Skip to content

Methodology

How InstantFocus generates, calibrates, and evaluates synthetic consumer panels. Every design decision is grounded in published research.

Three-stage pipeline

Three stages, executed per-request. No pre-generated data. Every panel is fresh.

01

Panel generation

Synthetic personas are generated with census-aligned demographics and IPIP-NEO Big Five personality traits. Each persona is a unique, internally consistent profile.

02

Per-persona evaluation

Each persona independently evaluates your stimulus via an LLM call, responding as their calibrated profile. Batched 10/batch with 5 concurrent batches.

03

Statistical aggregation

Individual responses are aggregated into distributions, scores, themes, and representative verbatims. Study-type-specific post-processing produces the final structured output.

IPIP-NEO Big Five

Each synthetic persona is assigned a Big Five personality profile — Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism — sampled from population-level norms.

Data source

IPIP-NEO normative dataset (Srivastava, John, Gosling & Potter, 2003). N=132,515 across ages 10–80+. Provides age-stratified means and standard deviations for each Big Five trait and its six sub-facets.

Calibration process
  • Look up the age-stratified mean and standard deviation for each Big Five trait from the IPIP-NEO normative table
  • Apply gender adjustments for Agreeableness (+0.5 SD for women) and Neuroticism (+0.4 SD for women) based on published sex differences
  • Sample each trait from a truncated normal distribution (clamped to 1–100 percentile range)
  • Enforce internal consistency: e.g., high Openness + Innovator tech adoption are positively correlated
  • Store the full personality vector as part of the persona profile for use in evaluation prompting

Big Five traits

TraitHigh endLow endRelevance
OpennessCreative, curious, novelty-seekingConventional, practical, routine-orientedDrives receptivity to new products and concepts
ConscientiousnessOrganized, detail-oriented, disciplinedFlexible, spontaneous, less structuredAffects price sensitivity and quality expectations
ExtraversionOutgoing, enthusiastic, assertiveReserved, independent, introspectiveInfluences social proof sensitivity and sharing intent
AgreeablenessCooperative, trusting, empatheticCompetitive, skeptical, directDrives response positivity bias — important to account for
NeuroticismAnxious, risk-averse, emotionally reactiveCalm, resilient, emotionally stableAffects risk perception and concern identification

Census-aligned distributions

Age

Distribution from U.S. Census American Community Survey (2023). Age-adjusted personality norms applied after demographic assignment.

Gender

Male/Female/Nonbinary distribution with gender-adjusted personality trait offsets for Agreeableness and Neuroticism.

Income

Five brackets mapped to Census income percentiles: low (<$30K), middle ($30–60K), upper-middle ($60–120K), high ($120K+).

Education

Three levels from Current Population Survey: high school, bachelor's degree, graduate/professional degree.

Region

US regional distributions (Northeast, Midwest, South, West) with planned expansion to UK, EU, and global panels.

Tech adoption

Rogers' diffusion curve: 2.5% innovators, 13.5% early adopters, 34% early/late majority, 16% laggards. Correlated with age and Openness.

LLM-powered response generation

Model

OpenAI gpt-4o-mini with structured JSON output mode. Selected for speed/cost balance at scale — full panel of 100 completes in ~3–5 seconds.

Temperature

0.7 — produces meaningful response diversity within persona constraints while maintaining coherence. Lower temperatures tested but produced too-uniform panels.

Prompt architecture

Study-type-specific system prompts inject the full persona profile (demographics + Big Five vector) as character context. The stimulus is presented as user content.

Batching

10 personas per batch, 5 concurrent batches. Balances throughput against rate limits. Full 100-persona panel completes in ~10 parallel rounds.

What this is not

  • Synthetic panels are not a replacement for real consumer research. They provide directional signal, not statistical proof.
  • LLM-generated responses may reflect training data biases that do not match real population sentiment on specific topics.
  • Published research shows 75–85% directional accuracy against real survey data (NORC, NNGroup). This means 15–25% of the time, the synthetic panel will disagree with reality.
  • Cultural nuance, regional dialect, and highly specialized domain knowledge are areas where LLM simulation is weakest.
  • Results should be treated as hypothesis generators, not hypothesis validators. Always follow up with real research for high-stakes decisions.