Glossary of Agentic Discovery

In short: Agentic discovery is how AI agents find, evaluate, and choose products for their users. This glossary defines the 37 terms the field actually uses, from llms.txt and AGENTS.md to ARD, ai-catalog.json, the search-and-fetch surface, and the stale window. Each is a standalone, quotable definition linked to the play that puts it to work.

This is a reference page, so the standard "Do this now" checklist is left off. Every entry links to the play or part where the action lives. Terms first defined by this guide are marked.

AEO

AEO (answer engine optimization) is the practice of optimizing content so answer engines (ChatGPT, Perplexity, Google's AI Overviews) cite it when generating direct answers. The acronym has a second sense too, "agentic engine optimization": optimizing for AI agents acting on a user's behalf. This guide treats the answer channel and the agent channel as separate, with different evidence standards. See Part 1.

Agent skills

Agent skills are installable instruction packages, defined by a SKILL.md file, that teach a coding agent a product-specific workflow such as a quickstart or a migration. Install rails include npx skills add. Adoption is measurable: top skill patterns show six-figure install counts on skills.sh. Skills write your conventions straight into the agent's environment. See Play 4.

Agent-grade docs

Agent-grade docs (term coined by this guide) are docs that meet eight testable properties: fetchable, addressable, indexed, chunked, versioned, directive, discoverable, and fresh. The definition is practical. Each property maps to a real example and a pass/fail test, so "agent-ready" becomes something you can audit instead of just claim. Five of the eight now ship as docs-platform defaults. See the agent-readiness audit.

Agentic discovery

Agentic discovery (term coined by this guide) is how AI agents find, evaluate, and choose products for their users. It spans the channels that form a default: the model's training prior, the web search-and-fetch surface it queries live, the retrieval layer it fetches mid-task, and the environment layer installed in its working context. It covers GEO and AEO, which only deal with the answer channel. See Part 1.

Agentic Resource Discovery (ARD)

ARD is an open specification, announced by Google and a Linux Foundation working group in late May 2026 (Apache-2.0, currently v0.9 draft), for discovering agentic resources across the web. Publishers describe their agents, MCP servers, skills, and APIs in an ai-catalog.json file; registries crawl those files and expose a search API so agents can find capabilities by need. It handles discovery only, before invocation, not the runtime. As of 2026-06-18, adoption is near zero. See the ARD guide.

ai-catalog.json

ai-catalog.json is the file at the center of the ARD spec, hosted at /.well-known/ai-catalog.json. It carries a specVersion, a host block, and an entries array; each entry has a domain-anchored URN identifier (urn:ai:<domain>:<namespace>:<name>), a displayName, an IANA-media-type, and either a url or inline data (never both). Think of it as sitemap.xml for agent capabilities. See the ARD guide.

AGENTS.md

AGENTS.md is a markdown file at a repository root that coding agents read at session start for project conventions and constraints. Scaffolders now write it by default. The create-next-app version states "Your training data is outdated. The docs are the source of truth." In our pilot E1, a one-paragraph AGENTS.md mandate flipped agent product choice 3/3. See Play 4.

Answer engine

An answer engine is an AI system that responds to a query with a written answer plus citations instead of a list of links. ChatGPT search, Perplexity, and Google AI Overviews are examples. It is one consumer of your content. Coding agents that select and integrate products are another, with different patterns (fetching markdown, grepping text). See Part 6, myth 7.

AX (agent experience)

AX, or agent experience, is the agent-facing version of developer experience: how easily an AI agent can fetch your docs, get credentials, call your API, and confirm success without human help. Poor AX shows up as failed integrations and agents picking competitors. The agent-readiness audit measures AX with eight pass/fail tests you can run in about 30 minutes.

B2A (business-to-agent)

B2A, or business-to-agent, is marketing and selling to software agents acting on behalf of buyers. The classic example is Upstash's page footer, written for no human: "For AI agents: a free Redis database is available via POST… No signup required." Roughly 2.6% of one major doc index's queries come from raw HTTP clients, custom agents nobody designs pages for. See Play 10.

Birdseye

Birdseye (Synscribe's free agent-observability Mac app) rebuilds an AI agent's research run as four findings layers you can inspect: per-page, per-search, per-agent, and synthesis. It shows the queries it wrote, the pages it opened, and the claims it killed. It is how the search-and-fetch findings in this guide were measured. See Play 1.

CLAUDE.md

CLAUDE.md is the project rules file Claude Code reads at session start: instructions, conventions, and tool preferences that persist across every session in that repository. Products now compete to be written into it. bun init auto-creates a CLAUDE.md that prefers Bun when it detects an agent environment. It is one of the highest-leverage files in the environment layer. See Play 9.

Content negotiation

Content negotiation is serving different formats at the same URL based on the requester: HTML to browsers, raw markdown to agents and non-browser fetchers. Next.js is the gold standard here. Its docs returned markdown to our non-browser fetcher on canonical URLs with no .md suffix. It makes every page agent-readable with zero URL changes. See Play 6.

