How we built AEO tracking for coding agents

AI has transformed information discovery, prompting businesses to understand how large language models (LLMs) search for and summarize web content. Vercel is developing an AI Engine Optimization (AEO) system to track how both standard models and coding agents discover, interpret, and reference their content. While standard models are straightforward to track, coding agents, which operate in development environments, present unique challenges, such as execution isolation and observability. Vercel addresses these by using ephemeral Linux MicroVMs that provide isolated environments for agents to run. The system follows a six-step lifecycle, which includes creating a sandbox, installing the agent CLI, injecting credentials, running the agent, capturing its transcript, and tearing down the sandbox. Different agents produce varied transcript formats, necessitating a normalization process to feed data into a unified pipeline. This pipeline captures, parses, enriches, and summarizes data for brand extraction, applicable to both models and agents. Early findings indicate coding agents often perform web searches and produce actionable code recommendations, highlighting the importance of optimizing content for these agents. Vercel plans to open-source its system and expand its coverage to include more agents and prompt types.