Agentic Engine Optimization (AEO) is the practice of structuring, formatting, and serving technical documentation so that AI coding agents — not just human readers — can discover, parse, and act on it. The term was introduced by Addy Osmani, engineering lead for Google's Chrome DX team, in a 2026 blog post that framed it as a discipline distinct from 📝Search Engine Optimization (SEO) and 📝Generative Engine Optimization (GEO), which target search crawlers and answer engines respectively. AEO treats agents such as 📝Claude Code, Cursor, Cline, and Aider as a distinct consumer class with their own access patterns, context budgets, and parsing constraints — the infrastructure layer that makes 📝agentic coding reliable at scale. It shares an acronym with 📝Answer Engine Optimization but refers to a separate concept: docs for coding agents rather than content for answer engines.
Five Dimensions
Osmani frames AEO around five questions a documentation site can fail or pass:
- Discoverability — can agents locate docs without rendering JavaScript?
- Parsability — is content machine-readable without visual-layout interpretation?
- Token efficiency — does content fit within 100K-200K token context windows?
- Capability signaling — does documentation convey what an API does, not just how to call it?
- Access control — does
robots.txtpermit AI agent traffic?
The AEO Stack
A six-layer implementation sequence:
- Access control — audit
robots.txt; consider anagent-permissions.jsonmanifest - Discovery — publish
/llms.txt, a Markdown sitemap with per-page token counts, kept under 5,000 tokens itself - Capability signaling — add
skill.mdfiles that declare tasks, inputs, constraints, and rate limits - Content formatting — serve clean Markdown alongside HTML, front-load outcomes in the first 500 tokens, place code examples immediately after the claims they illustrate
- Token surfacing — expose token counts via meta tags or HTTP headers so agents can budget context
- "Copy for AI" button — offer a clean-Markdown copy action for developers pasting docs into IDE prompts
Token Targets
Documentation exceeding practical context limits fails silently, gets skipped, or triggers hallucinated solutions. Osmani's recommended ceilings: quick start guides under 15,000 tokens, API reference pages under 25,000, conceptual guides under 20,000. Cisco's REST API guide at 193,217 tokens is cited as a cautionary example — a single page that nearly consumes an agent's entire usable context.
AGENTS.md
An emerging repo-root convention — AGENTS.md for agents as README.md is for humans. It documents project structure, API links, sandboxes, rate limits, codebase patterns, and 📝Model Context Protocol (MCP) endpoints. Cisco DevNet has adopted it as a default in their GitHub template. The closest analogue in the Claude Code ecosystem is the 📝CLAUDE.md file, though AGENTS.md aims to be cross-tool.
Agent HTTP Fingerprints
Each agent identifies itself via user-agent. Known fingerprints: Claude Code (axios/1.8.4), Cline (curl/8.4.0), Cursor (got (sindresorhus/got)), Aider (Playwright-driven Mozilla/Safari). Agents typically compress a human's multi-page browsing journey into one or two HTTP requests and bypass client-side analytics — making AI-traffic segmentation a prerequisite for measuring AEO impact.