Context7

Context7 is the leading public docs-retrieval index for coding agents: an MCP server (1.14M weekly npm installs as of 2026-06-11) that fetches current docs mid-task, plus published rankings of which libraries agents fetch most. Owners can claim their entry, tune parse config, and customize benchmark questions, which makes it one retrieval surface you can own among several. We use it as our main measurement window, not as the thing to optimize for. See Play 2.

Cursor rules

Cursor rules are instruction files in a project's .cursor/rules/ directory (.mdc format) that bias every Cursor session in that repository. They also work as a product-distribution surface. bun init writes one named use-bun-instead-of-node-vite-npm-pnpm.mdc, where the filename is the strategy. They are part of the environment layer, with the same disclosure and opt-out ethics as any injected rule. See Play 9.

Default-formation funnel

The default-formation funnel (term coined by this guide) is the path by which a product becomes an agent's habitual choice. Training prior, retrieval layer, and environment layer feed selection. Selection then feeds lock-in, as scaffolders write rules files that re-bias every future session. That closing loop is why the environment layer outweighs the other channels. See Part 2.

Deprecation eval

A deprecation eval is a test suite of prompts that used to pull up your old API, scored by whether an agent emits deprecated patterns with your docs in context. Target: under 5% emission. In pilot E3, directive docs took emission on a recent breaking change from 100% to 0%. Run it in CI and within two weeks of every model release. See Play 8.

Docs-MCP

A docs-MCP is an MCP server that exposes your docs as searchable tools (minimum: search_{product}), so agents query your current docs instead of relying on training data. It is the cheapest MCP to ship. Mintlify-class platforms generate one for you. It is the standard first step before a live-system MCP. See Play 3.

Environment layer

The environment layer (term coined by this guide) is everything installed in an agent's working context before any retrieval happens: AGENTS.md, CLAUDE.md, cursor rules, and skills. It is the strongest selection lever we measured. A rules file flipped product choice 100% in pilot E1. It works because it persists across sessions and skips retrieval ranking entirely. See Play 9.

Flip rate

Flip rate is the share of trials in which an agent's product choice changes when one of your surfaces (rules file, llms.txt excerpt, operability fact) is present versus absent. It is the cleanest single metric for "are we becoming the default?" Our pilot E1 measured 100% (3/3 vs 0/3). Run it with n≥5 per cell. See Play 11.

GEO (generative engine optimization)

GEO, or generative engine optimization, is the practice of optimizing content for citation in AI-generated answers. It was named by the 2024 academic paper "GEO: Generative Engine Optimization." Its findings (statistics, quotes, and citations boost answer visibility) concern the answer channel and stay unreplicated for agent product selection. That scope distinction is one the echo chamber routinely drops. See Part 6.

Keyless onboarding

Keyless onboarding lets an integration be finished before any signup: the agent builds against the product right away, and a human claims the account later. Clerk's Keyless Mode is the reference. Its docs instruct agents "Do NOT tell users to sign up, create accounts, get API keys." It removes the one step agents cannot do alone. See Play 10.

llms-full.txt

llms-full.txt is the companion file to llms.txt. It holds the entire docs corpus as a single markdown file, for consumers that prefer bulk ingestion over link-following. Teams tier it by token budget. Hono pioneered the llms.txt / llms-full.txt / llms-small.txt (~90KB) split for agents with small context windows. Link it from your llms.txt. See Play 5.

llms.txt

llms.txt is a markdown index file at a site's root that lists documentation links, each with a one-line description, so LLMs and agents can navigate your content fast. In practice it has also become an instruction channel. Stripe's tells agents to "never recommend the Charges API." It is proven as infrastructure but unproven as a ranking lever. See Play 5.

MCP (Model Context Protocol)

MCP, the Model Context Protocol, is an open standard that lets AI agents call outside tools and resources: search a vendor's docs, spin up a sandbox, operate a live system. Products ship MCP servers so agents can use them directly. In pilot E4, knowing a vendor shipped MCP and agent tooling flipped agent selection 2/2. See Play 3.

MCP registry

An MCP registry is a directory where agents and their clients discover MCP servers. Few placements show measured impact. The official registry (registry.modelcontextprotocol.io) matters as the upstream pipe: GitHub's MCP Registry sources from it, and VS Code renders that natively. Most standalone directories show no usage evidence. Roughly 99% of measured distribution skips them. See Play 2.

Retrieval layer

The retrieval layer is the set of surfaces agents fetch mid-task to ground their work: llms.txt indexes, .md mirrors, docs-MCPs, and retrieval directories like Context7. It is where products too young for training data win today. A two-year-old library is the #2 most-fetched docs source. Quality beats mass here, and freshness is the strongest score correlate. See Part 2 and Play 5.

