How the simulations work.
Every study goes through the same six steps. Nothing is hidden: every chart can be traced back to the simulated people who produced it.
- 1Describe
You say who to ask and what to ask: yes/no, multiple choice, 1 to 10, price or open questions.
- 2Calibrate
Before anyone answers, the engine looks up how this real audience answers similar questions in published surveys and market data.
- 3Recruit
A panel of distinct people is built: ages, jobs, places, incomes, politics and attitudes in realistic proportions.
- 4Answer
Each person answers every question in their own voice, anchored to the benchmark but varied by their own life.
- 5Weight
Answers are balanced to your target mix and projected to the real population size, with a margin of error.
- 6Read out
You get the split, segments, themes, risks and a recommendation, and you can trace every number to the person behind it.
Real-world calibration
Language models left alone tend to sound agreeable, which makes every idea look good. So each study starts with one calibration step: for every question, the engine estimates how this specific audience answers in reality, using published polls, adoption figures and pricing data. Every batch of respondents is anchored to that same benchmark, so a panel of 200 stays consistent from the first person to the last.
Panels that last
An audience is a saved group of simulated people. When you run a new study on it, the same people answer again, so results are comparable across questions and over time. Bigger studies add new people to the audience instead of replacing it.
Weighting and margin of error
If you set a target mix (for example by age and gender), answers are weighted so the panel matches it, then projected to the real population. Every result shows the panel size and a 95% margin of error, so a 48-person panel is never presented as if it were 10,000.
A/B tests
Two versions (prices, headlines, product ideas) are shown to matched halves of the same panel. The difference you see is caused by the version, not by who happened to be asked.
Your own customers
Paste a customer list or interview notes and the panel is built from those real people, keeping their real facts and filling in only what's missing.
Digital twins and focus groups
Any respondent can be saved as a twin who remembers what they said. You can interview them one to one, or put several in a moderated focus group where they react to and disagree with each other.
Follow-up questions
After results come in, ask the whole panel, or only the people who said no, why they answered the way they did.
What it is not
Simulated answers are a fast, cheap first read, not a replacement for a final survey of real people on a decision that matters. They are strongest on broad consumer and public-opinion questions and weaker on niche or brand-new topics with little public data.
Ten published polls, rerun on simulated people, gap shown for each.