Agentic Engine Optimization: Writing Docs for Coding Agents

A familiar acronym with a new job

Addy Osmani, who leads engineering work at Google Cloud AI, published an essay on April 11 introducing "Agentic Engine Optimization," or AEO 13. The acronym is already taken in marketing circles, where it usually means Answer Engine Optimization, the effort to get a brand cited inside AI-generated answers 4. Osmani's version targets a narrower and more technical problem: how AI coding agents fetch, read, and act on developer documentation 1.

The gap between those two meanings explains much of the confusion around the piece. Answer Engine Optimization is about visibility in chatbot responses and AI search features. Agentic Engine Optimization is about whether a tool like Claude Code, Gemini CLI, Cursor, or Copilot can pull the right information from a docs page and turn it into working code 1. One commenter in a Reddit discussion made the same point: Osmani is writing about documentation for coding platforms and integrable products, not general web content 2.

What Osmani actually proposes

Osmani builds his case around a familiar workflow. A developer asks an agent to implement a spec, and the agent retrieves documentation, grabs pieces of it, and parses it in ways that differ from how a person reads 1. He argues that teams writing docs only for human readers are missing a growing share of their audience. He also notes that the surface area now includes CLIs, MCP servers, and Skills alongside traditional docs 1.

Search Engine Land summarized his definition as structuring and serving technical content so agents can use it rather than just render it. That covers discoverability, parsability, token efficiency, capability signaling, and access control 3. The central constraint is the context window. Long, padded pages may be truncated, skipped, or chunked badly, which raises the chance of incomplete answers or hallucinated implementations 3. Osmani treats token count as a core optimization variable and advises putting the answer within roughly the first 500 tokens, because agents have little patience for preamble 3. Discussion of the essay also highlighted his support for Markdown versions of pages and LLM-oriented index files 2.

One of his more striking observations concerns measurement. When an agent consumes a page, the client-side signals teams rely on, such as scroll depth, time on page, clicks, tutorial completions, and form interactions, go dark, because the agent never triggers them 2. If agents become a primary docs audience, traditional analytics will steadily undercount real usage.

The caveats are part of the message

Osmani frames the essay carefully. He says it should be read as one practitioner's view from developer tooling, not as a Google-wide recommendation, and he acknowledges that optimizing for agents may diverge from broader Search guidance for the open web 1. That disclaimer matters. Google's public line on AI search has been that it is still fundamentally SEO 4. Osmani is not contradicting that for marketers. He is arguing that developer docs are a special case.

A crowded vocabulary

The naming collision lands in an industry that has not agreed on terms. Virgo Dev notes there is no standard label for AI search optimization: HubSpot leans toward AEO, the academic literature began with GEO (generative engine optimization), and the firm itself chose a hybrid label rather than pick a side 4. Ahrefs lists the definitional debate among its 2026 trends and says Osmani has given AEO a second meaning. It reports that "agentic engine optimization" is a breakout query, though at only about 80 monthly searches 5. The broader cluster is growing much faster, with "generative engine optimization" up 997% and "geo vs seo" up 982% over 18 months 5. Ahrefs also identifies agent optimization as a hot topic in its own right and points to community threads arguing that AEO and GEO repackage SEO rather than replace it 5.

Reading it right

The most useful takeaway is to separate the essay from the acronym. As a marketing term, "AEO" is now ambiguous, and Osmani's reuse adds to the noise. As engineering guidance, his argument is concrete and testable. Documentation that is front-loaded, available in plain Markdown, lean on tokens, and explicit about capabilities is more likely to produce correct code when an agent reads it. That differs from chasing brand mentions in AI Overviews, and teams should treat the two as separate workstreams.

For developer-tool companies, this looks like the start of a real discipline, even if its name changes. The analytics issue may prove the most consequential part. Teams that cannot see agent traffic cannot optimize for it, and building a way to measure it could matter more than any formatting change.