A new open-source framework invents fake online shoppers that are realistic enough to train AI systems almost as well as real customer data.
SimTrace uses a computer-use agent to generate synthetic multimodal clickstreams: clicks, page views, and context, grounded in anonymized real user sessions and a simulated copy of the target website. Researchers tested it on an e-commerce site and found it beat rival methods on seven of eight fidelity measures. Models trained purely on the synthetic trajectories matched real-data performance on tasks like purchase prediction and recommendation. Mixing synthetic data in with real data pushed next-action prediction accuracy up 11 percent over using real data alone.
The pitch is simple: most companies can't share real clickstream logs because of privacy rules, and smaller sites often don't generate enough traffic to build a useful dataset in the first place. A synthetic, shareable stand-in lets researchers and smaller teams build virtual client simulators for A/B testing and recommender systems without touching anyone's actual browsing history.
Fidelity numbers from one e-commerce testbed, built and graded by the same team that built the generator, are not the same as proof this scales to messier, real-world sites.