How Language Models Reshape Sponsored Search: Evidence from Google’s BERT Rollout

Natural language processing algorithms interpret search queries and support the real-time matching between advertisers and sponsored search ad auction opportunities on search engines. Improvements in interpretation capabilities can reshape how advertisers are ranked for ad space and what they pay. We investigate how improvements to query interpretation algorithms in sponsored search markets affect short-term cost-per-click (CPC) and the number of advertisers bidding for ad space (competition). We develop a conceptual framework that considers the types of information present within search queries. We identify two channels through which improved query interpretation can affect auction outcomes: topic understanding expands markets by enabling the platform to identify relevant advertisers with greater confidence, while context understanding creates dispersion in relevance scores that can lower prices for context-rich queries. This can lead to auctions with more bidders yet lower final clearing prices. We empirically test predictions from this framework using monthly data on the number of bidders and CPC for 12,000 queries, exploiting Google’s 2019 rollout of Bidirectional Encoder Representations from Transformers (BERT) as a natural experiment. Using a year-over-year difference-in-differences strategy, we find that the number of bidders competing for ad space increases across queries, CPC rises for short, context-poor queries, and falls for longer, context-rich queries. The finding that CPC declines for long queries despite increased competition is inconsistent with alternative platform changes. Our results illustrate how improved interpretation algorithms can reshape sponsored search market prices and competition.