Metadata practitioners and standards developers typically take classifications as given. Rarely do they look at how these were created and whether they make sense for respondents or data users. This talk will break tradition and discuss this issue. We considered a question in the Current Population Survey in the US on self-employment, called Class of Worker (COW). The COW question reads: "Were you employed by government, a private company, a non-profit organization, or were you self-employed (or working in the family business)?" The basic question was whether these four response options make sense together. Said another way, does the COW classification make sense for data users? Based on research done by the Small Business Administration and independent researchers, we suspected the answer was No. To investigate this, we paid 90 volunteers to come to BLS and classify a set of twelve job description vignettes based on two different groupings of the COW classification. The data show answering COW is a difficult task. Using research from cognitive psychology, we are now able to provide reasons for this, and we propose new roles for metadata practitioners and new considerations for metadata standards developers.