How we built AEO tracking for coding agents

Eric Dodds and Allen Zhou on extending brand tracking to a channel that does not behave like a chatbot. Vercel already runs what it calls an AI Engine Optimization system, tracking how models discover, interpret and reference Vercel and its sites. Coding agents needed a separate harness, because developers meet them in a terminal or IDE partway through real work rather than by asking a question cold. Their reason for caring:

Agent recommendations have a different shape than model responses. When a coding agent suggests a tool, it tends to produce working code with that tool, like an import statement, a config file, or a deployment script.

The mechanics are unremarkable in a useful way. Each run gets a Vercel Sandbox, an ephemeral Linux microVM: create it, install the agent CLI, inject credentials, run the prompt with a timeout so nothing loops forever, capture the transcript, tear it down. Each agent is a config object naming its setup and build commands, so Claude Code, Codex and OpenCode can be driven uniformly despite different CLIs. Credentials are never handed to the agent directly; base URLs are overridden so everything routes through Vercel’s AI Gateway for logging and cost tracking. Transcripts then go through a four-stage pipeline (capture, parse into canonical types, enrich with URLs and file paths, summarise) before brand extraction.

The number to remember is from early sampling: coding agents run a web search in roughly 20% of prompts. Most of the time they answer from training data, which means what a model already believes about your product matters more than what your docs say today.