Agent SEO: How to Get Recommended by AI Coding Agents

SEO optimized for a crawler that ranked links for a human to click. Agent SEO optimizes for a reader that doesn't click anything — it reads, decides, and writes the import statement itself. The job is no longer "rank on page one." It's "get picked."

Getting picked isn't one lever. An agent moves through four observable stages before it chooses you, and a weakness at any stage caps everything downstream. Here's the checklist, stage by stage.

1. Awareness — be in the corpus

If the model has never seen you, you can't be picked. This is the slowest stage to move because training data is baked months in advance, but it's not passive:

• Publish docs as clean, crawlable HTML — not locked behind JS-only rendering.
• Get written about in the places models train on: GitHub READMEs, Stack Overflow answers, well-linked technical blogs.
• Use consistent, unambiguous naming so the model associates your package with the job.

2. Discovery — be machine-readable now

This is the stage you can fix this week, and most tools fail it. Agents and their crawlers need to reach and parse you in real time:

• Ship an llms.txt that points agents at your key docs.
• Publish a real OpenAPI spec and, where it fits, an MCP server.
• Check your robots.txt isn't blocking the AI crawlers you want to reach you.
• Watch the Probe/Serve gap — the difference between bots hitting you and what you actually return. A high probe with a thin serve is a silent leak.

3. Recommendation — win the unbranded task

This is the outcome: your Pick Rate. The agent gets an unbranded job — "add auth," not "add Clerk" — and picks. To move it, find the prompts you lose and the rival taking them. In auth, for example, the race is tight: the leader wins around 15% of trials with a rival a couple points behind, so a handful of recovered prompts changes the ranking. Contrast payments, where the leader's ~78% means newcomers should target a niche, not the headline.

4. Adoption — make the pick stick

Getting imported once isn't winning if the agent rips you out on the next pass. Reduce friction to first success: a quickstart that works in one paste, sensible defaults, error messages a model can act on. Adoption is the signal that the pick held.

The mistake to avoid

Most teams pour everything into Awareness — more content, more mentions — and wonder why their Pick Rate doesn't move. Usually it's a Discovery or Recommendation problem: agents know you exist and still don't choose you. Measure first. See where you rank, find the stage that's actually capping you, and fix that one.