Two experiments tested whether emphasizing similarities between groups might reduce the harm of stereotype threat by changing performance expectancies. In Experiment 1, male and female university students were asked to generate similarities between men's and women's physical attributes, academic characteristics, and non-academic characteristics (stereotype threat), or were not asked to generate any similarities (control). Students were then asked to imagine that they were the only (man/woman) in a class of eight other (women/men), that they had to take a (Math vs. English) exam, and that they had to estimate how well themselves and their fellow students would do on the exam. Results showed that there were no theoretically significant effects in the domain of English. However, in the Math domain, women who generated similarities regarding academic characteristics of men and women had higher performance expectancies compared with the other conditions. In the other conditions, women but not men expected to perform more poorly than the other students. Experiment 2 assessed performance expectations and actual performance of women in the domain of math. After completing the similarity exercise from Experiment 1, students were told that they would take a math exam and that their results would be compared with those of a group of other male students. Expectations were highest in the shared academic characteristics condition, and performance was best in the shared academic characteristics condition and marginally better in the non-academic characteristics condition compared with the other two conditions. A mediation analysis suggested that performance expectations partially accounted for the relation between condition and performance. Specifically, generating similarities between men and women in Math-related academic settings raised performance expectations which in turn improved actual performance of female students.