How Yahoo enhances search retargeting using Amazon Bedrock

Connecting user search intent with relevant ad experiences across channels is a longstanding challenge in digital advertising. Advertisers need sophisticated ways to reach audiences based on their demonstrated interests and behaviors, particularly their search activity, which is one of the strongest signals of user intent. Traditional keyword expansion approaches often struggle with outdated vocabulary, limited semantic understanding, and inability to capture the nuanced meaning behind search queries.

The Yahoo omnichannel Demand-Side Platform (DSP) addresses these challenges by providing advertisers with advanced technology, premium supply access, vast scale, and trusted consumer relationships. A DSP lets advertisers purchase ad inventory across multiple exchanges and channels through a single interface. Yahoo DSP enhances this capability with sophisticated audience targeting powered by advertiser-defined rules, machine learning (ML), and generative AI.

At the heart of this targeting is the concept of an audience segment: a group of users who share specific interests, demographics, or behaviors and can be targeted collectively in advertising campaigns. These segments are passed to targeting systems for ad serving, making sure that users are matched with the most relevant advertiser campaigns.

In this post, we demonstrate how Yahoo implemented Amazon Bedrock to enhance their Search Retargeting (SRT) capabilities in the Yahoo DSP ad tech suite. SRT is a core audience targeting solution that helps advertisers reach users based on their historical search behavior, bridging search intent with display, video, and native advertising. Beyond targeting keywords entered on Yahoo Search, SRT uses AI to identify and engage users who demonstrate intent through search activity both on Yahoo and across integrated partner systems.

Previously, the Yahoo SRT architecture used the Word2Vec embedding model combined with locality-sensitive hashing (LSH) to expand advertiser keywords. This multi-step process involved canonicalizing keywords, generating embeddings, indexing them, performing LSH searches, filtering sensitive terms, and ranking the results.

However, the Word2Vec + LSH approach had several limitations:

We addressed these challenges by deploying generative AI–powered SRT keyword expansion using Amazon Bedrock and large language models (LLMs). This enhancement generates more relevant, semantically rich keyword expansions, helping advertisers reach larger and more precise audiences.

Before introducing the solution, it’s useful to review the legacy SRT workflow, illustrated in the following diagram. Advertisers begin by defining each segment through a set of target keywords. This metadata is stored in a SQL database. The backend system then performs keyword expansion, generating semantically related terms to increase audience reach and improve advertiser outcomes. Both the original and expanded keyword sets are persisted in an Amazon OpenSearch Service cluster. Another workflow called Batch Scoring evaluates Yahoo users’ search histories against this segment metadata to determine segment membership.

Recognizing the limitations of the legacy approach, our goal was to refresh keyword expansion with a modern generative AI solution. Amazon Bedrock made this work possible. By offering serverless access to a broad selection of leading foundation models (FMs), Amazon Bedrock helped our teams experiment with multiple LLMs, including Amazon Nova, Meta Llama, and Anthropic’s Claude, without building custom infrastructure.

In evaluating new approaches, we also factored in critical considerations for computational advertising, including:

The following diagram illustrates the updated architecture using Amazon Bedrock and generative AI.

During the LLM selection and evaluation process, we prototyped and assessed several models using Amazon SageMaker Studio, including Amazon Titan, Meta Llama, and Anthropic’s Claude. Ultimately, Anthropic’s Claude 3.5 Sonnet v2 was selected as the most effective model for search keyword expansion, balancing both expansion ratio and semantic similarity.

In the evaluation process, we observed hallucination behaviors from the LLMs. To reduce hallucinations and maintain quality, we introduced a verification step. Expanded keywords generated by the LLM are converted into embeddings and compared with the original set. Keywords falling below a similarity threshold are filtered out, making sure only the most relevant terms are retained. This validation step is central to maintaining accuracy and trust in the workflow.

When deployed through Amazon Bedrock, Claude 3.5 Sonnet v2 demonstrated substantially superior expansion capabilities compared to the legacy LSH method, frequently generating hundreds or even thousands of keywords in cases where LSH produced none. The median broad expansion ratio improved fivefold, and the maximum expansion ratio doubled, even under strict similarity thresholds. See the following graph.

Overall, generative AI–based keyword expansion increases the number of keywords retrieved and makes sure they are semantically relevant, significantly enhancing search user targeting.

LLM keyword expansion ratio vs. cosine similarity score.

When we first designed our SRT keyword expansion workflow, we focused solely on the expansion step, taking seed keywords and generating a broader set of related terms using generative AI models. This approach worked, but it also introduced challenges:

This highlighted the need for a quality control mechanism. To address this, we introduced embedding generation and similarity scoring:

By combining semantic similarity scoring with sensitive keyword filtering, we created a guardrailed keyword expansion pipeline. This approach preserves the breadth and creativity of generative AI–driven expansion while maintaining relevance, quality, and compliance.

This generative AI–based enhanced keyword expansion capability launched in production in Q1 2025.

Advertisers benefit from significantly broader and more precise audience reach. The LLM-powered expansion logic increases the pool of candidate keywords while filtering for quality, leading to campaigns that align more closely with user intent. This translates into higher relevancy, stronger engagement, and improved conversion rates.

From a system perspective, the new system delivers substantial scale improvements:

By modernizing keyword expansion with LLMs, Yahoo DSP is better positioned to drive value for advertisers while strengthening its competitive edge in the programmatic environment.

Amazon Bedrock provided serverless access to multiple leading FMs, including Amazon Nova, Meta Llama, and Anthropic’s Claude, without requiring the Yahoo team to build or manage complex infrastructure. This capability helped us experiment rapidly, scale our workloads, and deploy flexibly.

In this post, we highlighted how SRT serves as a core product offering on Yahoo DSP and how we transformed keyword expansion using generative AI. By using Amazon Bedrock, our teams gained serverless access to multiple leading LLMs without the overhead of managing infrastructure. This capability helped us expand advertisers’ original keywords into semantically rich, high-quality sets, increasing both the size and precision of the targeting audience.

Replacing the legacy Word2Vec and LSH workflow with Amazon Bedrock–powered LLMs delivered more accurate, contextually relevant keyword expansions. The new approach drives stronger campaign performance, higher conversion rates, and greater operational flexibility. Amazon Bedrock made this transformation possible, helping Yahoo DSP scale faster and more efficiently in AI-driven advertising.

To learn more, see the Amazon Bedrock detail page and the Amazon Bedrock User Guide. To start building, open the Amazon Bedrock console.