The Complete Playbook

Agentic Discovery,
end‑to‑end.

Agentic discovery is how AI agents find, evaluate, and choose products, with no human in the loop.

This guide shows you how to optimize APIs, MCP servers, and libraries to be chosen by default for coding agents like Claude Code, Codex, and OpenCode.

The Generative Engine Optimization (GEO) techniques are designed to also work for generic work agents like OpenClaw, Claude Cowork, Copilot, and more.

Start Reading →Get Free Tools
21 chapters/11 plays/8 free tools

Last verified: June 12, 2026 · re-tested quarterly

The Shift

From ranking pages to becoming the agent's default

Agents now make the call themselves. They search, compare, and pick a product with no human reading the results.

But SEO and GEO still optimize for the human doing the searching. Agentic discovery is the third box: being the default the agent reaches for.

The work moved from winning a position, to winning a citation, to winning the selection.

SEOSearch Engine OptimizationSince ~2000
What you winPositionTriggered byA human types a query into a search engine.You optimize forRanking your pages, and earning the click.Human in the loopReads & clicks.
Win the page.
GEOGenerative Engine Optimization · AEOSince ~2023
What you winCitationTriggered byA human asks an AI like ChatGPT, Perplexity, or AI Overviews.You optimize forBeing the source the answer cites. Fewer clicks, but still a reader.Human in the loopReads the answer.
Win the answer.
Agentic Discoverythe agent finds, evaluates & actsNow · unowned
What you winSelectionTriggered byAn AI agent searches, evaluates, and chooses. No human reading.You optimize forBeing the default it picks and invokes, end to end.Human in the loopNone. The agent acts.
Win the default.

Defaults are the new rankings.  ·  SEO → GEO → Agentic Discovery

The Framework

Four stages, twelve plays: the whole map

An agent runs four steps to pick a product: find → research → shortlist → act. Here is the whole method on one screen: what the agent does at each step, and what you do about it.

Most plays are tools that recur across steps, so the same name shows up in more than one column. Find is the entry gate. The later steps move the final pick the most.

Find · 1 of 4
1The agent findsFindGet into the running. Show up in its searches and training data.
Agent web searchPlay 1Rank for the agent’s own fan-out query; be in the results it scrapes.
Registries & directoriesPlay 2Get indexed so you surface in search; a second party naming you.
Evals & leaderboardsPlay 11A public leaderboard ranks in search and gets cited as a source.
Agentic Resource Discovery (ARD)Play 12Frontier: publish ai-catalog.json so registries can index your capabilities.
2The agent researchesResearchGet opened and read. It asks: “is this what I want?”
Agent web searchPlay 1Win the open decision: title, description, freshness. Survive the fetch cut.
Registries & directoriesPlay 2Corroboration across parties; coverage for more of its queries.
llms.txtPlay 5The index agents fetch first. Sectioned links, each described.
Markdown docsPlay 6Clean raw .md so what it opens is readable, not rendered soup.
Snippet densityPlay 7Self-contained, task-shaped examples across many use cases.
3The agent shortlistsShortlistWin the comparison: does it work, can I implement it, is it best?
Snippet densityPlay 7Prove the capability with runnable, complete examples.
MCP serverPlay 3Signal low-friction setup. The agent can operate you, not just read about you.
Markdown docsPlay 6Depth to compare you head-to-head against rivals in context.
Skills & AGENTS.mdPlay 4Encode “how to use us right” so you’re the easy choice.
4The agent actsActMake integration effortless, and lock in for next time.
Markdown docsPlay 6The implementation reference the agent writes against.
Directive layerPlay 8ALWAYS/NEVER rules that steer it past your deprecated patterns.
Skills & AGENTS.mdPlay 4In-context workflows that guide the integration.
Scaffolder rules / CLAUDE.mdPlay 9Persist into the repo → you’re the default next session.
Keyless onboardingPlay 10No signup, sandbox by default. An early win, zero human steps.
Evals & leaderboardsPlay 11Internal harness to tune on and prove the integration works.
The Guide

The Agentic Discovery Playbook

Twenty-three pages across six sections, plus ten company teardowns. Read front to back, or let the audit route you straight to the plays that move the needle.

