Purpose: Strategy selection and effort allocation in cognitive tasks are often mediated by metacognitive assessments of self-performance (e.g., Nelson & Leonesio, 1988). The current experiment examined the accuracy and role of predictive metacognitive judgments in a real-world visual search task. Method: Procedure was modeled after the judgments-of-learning task (e.g., Dunlosky & Nelson, 1992) commonly used to study metacognitive accuracy and control. In a pair of experiments, subjects performed a simulated luggage x-ray screening task, searching for knives hidden among varying numbers of background objects in passenger bags. Before performing the search task, subjects viewed each stimulus image without the embedded target item and rated how likely they would be to find the knife if it were hidden somewhere in that image. Ratings were made on a 5-point Likert scale. Experiment 1 used a 2IFC procedure for the visual search task. Experiment 2 used a speeded response procedure. Results: Experiment 1 produced a statistically significant but weak correlation between predicted and observed target detection performance (mean Goodman-Kruskal gamma =.161), indicating that subjects' predictive metacognitive judgments were only modestly accurate. Experiment 2 found that target-present RTs were similar across levels of predicted detection likelihood, but that target-absent RTs were longer for stimulus images in which target detection was predicted to be difficult. Conclusions: Results suggest that predictive metacognitive judgments are only modestly accurate, but that the information used in metacognitive judgments is nonetheless used to regulate criterion for terminating search when no target is detected.