Accuracy & honesty
We don't look up answers. We simulate people.
Ask a model “what's the president's approval?” and it will recall the poll for you. Go ahead, that's not our job. But “will people try this flavor?”, “does this tagline land with parents in Texas?”, “would you pay $9 for this?” have nothing to look up, because they haven't happened yet. That's what lewn is for. It simulates how a real population reacts, before you spend a dollar to find out.
Ask one AI
You get one answer, a single averaged guess. Tell it to “pretend to be 100 people” and it still collapses to that average. You never see who disagrees, or how strongly.
Ask a crowd with lewn
You get a distribution: hundreds of distinct, demographically grounded respondents, and how they actually split. The 18% who'd hate your tagline show up next to the 60% who'd love it.
How we know the simulated people are real enough to trust
We test where the truth is known but can't be looked up. Asked to estimate a specific city's approval of its mayor, a top model (Claude Opus, no search) misses by 17 to 45 points, because it is guessing. lewn simulates the actual residents and lands within about 10. Versus real human surveys, it averages ≈7.5% average error across 460 benchmark questions (10.7% on a fully held-out set).
For anything already public, a model with search will beat us, and you should use it. lewn earns its keep on the questions no one has answered yet: your product, your price, your customers. It doesn't try to remember the answer, it simulates the people and reads how they react.
Grounded
Each respondent reflects a real slice of the population (age, gender, politics, income, place), so “liberals in Columbus” really is liberals in Columbus.
Benchmarked
The underlying model is checked against Pew, Gallup, and academic panels — ≈7.5% average error across 460 benchmark questions (10.7% on a fully held-out set) vs. real human surveys.
Honest
Every result shows its confidence (calibrated panel vs. directional read) and flags when a thin segment is topped up with synthetic respondents.
What lewn is, and isn't
lewn is a fast, calibrated gut-check. Read a crowd in seconds before you spend on media, fielding, or a build. It's not a replacement for a full human study when the stakes are high. Synthetic respondents are estimates, not guarantees, and the confidence on every result tells you how much to lean on it.