Abstract: We track recruiters’ clicks on an online recruitment platform to estimate ethnic penalties for detailed ethnic and occupational groups. We link jobseekers to their entries in the Swiss unemployment register to observe search outcomes and skills unobserved by recruiters. Based on 3.4 million decisions, we estimate large differences in ethnic penalties between groups of immigrants. Many marginalized groups face large discrimination even if they are proficient of the local language and are culturally well integrated. The differences in penalties are hard to explain with statistical discrimination because the penalties are largely unrelated to observed and unobserved skills. We find mixed evidence for attention discrimination, and the economic effect is small. Discrimination against non-Europeans is smaller if recruiters have few candidates to choose from. In percentage points terms, discrimination is constant over the skill distribution, suggesting a similar rate of return to observable skills for natives and minorities. Finally, we provide evidence that discrimination prolongs unemployment. First, contact attempts missed out due to discrimination on the platform cause longer unemployment spells. Second, ethnic penalties are highly predictive of between-group differences in unemployment duration beyond a rich set of other predictors. Third, It is likely that jobseekers meet discriminatory recruiters. Estimating the full distribution of discrimination coefficients, we find that almost a third of recruiters discriminates against non-Europeans.