When and how do we decide whether the target is present in the scene or not? There are surprisingly few quantitative models that address this question. We propose a generative model of visual search. First, the observer obtains stimulus information (corrupted by additive white noise of variance σ) at each location and time instant. Second, information across scene locations is combined using optimal Bayesian inference to compute the instantaneous evidence for target presence in the scene (log likelihood ratio of target presence vs. absence at time t). Third, the observer accumulates evidence over time and decides ‘yes’ if the accumulated evidence exceeds criterion ζ+Δ, ‘no’ if the evidence falls below criterion −ζ+Δ, and otherwise continues to observe the scene. Δ is a criterion-shift which depends on the target frequency, and the reward for different responses. This model has only 2 free parameters (noise σ, criterion ζ), compared to many parameters in drift diffusion models. The model can explain several search phenomena including how accuracy and RT are affected by set-size, target-distractor discriminability, distractor heterogeneity, target frequency and reward. It can capture the shape of the RT distribution in target present and absent trials for different tasks like feature, conjunction and spatial configuration search. It explains that rare targets are missed (Wolfe et. al, 2005) because decreasing target frequency (e.g., from 50% to 2%) shifts the starting point of the decision process closer to ‘no’ than ‘yes’ criterion, leading to high miss error rates, and fast abandoning of search. Our model predicts that increasing penalties on miss errors will decrease these errors and increase RTs. We validated these predictions through psychophysics experiments with 4 human subjects. To summarize, we have proposed a generative model of visual search that with only 2 free parameters, can explain a wide range of search phenomena.