Aaru

Aaru builds simulated populations for companies to test decisions on.¹ Its site says it "grounds large-scale populations of AI agents in real demographic, behavioral, and outcomes data", so a client can try a product, a price, a message or a strategy on the populations that matter to it before committing.^1,2 It argues that people describe their own behaviour unreliably, and that its simulations often predict behaviour more accurately than asking people directly.^1,2 The company has offices in New York, Singapore and San Francisco.² Its about page names John Kessler as co-founder and chief technology officer, previously a simulation researcher at the MIT City Science Lab, and Redpoint as an investor.²

Its pages quote people at EY, Accenture Song and Interpublic Group, and show an EY Global Wealth Study case and Breakwater Conviction Advantage, a quarterly tracker of 71 judgments across 40,000 simulated investors.^1,2 Use cases listed are product innovation, marketing and brand, audience segmentation, scenario planning and strategic communications.¹

In September 2026 Aaru published an accuracy benchmark of its own.³ It gave the system each question's wording, its answer options and a description of the audience, and compared the predicted split of answers with the published results of 2,993 questions from 186 studies fielded in 2026, across nine industries.³ The measure is total variation distance: how many answers in 100 would have to move for the two distributions to match.³ Aaru's mean was 7.62 points, or 5.98 after its adjustment for sampling noise.³ For comparison, Aaru calculates from published studies that the same survey fielded twice with real people, two weeks apart, differed by 2.3 to 12.0 points; it calls such disagreement the "replication floor".³ In a second test it ranked brands within a category against purchase records from a credit card panel: 92.4 percent of brand pairs came out in the same order as the purchases, against 77.8 percent for YouGov's survey of which brands people would consider.³ The page says the core model that predicts each agent's behaviour is not a language model.³

The page states its own limits. Some confidential data sources were removed for the test, so it measures a reduced version of the system.³ Only questions that already had published answers were tested, and the brand comparison covers only the brands sampled.³ Aaru ran and reported the evaluation itself, with no peer review or independent check mentioned.³ None of these pages gives a price; access is by contact or account login.^1,2,3