01
Foundations
Start here
What agentic discovery is, and how an AI agent picks the product it reaches for.
02
Diagnose
Where you stand
One honest audit that turns the rest of the guide into a punch-list.
03
The Playbook
Twelve plays, in order
The tactical core. Run them top to bottom, or let the audit pick your first three.
Part 4The 12 Plays: Overview & SequencingThe whole method in ten minutes. Every play with who-needs-it, impact, and effort, plus how to sequence them by company type.12 min→
01Get Found in Agent SearchRank for the agent's own fan-out query.02AI Agent Registries & DirectoriesGet listed where agents look first.03MCP Server DistributionShip an MCP server agents can plug into.04Agent Skills & AGENTS.mdTell the agent how to use you.05llms.txtHand every model a map of your docs.06Markdown Docs for AI AgentsServe clean source, not rendered soup.07Code Snippets for AI AgentsCopy-paste examples that actually run.08Stop AI Using Deprecated APIsKill the old patterns models keep parroting.09Scaffolder Rules · CLAUDE.mdWin the moment a project is created.10Agent-First OnboardingA signup flow a bot can finish.11AI Evals & LeaderboardsProve you're the right default, on the record.12Agentic Resource Discovery (ARD)Frontier: Google’s new ai-catalog.json discovery spec.
04
Prove
Measure it, drop the rest
Instrument what agents actually see, and stop spending quarters on tactics the data busts.
05
Case Studies
Ten teardowns, with receipts
Who already did it? The breakouts, the engineered mechanisms, and the incumbent paradoxes.
06
Reference
Look it up
The vocabulary, the receipts, and the raw experiment runs. Everything the guide refers back to.
The Scope

What's published. And what's coming soon.

Every agent that chooses runs the same flow: find → research → shortlist → act. We prove it on one tier end-to-end; most of the same levers carry to the rest.

  • PublishedCoding agentsAPIs, MCP servers, libraries
  • Coming soonWork agentsSaaS connectors (Volume 2)
  • Coming soonShopping agentsconsumer products
  • Coming soonProcurement agentssuppliers
The agentic-discovery territory: the find → research → shortlist → act flow mapped against four tiers of agents: coding agents (Claude Code, Codex, OpenCode), work agents, consumer agents, and vertical enterprise agents. The coding-agent column is shaded as published and tested; the remaining tiers are marked coming soon.
Click to expand. The shaded map is both our evidence and our roadmap.

Subscribe to the newsletter to get the next volume the day it ships.

Free Tools

Free agentic-discovery tools you can run today

Free Tools for SEO & GEO
Mac App · BirdseyeBirdseyeBirdseye is our free agent-observability app for the Mac. It replays a coding agent's research run as four debuggable layers: every search it typed, every page it opened (or only saw), every claim it kept or killed. It's the instrument behind every number in this guide: the 57× search spread, the ~6% fetch rate, the ~48% claim-kill rate. Point it at your own category to see whether agents pick you.Download →
WorksheetAgent-Readiness ScorecardRun the Part 3 audit yourself. Eight checks, one score.Open →Template Packllms.txt Template PackAnnotated + minimal llms.txt, five directive snippets, a linter.Open →RunbookRegistry Submission RunbookFour Tier-1 checklists and a tracking grid.Open →Prompt PackDeprecation Eval Prompt PackTwenty prompts to catch models shipping your old API.Open →TemplatesRules-File TemplatesCLAUDE.md / AGENTS.md starters, ethics box included.Open →RepoEval-Harness StarterA skeleton to score yourself the way leaderboards do.Open →SpreadsheetVisibility Tracker Template66 formulas, zero setup. Track share-of-model weekly.Download →
Questions

Frequently asked questions

What is agentic discovery?

Agentic discovery is how AI agents find, evaluate, and choose products on behalf of their users. It is the full flow from first awareness to completed integration or purchase, with no human in the loop. It runs across four channels: the model's training prior, the web search-and-fetch surface, the retrieval layer (llms.txt, Context7, MCP), and the environment layer (AGENTS.md rules files).

How is agentic discovery different from GEO or AEO?

GEO and AEO optimize the answer an AI shows a human. That is a citation layer. Agentic discovery covers what happens when the agent itself decides and acts: the queries it writes for itself, the pages it opens, the claims it verifies, and the product it silently installs. Mention is not selection, and this guide measures selection.

Which agents does this playbook cover today?

This edition is proven on coding agents (Claude Code, Codex, OpenCode, Cursor, Copilot) choosing APIs, MCP servers, libraries, and developer tools. Roughly 60–70% of the levers (registries, llms.txt, markdown docs, MCP, evals) serve any agent tier. Volume 2 extends the instrumentation to work and computer-use agents.

What is Birdseye?

Birdseye is our free Mac app for agent observability. It replays an AI coding agent's research run as four layers (every query, every fetch, every extracted claim, every verdict) so you can see exactly why an agent did or did not pick you. It produced every search/fetch statistic in this guide.

What is Agentic Resource Discovery (ARD) and should I publish an ai-catalog.json?

ARD is an open spec from Google and a Linux Foundation working group (announced late May 2026) for discovering agentic resources: you publish a `/.well-known/ai-catalog.json` listing your MCP servers, agents, and APIs, and registries index it so agents can find you. If you ship one of those resources, publishing the file is a cheap (~1 hour) first-mover option, our 2026-06-18 census found 0 of 39 top sites doing it yet. But there is no agent-queried ARD registry today, so treat it as a forecast, not a working channel. Full guide: Agentic Resource Discovery (ARD).

Get a diagnosis

Are you the default an AI agent reaches for?

Get an agent-readiness diagnosis of your product, especially API and MCP products, plus the punch-list to become the one agents pick by default.

Subscribe to research