Seeful

AI agents matched real humans' survey answers at 86% accuracy — on par with people re-answering the same questions themselves two weeks later.

Researchers at Stanford HAI, Google DeepMind, and Northwestern University built generative agents from 2-hour interviews with 1,052 real Americans, then tested them against held-out General Social Survey items. The agents' responses aligned with participants' own re-test scores — suggesting simulated audiences can genuinely replicate how real people think.

LLMs reproduced consumer purchase intent at 90% of human test-retest reliability, maintaining realistic response distributions across 57 product surveys.

Researchers at Stony Brook University and PyMC Labs introduced a semantic similarity method that elicits rich LLM text responses and maps them to consumer rating scales. Tested on 57 personal care product surveys with 9,300 real human respondents, the synthetic panels closely replicated human rating distributions — opening a scalable alternative to traditional consumer panels.

AI personas replicated 76% of published experimental findings from the Journal of Marketing — across 133 studies and 19,447 simulated consumers.

Researchers at Stanford University and Boston University created LLM-based AI personas and tested whether they would reproduce established effects from 133 experiments published in the Journal of Marketing. The 19,447 simulated consumers successfully replicated 76% of the main effects — providing strong independent evidence that AI audiences mirror real human behavioral responses.