Sandbox MCP

A sandbox MCP is an MCP server pointed at a test environment. It lets agents complete a full integration (accounts, transactions, webhooks) without touching real money or data. Polar's production-plus-sandbox split is the reference build. For transactional products it is the difference between being cited and being integrated. See Play 3.

Scaffolder injection

Scaffolder injection is a CLI writing agent rules files into the projects it creates. bun init detects Claude or Cursor and writes CLAUDE.md plus a cursor rule preferring Bun. create-next-app writes AGENTS.md by default. It is the most aggressive verified growth lever in agentic discovery. To stay legitimate it must be disclosed in output and easy to opt out of. See Play 9.

Search-and-fetch surface

The search-and-fetch surface (term coined by this guide) is the layer between the training prior and the retrieval layer. Here an agent runs its own web searches, opens a fraction of the results, and checks the claims it reads against primary sources. It is the entry gate for any product the model doesn't already trust. It is also the one layer where the prior gets overturned mid-task. See Play 1.

SKILL.md

SKILL.md is the file format that defines an agent skill: name, trigger description, and the step-by-step instructions an agent loads when the skill is invoked. The spec (agentskills.io) is an open Anthropic standard. Vercel's skills CLI distributes any public repo containing SKILL.md files, with no submission or review step required. See Play 4.

skills.sh

skills.sh is Vercel's directory and leaderboard for agent skills, backed by the skills CLI (1.38M npm downloads/week as of 2026-06-11; one command installs to ~55 agent clients). There is no gatekeeper: any public repo with SKILL.md files becomes installable, and the leaderboard auto-populates from install telemetry. Your own funnels drive installs. The directory is the scoreboard. See Play 4.

The four stages

The four stages (framework coined by this guide) are the steps every agent runs to pick a product. Find: appear in its searches and training prior. Research: get opened and read ("is this what I want?"). Shortlist: win the head-to-head comparison (does it work, can I build it, is it best?). Act: make integration easy, then stay in its environment for next time. Most of the twelve plays are tools that work across more than one stage, not items in a single box. Find is the entry gate. The later stages move the final pick the most, and Act feeds back into the next Find, so defaults compound. See Part 4.

The stale window

The stale window (term coined by this guide) is the time between an API change shipping and the next model-training cycle absorbing it. These are the months when agents confidently emit your old patterns. It is the exact spot where directive docs work: 100%→0% deprecated emission in pilot E3, versus no effect on long-absorbed deprecations (E2). It returns with every release you ship. See Play 8.

The fetch cut

The fetch cut (term coined by this guide) is the gap between being surfaced in an agent's search and actually being opened by it. In one instrumented run, agents surfaced 215 domains but fetched only ~13 (~6%). The choice to open is driven by title, snippet, domain authority, and freshness, not your on-page content. See Play 1.

The verification cut

The verification cut (term coined by this guide) is the share of claims an agent kills when it tries to refute them against primary sources before believing them. In one instrumented run, a three-voter skeptic panel killed 48% of the claims it checked. Self-reported vendor numbers fared worst, while claims backed by a primary source survived. See Play 1.

Zero-human integration

Zero-human integration is the north-star test for agent-first onboarding: can an agent go from the instruction "integrate this product" to a passing test with no human action (no signup, no key copying, no dashboard visit)? It is measured as a pass/fail eval plus time-to-first-successful-API-call. Keyless onboarding and sandbox-by-default keys are the enablers. See Play 10.

Reference note: every definition above traces to the same evidence base as the rest of this guide: the audit of 18 developer products, live Context7 rankings, the directory-landscape study, and pilot experiments E1–E5 (single model family, n=2–3 per arm, directional, not population estimates), all collected 2026-06-11. The quotable numbers, with methodology, live in the Data Room.

FAQ

What is the difference between GEO, AEO, and agentic discovery? GEO and AEO both target the answer channel: getting cited when AI systems generate answers for humans. Agentic discovery is the umbrella covering how agents find, evaluate, and choose products across all channels: training prior, retrieval, environment, and the product surface itself.

Which terms does this guide coin? Nine: agentic discovery, the four stages, the stale window, default-formation funnel, agent-grade docs, environment layer, the search-and-fetch surface, the fetch cut, and the verification cut. Each is marked "coined by this guide" in its entry and defined so it can be cited standalone.

How current are these definitions? Every entry was verified 2026-06-11 and is re-tested quarterly alongside the rest of the playbook. Status changes (terms, numbers, and verdicts) are logged in the Data Room.

---

Last verified 2026-06-12. We re-test the claims on this page quarterly. Changes are logged in the Data Room.

Part of The Complete Playbook to Agentic Discovery.

Stay ahead of the agents. We re-test this playbook quarterly and publish what changed: new data, busted myths, ranking shifts. Get the update digest →

Want this done for you? Synscribe runs agentic-discovery programs for B2B SaaS and developer platforms. Talk to us →