We conduct a choice experiment where the number of labels vertically differentiating Chianti wines (Chianti, Chianti Classico, Chianti Classico Riserva, Chianti Classico Gran Selezione) is augmented incrementally in a between-subject design, eliciting both quality perceptions and wine choices. We find that quality expectations are endogenous to the labeling regime, and adding a high-quality label (e.g., Chianti Gran Selezione) decreases the perceived quality of all other Chianti wines (comparative stigma). A model conditioning on subjective quality perceptions with heterogeneous WTP for quality is then proposed, and estimated via random parameter multinomial logit. The endogeneity problem arising from using subjective beliefs as regressors is addressed by means of a control-function approach. Results are compared to reduced form approaches where the marginal utility of quality and subjective perceptions are confounded in a single label-specific estimate, and the model is used to determine how much of the cannibalization observed after introducing higher-tier quality standards is attributable to restructuring of perceptions and comparative stigma.