Documentation
How Reynard AIreally works.
A complete explanation of the technology behind the US population simulation — from Census Bureau data to AI reaction.
This page describes the United States panel and its sources (US Census Bureau data, eight US value segments, the nine census divisions). A separate Dutch panel follows the same method, but is weighted to CBS statistics and the Motivaction Mentality segmentation.
How AI reactions cluster
Each dot is a simulated panel member from the US panel. Live example: reactions cluster into positive, neutral or negative
What is Reynard AI?
Reynard AI is not a survey. You don't send out forms and wait weeks for panel responses. Instead, we simulate how the US population reacts to your question — within minutes, weighted to the true demographic distribution from Census Bureau data.
At the core of the system is a pool of 1,020,001 simulated Americans, built from the most recent Census Bureau microdata. That pool forms a statistical mirror of the country: age, education, division, income, household type, occupation — all in the correct proportions.
Reynard combines that population data with large language models to generate an authentic, in-character reaction for each demographic segment. The result: weighted sentiment, verbatim quotes and adoption intent — representative of the whole country or a specific target audience.
1,020,001 simulated Americans
The foundation of everything is the persona_pool: a table of exactly 1,020,001 records, each representative of a real group of people. The records aren't made up — they're built using Iterative Proportional Fitting (IPF), a statistical technique also used by national statistics agencies to calibrate census data.
Each record includes: age, gender, education, income, occupation category, household type, living situation, state, census division and a value segment. The distribution checks out: if 15.2% of the country belongs to the Heartland Traditionalists, then exactly 15.2% of our pool is Heartland Traditionalists.
| Name | Age | Division | Value segment | Education | Income |
|---|---|---|---|---|---|
| Ray T. | 58 | East South Central | Heartland Traditionalists | High school | Median |
| Maya E. | 29 | Pacific | Progressive Urbanites | Bachelor's | Median+ |
| Alicia R. | 44 | West South Central | Community Anchors | Some college | Median |
| Warren S. | 71 | New England | Market Optimists | Graduate | Affluent |
| Devon M. | 22 | Middle Atlantic | Striving Climbers | Bachelor's | Tight |
The 8 American value segments
The United States is not one audience. Reynard works with eight value segments built from US Census Bureau demographics and national values research, so a result is never just an average of everyone.
Progressive Urbanites
Educated, socially liberal, city-dwelling. Values-driven buyers who research before they commit and notice when a claim doesn't hold up.
Community Anchors
Rooted in local life — church, school, neighborhood. Trust word of mouth over advertising and reward brands that show up locally.
Heartland Traditionalists
Small-town and rural, self-reliant, cautious about change. Loyal to what has worked, slow to switch, skeptical of novelty for its own sake.
Comfort Seekers
Suburban and pragmatic. Take the easy option, avoid friction, and will pay a little more for something that simply works.
Market Optimists
Business-minded and upbeat about the economy. Comfortable with new products and services, and quick to see the upside.
Stretched Survivors
Working hard on a tight budget. Price is the first filter and the last one; any increase lands immediately and is felt.
Striving Climbers
Ambitious, career-focused, status-aware. Buy signals of progress and expect quality to match the price.
Independent Skeptics
Self-reliant and wary of institutions, corporations and messaging. Have to be convinced, and won't be rushed.
The percentages are derived from US Census Bureau data and are updated whenever new estimates are published.
Segment cells: the intersection of everything
The million-plus people aren't queried one by one — that would be too expensive and too slow. Instead, Reynard works with segment cells: unique combinations of value segment × life stage × census division × education × income. Each cell carries its own population weight — the percentage of the country that combination represents.
Community Anchors × Family Building × South Atlantic × Some college × Median
Heartland Traditionalists × Retired × East South Central × High school × Median
Progressive Urbanites × Young Adult × Pacific × Bachelor's × Tight
For each simulation, Reynard selects the most representative cells (8 to 60, depending on the plan), guaranteed to include representation from all 8 value segments and all 9 census divisions.
The simulation process step by step
Scenario input
You describe your product, price change, campaign or policy measure in plain language. Optionally, you can add extra context.
Example: "We're launching a mobile strategy game called CIPHER. The basic version is free, but full access costs $4.99 per month."
Cell selection & diversity guarantee
The system selects the most representative segment cells. Proportional to Census Bureau data: value segments with a larger share of the population get more slots. At the same time, an algorithm guarantees that all 9 census divisions and all 8 value segments are represented — so no part of the country is missing from the report.
Identity construction (deterministic)
A representative identity is assembled per cell: name, age, city. Fully deterministic based on cell attributes, never generated by the AI. Guardrail A1: the AI never invents a person.
Example: Maya Ellison, 27, Seattle — Progressive Urbanites, college-educated, tight income
AI reaction per cell
For each cell, Reynard sends an extensive system prompt to the language model with the full demographic profile, behavioral guidance and calibration instructions. The AI generates: initial reaction, concrete behavior, a verbatim quote, sentiment (−1 to +1), adoption intent (yes/maybe/no) and top concerns & benefits.
