Lewsearch

Lewsearch is a synthetic research panel calibrated to U.S. Census and ACS demographic targets. Studies field through the Lewis model, a proprietary fine-tune with city-level granularity. Lewsearch has been tested against 460 ground-truth public polling questions sourced from Pew Research Center, Gallup, the University of Texas / Texas Politics Project, and PPIC. After calibration, pooled MAE is 7.47 percentage points across the full benchmark set, 4.88pp on the Texas UT/TPP panel, and 0.40pp on the best single question. Critically, Lewsearch is not provided with published toplines during fielding, the synthetic panel generates the response distribution independently, then is scored against the ground truth post-hoc. Lewsearch publishes timestamped forecasts against future recurring public surveys. The current release includes 28 frozen forecasts against the next comparable Pew Research Center waves on AI attitudes and top national problems. Each forecast is GitHub-timestamped, baseline-anchored to the most recent Pew publication, and will be publicly audited when comparable future waves release. See lewsearch.com/predictions. Use Cases - concept and message screening, early-stage idea filtering, brand-perception hypothesis testing, message resonance triangulation, sensitive-topic exploration without PII collection, and segment-level directional reads. Lewsearch is designed as a fast pre-fieldwork layer in a hybrid synthetic-plus-human research workflow, not as a replacement for probability sampling. Each study returns a 10-12 page PDF including topline distributions, demographic crosstabs, respondent voice, and an analyst decision brief. Free Sandbox+ studies (n=200) return in seconds; full studies (up to n=10,000) return within minutes to one hour depending on configuration. Self-serve access without demo gating. Methodology published at lewsearch.com/methodology. Built by Swarmgram, Inc. (DE), HQ in Columbus OH.