AI is reshaping how businesses understand and interact with information. Vercel is developing an AI Engine Optimization (AEO) system to track how large language models (LLMs) discover and reference their web content. Initially focused on standard chat models, they quickly expanded to include coding agents, which are more common among Vercel users who engage with AI through terminals or integrated development environments (IDEs). Early data showed that coding agents perform web searches in about 20% of their interactions, highlighting the need for effective tracking of both response quality and source accuracy. The AEO lifecycle for coding agents differs significantly from standard model testing. Coding agents operate within a project environment and require access to filesystems and shell commands. Vercel Sandbox addresses this by creating ephemeral Linux MicroVMs for each agent run. The process includes creating a sandbox, installing the agent's CLI, injecting credentials, running the agent with a prompt, capturing the transcript, and tearing down the environment. Each agent is treated as a config object, making it easier to add new ones to the system. The use of the AI Gateway streamlines cost management and logging by routing all requests through a centralized system. This allows agents to function as if connecting directly to their providers while maintaining oversight. A significant challenge arises from the differing transcript formats produced by various agents. Vercel built a normalization layer to handle these discrepancies, ensuring that data from different agents can be aggregated and analyzed uniformly. This includes capturing the transcript, parsing it to standardize tool names, enriching data with structured metadata, and ultimately summarizing the results.