AI-powered contextual targeting reshapes ad delivery without cookies
New technical paper reveals how Amazon uses machine learning to match ads with relevant content in real-time without tracking.
The shift away from third-party cookies has accelerated the adoption of contextual targeting solutions powered by artificial intelligence, marking a significant change in digital advertising. According to a technical paper released on February 11, 2025, Amazon has developed an advanced system that analyzes content in real-time to deliver relevant ads without relying on tracking identifiers.
According to the document authored by Amazon's Applied Science team, the company processes over 8.5TB of webpage data daily to understand content context and match it with appropriate advertisements. This massive data processing operation employs sophisticated large language models (LLMs) and transformer architectures to comprehend content relationships beyond simple keyword matching.
The technical implementation combines two distinct approaches. First, offline batch processes categorize content based on keywords using encoder-decoder LLMs. Second, the system employs embeddings - numerical representations of content - to rapidly identify relevant ads in real-time using Approximate Nearest Neighbors algorithms.
Amazon's system demonstrates significant performance improvements compared to traditional methods. According to internal tests cited in the paper, contextually targeted impressions on meaningful pages increased 1.3x compared to targeting based on ad identifiers. The system also showed that bid requests without ad IDs converted twice as frequently on content-rich pages versus less substantial ones.
The technology has delivered measurable business impact across various metrics. Brand campaigns saw a 23% reduction in cost per thousand impressions (CPM), while performance campaigns achieved a 15% decrease. Overall engagement costs dropped by 24%, and return on investment increased by 10%.
Major brands have reported positive results from implementing the technology. PepsiCo's Prime Day 2024 campaign achieved three times higher return on ad spend compared to behavioral targeting approaches, while reducing cost per action by 62%. Similarly, Smiles credit card promotion saw a 47% reduction in cost per click on desktop and 27% on mobile platforms.
The system incorporates several technical innovations. It uses a hybrid approach combining human annotators with automated systems powered by conversational agents. This method proved 25% more accurate than crowd-sourced annotations while maintaining scalability. The architecture employs tenet-based chain-of-thought prompting, a novel technique that guides AI reasoning by breaking complex tasks into smaller components.
Amazon processes the massive volume of web content through its AdBot system, which respects publisher preferences and robots.txt instructions. The company focuses on regularly updated content, refreshing frequently changing pages in near real-time while revisiting more static content every 48 hours.
Looking toward 2025 and beyond, Amazon plans to enhance these capabilities further. The company aims to develop models that leverage LLMs to understand targeting holistically at an ad line level, potentially incorporating AI-powered context understanding into closed measurement loops.
The technical paper represents a significant contribution to the field, documenting how advanced AI techniques can maintain ad relevance without compromising user privacy. The results demonstrate that contextual targeting, enhanced by modern machine learning approaches, can effectively replace cookie-based tracking while improving key performance metrics.