works today?

Common questions

There's no single best way; it depends on what you already have and what you're deciding. Three patterns cover most teams. Synthetic first, then human: screen many ideas cheaply with synthetic, then validate the survivors with a panel. Human first, then synthetic: run a few real interviews, then use those transcripts as grounding evidence to extend coverage with synthetic personas. Synthetic on a loop: ask quick, ongoing questions of an existing synthetic audience without the recruiting and scheduling overhead of a full study. Most teams end up mixing them, running synthetic before, after, and between human studies. The sequence should bend to the question rather than the question being forced through one fixed process.

Yes, and it's one of the strongest ways to use synthetic research. If you've run real interviews, you can upload the transcripts as grounding evidence, generate personas grounded in what real people actually said, and then interview synthetic personas for broader coverage than you could afford to recruit. The handful of real conversations set the direction; the synthetic layer stretches it. This also works for pressure-testing: when a small sample surfaces a surprising result, synthetic can quickly indicate whether the pattern holds across a wider grounded audience before you commit to another round of recruiting. The quality ceiling is set by the interviews you ground on, so a few rich conversations make a much stronger base than one thin one.

Once you've built an audience, yes. Generating the audience is a one-time, evidence-grounded step that isn't instant, but after that you can open a free-form chat and ask your personas new questions whenever they come up, with no recruiting, scheduling, or survey fielding. The personas remember earlier sessions, so you can return days later and continue the thread rather than starting cold. This makes ongoing, low-overhead questioning practical for the small questions you'd never stand up a formal study or survey to answer. One limit to know: personas are scoped to a single study, so "over time" means across the life of a research project, not a permanent panel you query indefinitely.

Whenever being wrong is expensive and hard to undo, and in a few specific cases regardless of stakes. Statistically-bounded numbers (a price-switch percentage with a confidence interval) need real-respondent sample size. Claim substantiation for regulatory, label, or advertising use needs real-respondent data. Anything that depends on reacting to a visual mockup, prototype, or interactive flow needs real users, because synthetic personas respond to text, not designs. And actual in-product behavior is captured by analytics and real usability sessions, not simulated. In every workflow pattern, synthetic handles direction, breadth, and speed, and the decisive, irreversible calls still go to humans.

Not strictly. Two of the three common patterns pair it with human research (synthetic before human, or human before synthetic), and that pairing is where most high-stakes value comes from. But the third pattern runs synthetic on its own: a steady loop of quick, lower-stakes questions against an existing audience, where the realistic alternative isn't a panel, it's not asking the question at all. The honest framing is that synthetic is an augmentation layer. It's genuinely useful standalone for fast, directional, ongoing questions, and it should not be the last word on a decision where being wrong is costly. Match the pattern to the stakes and it earns its place either way.

Candor is in development.

Be the first to know when it launches.

No spam. Just a note when Candor is ready. Powered by Highline Beta.