Executive Summary
Vercel has developed an internal AI Engine Optimization (AEO) system to track and analyze how AI coding agents discover, interpret, and reference its web content. To overcome the unique challenge of testing agents that require a full development environment, the system utilizes Vercel Sandbox to provide isolated, ephemeral MicroVMs for each test run. By routing all agent API calls through the Vercel AI Gateway, the company achieves unified logging, cost tracking, and observability, enabling a comprehensive understanding of its content's performance in AI-driven developer workflows.
Key Takeaways
* AEO for Coding Agents: Vercel is pioneering an "AI Engine Optimization" (AEO) system specifically to analyze how AI coding agents (e.g., Claude Code, OpenAI Codex) interact with and cite web content during development tasks.
* Vercel Sandbox for Isolation: The system's core is Vercel Sandbox, which creates ephemeral Linux MicroVMs to safely run each agent in a full, isolated runtime environment, solving the problem of agents needing to execute arbitrary code.
* Standardized Agent Lifecycle: A uniform six-step lifecycle is applied to test any agent: create a sandbox, install the agent's CLI, inject credentials, run the prompt, capture the transcript, and tear down the sandbox.
* Centralized Observability via AI Gateway: All LLM calls from the various agents are routed through Vercel AI Gateway. This provides centralized logging, rate limiting, and cost tracking, even when agents use different underlying AI providers.
* Transcript Normalization Pipeline: The system uses a four-stage process to handle the inconsistent output formats of different agents. It captures raw transcripts, parses them into a unified event type, enriches the data, and finally performs analysis and brand extraction.
Strategic Importance
This initiative positions Vercel as a thought leader in the nascent AEO field while simultaneously serving as a powerful, real-world demonstration of its own products—Vercel Sandbox and AI Gateway—for managing and observing complex AI agent workloads.