How we built AEO tracking for coding agents
By Steven Van ·
Vercel details the sandboxed pipeline it built to track how coding agents like Claude Code and Codex search for and cite its content.
Vercel has extended its AI Engine Optimization (AEO) tracking, the system it uses to monitor how LLMs discover and cite Vercel's content, to cover coding agents rather than just standard chat models. Early sampling found that coding agents perform web searches in roughly 20% of prompts, and because those searches happen inline with real development work, Vercel wanted to check both response quality and source accuracy from CLI-based agents, not just API responses.
Each agent run gets its own Vercel Sandbox, an ephemeral Linux MicroVM, where the agent's CLI (such as @anthropic-ai/claude-code or @openai/codex) is installed and run with a prompt. Credentials are injected as environment variables that route each agent's LLM calls through Vercel AI Gateway instead of a direct provider key, giving unified logging, rate limiting and cost tracking across agents that use different underlying providers.
Because Claude Code, Codex and OpenCode each produce transcripts in different formats, Vercel built a four-stage normalization layer, parsing, enrichment and summary steps, that turns each agent's tool calls and outputs into a single unified event type before feeding it into the same brand extraction pipeline used for standard model responses. The full technical breakdown is in Vercel's writeup.
