Skip to content

Case studies

How close does a simulated panel get?

Published benchmarks of HumanPanel against real survey data. The misses are published alongside the hits, with every number tagged by where it came from.

  • Benchmark · September 2026

    HumanPanel vs. real survey data: five coffee-market questions

    Five questions from a published Verasight study with 1,519 real U.S. adults, rerun on 100 simulated respondents. How close does a cheap, fast synthetic panel get, and where does it break? Every miss is published alongside every hit.

    mean absolute error, HumanPanel
    9.6 pts
    mean absolute error, GPT-5 baseline
    23.8 pts
    results indistinguishable from real
    2 of 5
    Read the case study

Have real data we could benchmark against?

If you have published survey results and want to see how HumanPanel compares on the same questions, we would like to run it and publish the result, hits and misses included.

Get in touch