Recent reviews have summarized the rich history of drift-feeding models, including their successful application to varied theoretical and management issues. Unfortunately, their predictive success has been limited, and clear evidence now contradicts some of their foundational assumptions. For example, current models reproduce the important decline of reaction distance with increasing water velocity by inaccurately assuming that fish 1) always detect prey on the surface of a hemispherical reaction volume and 2) only capture prey they can reach at their maximum sustainable swimming speed before it passes downstream of their focal point. To clarify any such conflicts between drift-feeding models and data, and to identify a broad base of metrics for evaluating such models in the future, we reviewed empirical tests of past models and other studies of the spatial behavior or diet composition of drift feeders. We then developed a new mechanistic model in which the primary characteristics of drift-feeding behavior follow from universal cognitive constraints on the rate at which animals can process visual information. This new drift-feeding model treats prey detection as a random (Poisson) process, which permits a more realistic depiction of prey detection locations and probabilities. It also incorporates signal detection theory to describe the effects of tradeoffs between search speed (volume per unit time) and accuracy in discriminating prey from inedible debris. This model replicates and exceeds the qualitative successes of past models without using their falsified assumptions. We are currently testing it with laboratory and field data using three Alaskan drift-feeding salmonids.