We develop a procedure for adjudicating between models of firm wage-setting conduct. Using data on workers' choice sets and decisions over real jobs from a U.S. job search platform, we first estimate workers' rankings over firms' non-wage amenities. We document three key findings: 1) On average, workers are willing to accept 12.3% lower salaries for a 1-S.D. improvement in amenities. 2) Between-worker preference dispersion is equally large, indicating that preferences are not well-described by a single ranking. 3) High-paying firms have better amenities. Following the modern IO literature, we use these estimates to formulate a test of conduct based on exclusion restrictions. Oligopsonistic models incorporating strategic interactions between firms and tailoring of wage offers to workers' outside options are rejected in favor of simpler monopsonistic models featuring near-uniform markdowns. Misspecification has meaningful consequences: while our preferred model predicts average markdowns of 19.5%, others predict average markdowns as large as 26.6%.