Methodology
Research you can audit.
Seven principles behind every study. Enough to judge the method, and to see why it holds up.
01Grounded
Built on official statistics.
Every synthetic panel starts from official population statistics and is quota-matched to your target group, so it looks like the people who actually live in your market.
02Deep
Every respondent gets a life story first.
Before a synthetic consumer sees your study, it gets a complete life story: daily life, household, habits, money and values. By the time your concept arrives, it has a history to answer from. The approach builds on research showing that life stories capture people considerably better than demographic profiles (Park et al., 2026).
03Checked
Every respondent passes a quality gate.
An independent check reviews each respondent for consistency and filters near-duplicates. Respondents that fail are rebuilt or dropped.
04Real
Digital twins of real people.
When a study needs real people behind the answers, digital twins respond. They represent real consumers, built from in-depth personal interviews, and can be asked again and again across many studies. In published research, agents grounded in such interviews simulated real individuals considerably more accurately than demographic profiles (Park et al., 2026).
05Measured
The number comes from the answer.
Ask a language model for a rating from 1 to 5 and you get suspiciously tidy numbers. So we don't. Respondents answer in their own words, and each answer is placed on the scale by its semantic similarity to reference statements, one for each point on the scale. In published research, this method, Semantic Similarity Rating, reaches 90% of human test–retest reliability (Maier et al., 2025).
06Evidenced
No hallucinations in reports.
Findings are built up from the answers. Every number links to the answers behind it, and every quote is verified verbatim by code. Claims without evidence don't make it into the report.
07Transparent
Labelled for what it is.
Every result shows which panel it rests on and how it came about. So you can weigh it correctly, stand behind it in front of your team and compare studies months later.
08Semantic Similarity Rating
How an answer becomes a number.
Asked for a rating, language models keep landing on the same few values, and the spread of real opinion gets lost. Semantic Similarity Rating keeps it. In published research, the method reaches 90% of human test–retest reliability (Maier et al., 2025).
01
Answer in their own words
Each respondent answers the scale question in free text, for example: “I’m somewhat interested. If it works well and isn’t too expensive, I might give it a try.” Nobody is asked for a number.
02
Compare with reference statements
Every point on the scale has a reference statement, from “definitely not” to “definitely yes”. The answer and the statements are turned into embeddings, numerical representations of meaning, and compared. Each question uses six differently worded sets, so no single wording decides.
03
Similarity becomes a distribution
The closer the answer is in meaning to a statement, the more weight that point on the scale gets. Averaged over all six sets, every answer yields a distribution across the whole scale instead of one forced value.
04
Distribution becomes a result
Each answer gets a value on the scale, drawn reproducibly from its distribution. Top-2-box, mean and results per segment are then calculated in code, and every figure links to the answers behind it.
09Open answers
How we code open answers.
Open answers show why people react the way they do. We turn them into themes you can count and check, and every theme leads back to the exact words behind it.
01
Follow-up questions
When a first answer leaves something open, the study asks up to two follow-up questions, guided by the study’s probing instructions. Once an answer is clear, it stops.
02
Findings with evidence
A model reads the answers and records findings. Each finding has to point to at least one specific answer. Findings that merely repeat a quote are discarded.
03
Themes from the answers
Findings are grouped into up to six themes per question. The themes emerge from the answers themselves, without a predefined codebook. How often a theme occurs is counted in code: the number of distinct answers behind it.
04
Verified quotes
Every quote is cut from the stored answer again by code, word for word. A theme only makes it into the report when every finding in it has a verified quote.
10Research
The research behind the method.
Our methodology builds on two studies. Both are publicly available, so you can check the foundations yourself.
Park et al., 2026
LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals
Joon Sung Park, Carolyn Q. Zou, Jonne Kamphorst et al.
Agents grounded in two-hour interviews about people’s lives reproduced the attitudes and behaviour of 1,052 real individuals considerably more accurately than agents built from demographics alone.
The basis for life stories and digital twins
Read the paper on arXivMaier et al., 2025
LLMs Reproduce Human Purchase Intent via Semantic Similarity Elicitation of Likert Ratings
Benjamin F. Maier et al.
Free-text answers mapped onto a rating scale by semantic similarity reach 90% of human test–retest reliability (Maier et al., 2025), with realistic response distributions.
The basis for Semantic Similarity Rating
Read the paper on arXiv
11Why it's different
Statistics tell you who. Interviews tell you why.
Most synthetic research builds a respondent from a statistical profile and a prompt. That mostly reproduces what the statistics already know. With us, every respondent has a life story before it answers. Research backs this: agents grounded in interviews about people’s lives simulate individuals considerably more accurately than agents built from demographics alone (Park et al., 2026).
| Traditional research | Generic AI | Statistical synthetic panels | theconsumer.ai | |
|---|---|---|---|---|
| Time to result | 2–6 weeks | Seconds | Minutes | Minutes |
| Grounded in | Real respondents | Whatever the model learned | Statistical profile and prompt | Official statistics and a life story per respondent |
| Depth per respondent | High, but costly | None | A row of attributes | A coherent life story |
| Real people behind it | Yes | No | No | Optional, with digital twins |
| Scale answers | Measured | Picked by the model | Often picked by the model | Measured with SSR |
| Traceability | Report | None | Summary | Every number and quote linked to its answers |
Comparison of typical approaches, not of individual providers.
See the method on your own category.
We'll walk you through a real study, from the panel to the evidence.