Synthetic Signals
← Platform
Interviews & surveys

Interview a synthetic audience.

Survey the whole city, run a focus group, or concept-test copy and pricing on thousands of Census-grounded synthetic users — answers broken down by cohort, in minutes. No agent required.

Survey · 1,000 respondentscohort split
“Would you pay $12/mo for the premium plan?”
answered yes
Overall
34%
Age 18–34
52%
Age 35–64
31%
Age 65+
12%
Spanish-pref
22%
Every cohort, not an average — in minutes
Minutesfrom question to cohort-split answers
1k–5ksynthetic respondents per audience
By cohortanswers broken down, not averaged
No agentinterviews run straight from the app
Surveys

Survey the whole city at once

Ask one question — or a structured set — across the entire audience, or any cohort of it. A thousand Census-grounded synthetic users answer in character, in minutes, and the answers come back as distributions you can slice, not a wall of transcripts.

  • One question × 1,000 respondents, in parallel
  • Target the whole city or a single cohort
  • Structured answers, aggregated automatically
Survey · whole cityrunning
Asked
Answered
Aggregated
1,000 answers · minutes, not weeks
Focus groups

Run a focus group without recruiting one

Pull a handful of demographically distinct citizens into a moderated session and dig into the why behind the numbers. The skeptic, the budget-conscious retiree and the early adopter push back in their own voices — the texture a survey can't give you.

  • Moderated sessions with the citizens you choose
  • Demographically distinct voices, in character
  • Full transcripts — quote them in your report
Focus group · 6 citizensmoderated
What would make you switch pharmacies?
Grace W · 71 · ESLIf someone explain my refills in plain words.
Marcus M · 34Honestly? Whoever texts me when it's ready.
Priya N · 58Price. I compare every single time.
Concept tests

Test the change before you build it

Show the audience a concept — pricing, copy, a feature description, a positioning line — and collect reactions before you commit a sprint to it. Compare variants head-to-head and see which cohorts each one wins, and which it loses.

  • Copy, pricing, features — test any stimulus
  • Compare variants head-to-head
  • See which cohorts a change wins — and loses
Concept test · pricing copyA vs B
“Pro plan — $12/mo, cancel anytime.”
A · “cancel anytime”
58%
B · “no contract”
42%
65+ prefers A two-to-one — reassurance wins
The report

Answers broken down by cohort

Every interview ends in the same cohort lens that powers agent test coverage: answer distributions split by age, income, language and life situation, with quotes to back the numbers. And the confidence caveats are baked in — a synthetic panel is directional signal, not a substitute for a properly powered human study, and the report says so.

  • Cohort splits: age, income, language, life situation
  • Distributions and quotes, not a black-box score
  • Honest-limits caveats baked into every report
Survey · 1,000 respondentscohort split
“Would you pay $12/mo for the premium plan?”
answered yes
Overall
34%
Age 18–34
52%
Age 35–64
31%
Age 65+
12%
Spanish-pref
22%
Every cohort, not an average — in minutes
How it works

From a question to a cohort-split answer.

No panels to recruit, no incentives to pay, no three-week field time — the audience is already there.

1
Pick the audience

The whole city, a saved audience, or a single cohort — Census-grounded either way.

2
Ask your question

A survey, a focus-group topic, or a concept to react to.

3
They answer in character

Each citizen replies from their own demographics, personality and memory.

4
Read the cohort report

Distributions split by cohort, with quotes to back the numbers.

Synthetic respondents are directional signal — fast, cheap and honest about being synthetic. Validate the bet-the-company decisions with real humans.

Questions

When to trust a synthetic panel — and when not to.

When should I trust a synthetic panel — and when not?

Use it for direction: ranking options, killing obvious losers, spotting which cohorts react differently, and sharpening the questions you'll take to real users. Don't use it as a substitute for a properly powered human study on a bet-the-company decision — every report carries that caveat on purpose.

Are these real survey respondents?

No — every respondent is fully synthetic, generated from public, aggregate Census data. That's the point: no recruiting, no professional survey-takers, no PII, and answers in minutes instead of weeks.

How can an LLM answer for a 71-year-old retiree?

Each citizen is driven by an LLM conditioned on that citizen's demographics, Big Five personality, current context and memory — so answers are generated in character and vary across the population the way real answers do. Research on generative agents shows usefully human-like response patterns; it still isn't a human, which is why the caveat stays on every report.

Do I need to connect an agent to run interviews?

No. Interviews run from the app against the audience you built. Connecting your agent over MCP or REST is only for the agent-testing side of the platform.

Can I re-interview the same people later?

Yes. Citizens keep memory across sessions, so you can run longitudinal research — ask the same panel next week, after a pricing or message change, and see what moved.

How is this different from asking a chatbot to role-play users?

One prompt role-playing "a user" gives you one averaged voice. Here every answer comes from a distinct, demographically correct person in a Census-grounded population — so you get distributions and cohort differences, not vibes. And the same panel comes back next run, so results are reproducible.

Ask a thousand people — before you build it.

Build a Census-grounded audience and get cohort-split answers in minutes — no recruiting, no field time.