hugo
    back to site
    Validation/Simulations/Open accuracy log

    How accurate are Hugo's simulated polls?

    This page validates Hugo Simulations, the poll feature that asks a simulated panel a question and returns a distribution of answers. It does not cover Hugo's research feature, which is evaluated separately.

    Every release of the simulator is benchmarked against published real-world surveys it has never seen. We update this page with every run.

    Headline metric

    71%
    got right
    (top answer)
    Questions
    70
    held-out, never seen
    Audiences
    19
    11 countries · 5 languages
    Mean JSD
    0.039
    distribution distance · 0 = identical
    Run
    June 29, 2026
    updated every release
    Across 70 held-out questions covering 19 audiences in 11 countries, Hugo gets the top answer right 71% of the time and the full distribution matches reality to within 0.039 JSD on average. Hugo tells you what an audience will pick and how strongly they lean, and it's sharpest where it matters most for commercial decisions: brand and product taste, and behavior you can read off content.
    Metric 1: Got right

    A simple yes/no per question: did Hugo pick the same #1 answer as the real survey? Only looks at the top choice.

    Shown on the headline above, in the Scorecard (questions Hugo got right per audience), and on each question in the All questions section as a green check or red x.

    Metric 2: JSD

    Jensen–Shannon divergence: how different the full answer distributions are. 0 = identical, 1 = opposite. Under 0.1 is a tight match.

    Got right only checks the winner; JSD compares the whole shape, so Hugo can miss the #1 but still land close if the split is similar, or get #1 right but be off on the rest.

    ·

    How close is close?

    A second view: how far each audience lands from reality on the 0 → 1 distribution scale

    Hugo
    0
    0.025
    0.050
    0.10
    0.15
    0.20
    0.50
    1.00
    0 · identical to reality1 · opposite
    Per audience
    Brazilian evangelical youth
    0.010
    US Gen-Z gamers
    0.012
    Swedish climate youth
    0.021
    US outdoor & hiking enthusiasts
    0.024
    Brazilian fitness women
    0.026
    Swedish Gen-Z
    0.026
    US Gen-Z retail investors
    0.028
    Finnish Gen-Z
    0.028
    Excellent[0.00, 0.05)

    As close as two real surveys would differ

    Strong[0.05, 0.10)

    Directionally tight

    Useful[0.10, 0.20)

    Right shape, soft on magnitudes

    Weak[0.20, 0.40)

    Trends only

    Poor[0.40, 1.00]

    Don't trust

    01

    Scorecard

    Audiences from the 19-audience run, best to worst

    AudienceAccount responsesQuestionsJSDGot right
    Brazilian evangelical youth
    Brazil · faith-driven Gen-Z
    10920.0102/2
    US Gen-Z gamers
    United States · ages 13–25
    6140.0123/4
    Swedish climate youth
    Sweden · climate-engaged
    10540.0212/4
    US outdoor & hiking enthusiasts
    United States · outdoors community
    10230.0241/3
    Brazilian fitness women
    Brazil · body-positive fitness
    9530.0263/3
    Swedish Gen-Z
    Sweden · 16–25
    70120.0268/12
    US Gen-Z retail investors
    United States · personal finance · 18–29
    6330.0282/3
    Finnish Gen-Z
    Finland · 16–30
    6040.0283/4
    Mean700.03950/70
    02

    Findings

    What it gets right, where it's still developing

    Good at

    • Brand and product taste. The core commercial use case (US beauty: 100% of favorite-brand winners correct across fragrance, skincare, makeup, retailer).
    • Content-legible behavior. Gaming habits, platform-of-origin usage, exercise.
    • Attitudes, values, beliefs. Vote intention, EU impact, societal outlook, religion, mental-health prevalence.
    • Generalizing across geography and language. Finland (Finnish-seeded) JSD 0.028, on par with US audiences.

    Temporal limitations

    • Off-content options
      Personas built from a single seed channel can over-index that channel when asked about cross-platform behavior.
    • Niche posted behavior
      Communities built around a single interest can over-represent that interest in their day-to-day reporting.
    • Flattens magnitudes
      Hugo tends to pick the right #1 answer but understates how dominant it is, and defaults attitudes toward the middle of the scale.
    Trust
    • · Brand, product, content and taste preference
    • · Attitude and value direction
    • · Which option an audience leans toward
    • · Rank-ordering ideas and concepts
    Use with caution
    • · Which platform or channel an audience uses
    • · Absolute prevalence of off-content behavior
    • · Exact percentages. The ranking is more reliable than the spread
    03

    All questions

    Real vs simulated, per audience

    04

    Method

    For each audience, Hugo builds a panel of 50–500 simulated personas from public social content and has them answer the survey question. We then compare Hugo's answers to a real, published survey.

    Each persona is grounded in a public account's actual activity, what they post, what they engage with, the language they use. Hugo receives that profile and answers the way the person plausibly would.

    Hugo never sees the real numbers. Every survey is verified against its original source and linked on each question, so the comparison can be audited end-to-end.

    Audiences are chosen to stress-test the system across geographies, languages, and interests, from US gamers and beauty shoppers to Kenyan Gen Z, Brazilian fitness women, and Finnish students.

    Most of these surveys were published after the model's training cutoff, so Hugo can't have memorized them. Recall doesn't explain the accuracy.

    Bottom line

    Tested on 70 held-out questions across 19 audiences in 11 countries. Hugo tells you what an audience will pick and how strongly they lean, in minutes instead of weeks.

    05

    Questions about the method

    Run your own

    Want to validate Hugo on your customer?

    Bring a survey you trust. We'll run the same held-out test on your audience, live, in the demo.

    Book a demo