How much do opposing sources distort their private information to sway non-committed voters and how do voters act on contradictory information regarding candidate quality? This paper studies a two-candidate voting competition in which voters receive estimate of candidate quality from two opposingly biased sources who can derive unbiased estimate based on their private information. We derive and test an “average opinion” hypothesis that proposes that sources would distort their estimates symmetrically and in opposite directions due to competition and voters would split the estimate difference to decipher the unbiased estimate. We find strong evidence supporting the hypothesis, showing that information aggregation by voters is indeed informative and they benefit from the competition (compared to a control treatment without competition). When sources are allowed to lie, however, sources are predicted to babble, and voters are expected to ignore the information. In contrast, data suggest that sources still convey informative messages and voters average them to benefit as in the original hypothesis which we explain using a level-k model. We do not find any effect of weak partisanship of voters either in prediction or data.