We study the microstructure of the U.S. housing market using a novel data set comprising housing search and bargaining behavior for millions of interactions between sellers and buyers. We first establish a number of stylized facts, the most prominent being a nearly 50--50 split between houses that sold below final listing price and those that sold above final listing price. Second, we compare observed behavior with predictions from a large theoretical housing literature. Many predictions on the relationship between sales price, time on the market, listing price and atypicality are borne out in the data. However, existing models do not adequately explain the spread of the sales price around the final listing price. Using a modeling strategy that treats listing price changes as revisions of expectations about the sales price, we find sellers under-react to information shocks in estimating the sales price. Last, we find that the bargaining outcomes are influenced by previously undocumented buyers' bid characteristics, e.g., financing contingencies and escalation clauses, that signal a buyer's ability to complete or expedite the transaction. This suggests an important role for buyer bid characteristics, which are not explained by existing theories, in affecting bargaining power and surplus allocation in bilateral bargaining in housing transactions.