We study how incorrect and uncertain beliefs about product quality can persist in equilibrium, using the example of fertilizer in East Africa. Farmers believe much local fertilizer is counterfeit or adulterated, but are uncertain of the rate of bad fertilizer; however, multiple studies find little evidence of poor quality fertilizer. We develop a learning model to explain how these incorrect beliefs persist. We show that when the production process is stochastic, agents misattribute idiosyncratic outcomes to bad inputs. Variable outcomes also interfere with updating, and allow beliefs to remain uncertain. Our learning model and simulations show that learning about quality is not possible when misattribution and multiple priors are present.