DEV Community

Siddhesh Surve
Siddhesh Surve

Posted on

🤖 The New SEO: Why You Need to Optimize for Coding Agents

#ai #codingagents #webdev #engineeringmanagement

Forget "Search Engine Optimization." Welcome to the era of AEO (AI Engine Optimization).

Vercel noticed something wild: Coding agents (like Claude Code and OpenAI Codex) perform web searches in roughly 20% of all development prompts. These aren't just casual searches; they are deep dives by AI looking for the best libraries, deployment scripts, and config files to solve a dev's problem in real-time.

If your documentation or product isn't "agent-friendly," you simply don't exist in the modern dev workflow.

🏗️ The Problem: Agents Don’t Play by the Rules

Unlike a standard LLM chat (where you send a prompt and get a text response), coding agents are autonomous. They expect:

  1. A Filesystem: They want to read and write files.
  2. Shell Access: They want to run commands.
  3. A Sandbox: They need a live environment to execute code.

Vercel realized they couldn't just "ping an API" to see how they ranked. They had to build a lab.

🧪 How Vercel Built the "Agent Lab"

To track how these agents see the world, Vercel built a pipeline using Vercel Sandbox and AI Gateway. Here is the 6-step lifecycle of an AEO test:

  1. The Sandbox: Spin up an ephemeral Linux MicroVM (Node, Python, etc.).
  2. The Install: Globally install the agent CLI (e.g., @anthropic-ai/claude-code).
  3. The Interception: Use environment variables to route all agent calls through Vercel AI Gateway. This lets Vercel see exactly what the agent is asking the LLM without the agent knowing.
  4. The Prompt: Run the agent with a specific query (e.g., "Deploy this Next.js app to Vercel").
  5. The Transcript: Capture the raw JSONL or stdout data of every tool the agent called.
  6. The Tear Down: Kill the sandbox to save resources.

🧹 The "Transcript Mess" (and How to Fix It)

The hardest part of AEO? Every agent speaks a different language.

  • Claude Code calls a file read: Read.
  • Codex calls it: read_file.
  • OpenCode calls it: read.

Vercel built a Normalization Layer that collapses these into canonical names like file_read or shell. This allows their marketing and engineering teams to see a unified "Share of Voice" across all agents.

💡 3 Lessons for 2026 Devs & Founders

  1. "Demo, Don’t Memo" Applies to Content: Coding agents don't just "mention" tools; they write code with them. If your documentation doesn't have clear, copy-pasteable snippets, the agent will skip you for a competitor who does.
  2. The Terminal is the New Browser: Your "Brand Awareness" now happens in a terminal prompt. If an agent recommends a tool, it's often providing the import statement and the deployment script. That is a 100% conversion rate.
  3. AEO is a Technical Problem: You can't just throw keywords at a page. You need to understand how agents use WebFetch and Grep to parse your docs.

🚀 What's Next?

Vercel is planning to Open Source this AEO tracking tool. Soon, every startup will be able to run "Agent Evals" to see if their product is the first choice for the next generation of AI-driven development.

Are you ready for an agent to be your biggest customer?

Top comments (0)