Guide
Synthetic respondents: when to use them and when not to
TL;DR
- A synthetic respondent is a simulated person who answers questions the way someone in your target audience would. A synthetic panel is a group of them.
- Neon Blue builds simulated audiences: AI models that answer the way a defined group of people would, grounded in Neon Blue's own paid studies and AI-moderated interviews with real people.
- They're most useful for decisions you can't test live or can't get enough people for: a fund's thin survey cell, a low-traffic pricing page, a B2B message before it goes to a list, twelve concepts with budget for three.
- For final calls, new categories and sensitive topics, confirm with real people.
- Before you rely on one, find out what it's grounded in and how far your questions sit from that data.
When should you use synthetic respondents, and when not?
| Job | Who usually asks | Fit |
|---|---|---|
| Fill thin cells in a consumer survey | Fund analyst | Good with care: label it and check it against held-back real responses |
| Pick a price range or paywall design before a live test | Product or growth team | Good: finds the range worth testing |
| Test value props on a synthetic ICP panel before outbound | B2B GTM team | Good: no list gets burned |
| Screen 10 to 30 concepts or claims to a shortlist | Enterprise marketing or product lead | Good: rank order matters more than exact levels |
| Final go/no-go on a major launch | Anyone | Poor alone: confirm with real people |
| A brand-new category | Anyone | Poor: nothing to ground in |
| Health, money or other sensitive behavior | Anyone | Poor alone: stated answers differ in private |
| Legal, regulatory or claims substantiation | Anyone | No: needs real respondents |
The poor-fit rows above need a real check. Before you rely on a simulated read there, compare it with real people on those questions.
How do you mix synthetic and real respondents?
Screen, then confirm. Run all options on a synthetic panel and take the top few to real people with the same questions. It's the cheapest way to cover a wide set of ideas.
Real core, synthetic boost. Field a real sample, then use a simulated audience for the segments you couldn't afford to fill. Keep the boost labeled and reported separately, and hold back a few real respondents in each thin cell to check it against.
Synthetic pre-read on the question set. Before a real study or a live test, run your questions on a synthetic panel. If everyone gives the same answer, the question is probably leading. If open-ended answers wander, it was probably misread.
What should you ask before relying on synthetic respondents?
- What are the respondents grounded in: interviews, surveys, public data, or a prompt?
- Who are the people behind the grounding data, and how recent is it?
- Can you show how far my questions sit from that data?
- What happens when a question is out of range? Is it flagged or answered anyway?
- Can I see raw, per-respondent output as well as the summary?
- Can I run a paired test against real respondents on my own questions?
- Is simulated output always labeled and kept separate from real responses?
What does a synthetic respondent test look like for a B2B message?
Hypothetical: a GTM team at a workflow software company is moving upmarket and has four value props for operations leaders at 500 to 2,000 person companies. Sending all four to the live list would burn it.
- Build a synthetic ICP panel: heads of operations and their direct reports in that size band, split by industry.
- Show each value prop as the first two lines of an email, with the same subject line.
- Ask how relevant it is to a problem they have this quarter (5-point scale), what they'd need to believe to take a meeting, and which current tool or process it would replace.
- Rank the four by relevance, and read the "need to believe" answers for objections the sales deck doesn't handle yet.
- Send the top two to a small slice of the real list and compare reply rates before rolling out.
Want this run on your own question? Neon Blue's first study is free: start a free first study.
How does Neon Blue run synthetic respondents?
Neon Blue runs real and synthetic in one place: AI-moderated interviews with real people, and simulated audiences grounded in those interviews and its own paid studies. Simulated results are always labeled. Per-question confidence scoring is in build. For how a simulated audience compares with a traditional panel, see synthetic panels vs traditional panels.
Common questions about synthetic respondents
Do synthetic panels replace real respondents?
They decide where to spend real sample and cover questions you'd otherwise skip. For decisions that are expensive to get wrong, confirm with real people.
Are synthetic respondents the same as survey bots or fraud?
No. Fraudulent respondents pretend to be real people inside a real panel. Synthetic respondents are openly simulated and labeled as such.
How many synthetic respondents do I need?
Adding simulated respondents doesn't add information beyond the grounding data. Check how well your segments are covered in that data instead.
Can I mix synthetic and real responses in one number?
Report them side by side. Blending hides which part of the result came from real people.
Can I build a synthetic panel of B2B buyers?
Yes, if the grounding interviews cover that role and company size. Narrow B2B roles are where coverage matters most, so check it first.
How do I start a free first study?
Bring the decision you'd otherwise test on a live list or skip. Neon Blue's first study is free. Start a free first study