Yahoo upgrades ad targeting with Amazon Bedrock LLMs

Yahoo has upgraded its Search Retargeting (SRT) capabilities in its omnichannel Demand-Side Platform (DSP) by integrating Amazon Bedrock's generative AI models, replacing its outdated Word2Vec-based keyword expansion system. The legacy approach struggled with outdated vocabulary, syntactic similarity, and zero-expansion failures, limiting advertisers' ability to target users based on nuanced search intent. By deploying LLMs via Amazon Bedrock—including models like Amazon Nova, Meta Llama, and Anthropic's Claude—Yahoo now generates semantically rich keyword expansions that align better with user interests, bridging search intent to display, video, and native ads.

The new architecture leverages Amazon Bedrock's serverless access to foundation models, eliminating the need for custom infrastructure while ensuring compliance with sensitive keyword filtering, denylists, privacy preferences, and policy requirements. Advertisers define target segments via keywords, which are expanded into semantically related terms and stored in Amazon OpenSearch Service. Batch Scoring then evaluates user search histories against these segments to determine membership, enabling more precise and scalable ad targeting across Yahoo's premium supply and partner networks.