The increasing spread of false stories (“fake news”) represents one of the great challenges societies face in the 21st century. A little-understood aspect of this phenomenon and of the processing of online news in general is how sources influence whether people believe and share what they read. In contrast to the predigital era, the Internet makes it easy for anyone to imitate well-known and credible sources in name and appearance. In a preregistered survey experiment, we first investigate the effect of this contrast (real vs. fake source) and find that subjects, as expected, have a higher tendency to believe and a somewhat higher propensity to share news by real sources. We then expose subjects to a number of reports manipulated in content (congruent vs. incongruent with individuals’ attitudes), which reveals our most crucial finding. As predicted, people are more likely to believe a news report by a source that has previously given them congruent information. However, this only holds if the source is fake. We further use machine learning to uncover treatment heterogeneity. Effects vary most strongly for different levels of trust in the mainstream media and having voted for the populist right.