This study investigates how individuals update their beliefs in response to new information, testing the standard assumption of Bayesian rationality. Through a controlled lab experiment, we uncover significant and systematic deviations from the Bayesian benchmark in two key dimensions: (i) posterior mean and (ii) posterior certainty level (a counterpart of posterior variance). First, we find that the population is heterogeneous in its updating rules for mean beliefs. While the majority of subjects (80%) exhibit a mean updating rule largely consistent with Bayesian principles, a substantial minority (20%) violate the fundamental Bayesian rule of diminishing return of signals. Second, we document a clear trajectory in certainty level updating: most subjects initially underreact to information (i.e., their certainty level improves not as much as their Bayesian counterpart), but they become more reactive over time, with their certainty level ultimately converging to or exceeding the Bayesian level. Third, all participants consistently reduce their posterior certainty level in response to surprising signals, presenting a stark contrast to Bayesian updating where surprise plays no role. These results demonstrate that belief updating is characterized by significant individual heterogeneity and some systematic biases, which have not been documented before.