Calibration that's always included: "Most Americans are indifferent to small changes. Behavior lags behind opinion. Weigh the magnitude of the change."
Magnitude-echo check
All AI output is automatically checked: does the AI mention an amount or percentage that isn't in the scenario? Then the cell is re-run (max 1 retry). If it still fails, the output is flagged as
value_driftand counted in the statistics, but no quotes are cited.Weighted aggregation
The results from all cells are weighted by population share from Census Bureau data. A cell representing 2% of the country carries more weight than a cell representing 0.3%. The end result: percentages that are representative of the whole United States (or your target audience). The analysis summary is generated by Claude Sonnet.
Following up with a respondent
A number tells you what people think, not why. When a reaction catches your eye, you can put a follow-up question straight to that respondent. They answer in their own register and stay consistent with what they said earlier — amounts and percentages are taken verbatim from your scenario. How many respondents you can reach depends on your plan: Fast gives you 1 respondent with 1 answer, Extensive 3 respondents with 1 answer each, and In-depth 3 respondents with 3 answers each so you can genuinely dig in. On their final answer the respondent closes off the conversation themselves. On the free plan this feature is locked.
Example: "You say you'd stop shopping downtown — what would actually win you back?"
The 9 census divisions
Reynard works with the nine official US Census Bureau divisions. Every simulation guarantees at least one representative per division.
- 1. South AtlanticAtlanta, Miami, Washington DC, Charlotte
- 2. PacificLos Angeles, San Francisco, Seattle, Portland
- 3. East North CentralChicago, Detroit, Columbus, Indianapolis
- 4. Middle AtlanticNew York, Philadelphia, Pittsburgh
- 5. West South CentralHouston, Dallas, San Antonio, New Orleans
- 6. MountainPhoenix, Denver, Las Vegas, Salt Lake City
- 7. West North CentralMinneapolis, St. Louis, Kansas City, Omaha
- 8. East South CentralNashville, Memphis, Louisville, Birmingham
- 9. New EnglandBoston, Providence, Hartford
What's in the report?
Every simulation includes
- → Weighted sentiment score (−1 to +1)
- → Positive / Neutral / Negative breakdown (%)
- → Adoption intent: yes / maybe / no (%)
- → Market Readiness Score (1–10)
- → Top 8 concerns & top 8 benefits (weighted)
- → Executive summary (3–5 sentence analysis)
- → 4 concrete recommendations for the US market
- → Verbatim quotes per value segment (name, age, city)
- → Segment breakdown per value segment (% positive per group)
Extra with In-depth
- Alternative scenario: the system automatically generates an adjusted version of your scenario addressing the biggest concern, and tests it on 16 representative cells.
- Social round: all participants hear each other's opinions and can revise theirs — measuring opinion change under social pressure.
The four plans
| Free | Fast | Extensive | In-depth | |
|---|---|---|---|---|
| Segment cells | 8 | 20 | 40 | 60 |
| All 8 value segments | ✓ | ✓ | ✓ | ✓ |
| All 9 census divisions | Basic | ✓ | ✓ | ✓ |
| Executive summary | — | ✓ | ✓ | ✓ |
| Alternative scenario | — | — | — | ✓ |
| Social round | — | — | — | ✓ |
| Speed | ~30s | ~1 min | ~2 min | ~4 min |
| Price | $0 | $12 | $25 | $39 |
Example scenario: CIPHER mobile game
"We're launching CIPHER, a mobile strategy game. The basic version is free, but full access costs $4.99 per month. The game targets players who love puzzles and strategy."
"I don't get why it can't just be free, there are already so many apps like this."
"Finally something that isn't purely reflex-based."
Reynard automatically generated: "CIPHER is free with an optional $2.99/month for extra levels."
→ Adoption intent rises from 22% to 31% yes. Resistance drops from 25% to 18%.
Safeguards & reliability
Identity from database
Names, ages and cities are deterministically assigned from fixed lists based on cell attributes. The AI never invents a person.
Magnitude-echo check
Every percentage or amount in the AI output is compared against the scenario. Does the AI introduce a number that isn't there? Automatic retry. If it still fails after one retry it is flagged as value_drift and not quoted.
Language check
All output is in US English. An automatic heuristic detects off-language output and retries the cell.
No inflated claims
'1,020,001 simulated Americans' refers to the size of the persona pool — the statistical foundation. The actual number of AI calls per simulation is 8 to 60, depending on the plan. We never overstate.
Frequently asked questions
No. A survey asks real people. Reynard simulates based on demographic segments calibrated against US Census Bureau data. It's faster and cheaper, but not a replacement for qualitative research on niche audiences. Reynard is strongest for broad consumer questions and market choices.
Ready to gauge the market?
Start with a free simulation and see how your idea lands in 30 seconds.