Yahoo has modernized its search retargeting (SRT) capabilities by integrating Amazon Bedrock, a service that provides access to large language models (LLMs). This integration helps advertisers better connect with users based on their search behavior, improving the relevance and effectiveness of targeted ads. The SRT solution now leverages generative AI to expand keywords, making it easier to reach larger and more precise audiences.
Previously, Yahoo used the Word2Vec embedding model with locality-sensitive hashing (LSH) for keyword expansion, but this approach had limitations such as outdated vocabulary and poor semantic understanding. By switching to Amazon Bedrock, Yahoo was able to significantly enhance keyword expansion, with the median broad expansion ratio improving fivefold and the maximum expansion ratio doubling. The new system also introduced a verification step to filter out irrelevant or sensitive keywords, ensuring the generated terms remain semantically relevant and compliant with privacy and policy requirements.
The updated SRT workflow involves defining target keywords, expanding them using generative AI, and then evaluating the results through embedding and similarity scoring. This ensures that only the most relevant terms are retained for targeting. The solution also incorporates checks for sensitive keywords, both before and after inference, to maintain quality and compliance. Yahoo’s implementation of Amazon Bedrock marks a significant step in improving the efficiency and accuracy of audience targeting in digital advertising.
Source: awsml
