In sponsored search, advertisers bid on phrases representative of offered products or services. For large advertisers, these phrases often come from quasi-algorithmically generated lists of thousands of terms prone to poor linguistic construction. A bidded term by itself is usually unsuitable for direct insertion into an ad copy template; it must be rephrased and capitalized properly to fit the template, possibly with additional language to avoid semantic ambiguity. We develop a natural language generation system to automate these steps, preparing a list of terms for insertion into an ad template. For each input term, our system first finds a proper word ordering by mining a corpus of Web search query logs. Next it determines whether the term is ambiguous and--if semantics dictate--attaches a clarifying modifier culled from query logs. Finally, it applies proper capitalization by analyzing pages from Web search engine results. Each step yields a plausible set of displayable forms from which a machine-learned model selects the best. The models are trained and tested on a large set of human-labeled data. The overall system significantly outperforms baseline systems that use simple heuristics.