Agentic Commerce12 min

Agentic Auto-Discovery: Why 90% of Platforms Are Invisible to AI Agents

We ran a simple test: we sent an AI agent to discover the top 10 agentic commerce platforms. 9 of 10 returned 404 on llms.txt. 10 of 10 had no agent-policy.json. Only one was autonomously discoverable.

Executive summary

Published

2026-03-20

12 min

Author

Platform Strategy Team

Commerce strategy analysts

The platform strategy team translates AI, commerce, and protocol shifts into actionable guidance for operational teams.

View profile

The discovery test: what we checked and what we found

For each platform we checked four layers of agent discoverability. First, llms.txt at the root — a plain-text file that tells language models what the site offers and where to find key resources. Second, agent-policy.json in the .well-known directory — a machine-readable file that declares what actions agents may take, what requires confirmation, and what is blocked. Third, mcp.json — a manifest that declares available MCP tools, endpoints, and capabilities. Fourth, JSON-LD structured data embedded in HTML pages — Product, Organization, FAQPage, and similar schemas that agents use for understanding context.

1
**llms.txt**: 9 of 10 platforms returned 404. Only one had a valid llms.txt with resource links and capability descriptions.
2
**agent-policy.json**: 0 of 10 platforms had this file. No platform declared machine-readable governance rules for agents.
3
**mcp.json**: 1 of 10 platforms had a partial MCP manifest. The rest had no MCP discovery surface at all.
4
**JSON-LD**: 4 of 10 platforms had some JSON-LD, but only 2 had complete Organization and Product schemas.

If an AI agent cannot discover your platform autonomously, it will either scrape your HTML — with all the fragility that implies — or simply skip you in favor of a platform that is machine-readable.

Essential insight

What agents look for during discovery

AI agents approach platform discovery differently from humans and from traditional web crawlers. They do not care about visual design or navigation menus. They look for machine-readable manifests that answer four questions: What does this platform do? What actions can I take? What are the rules? How do I authenticate? Each of the four discovery files addresses one or more of these questions. Without them, the agent must infer everything from unstructured HTML, which is unreliable and often incomplete.

Why most platforms fail at agent discovery

1
Most platforms were designed for human visitors and developer integrations, not for autonomous AI agents.
2
Discovery standards like llms.txt and agent-policy.json are new and not yet widely adopted.
3
Teams underestimate how much agents depend on explicit manifests rather than inferring capabilities from HTML.
4
There is no SEO equivalent for agent discoverability yet — no Google Search Console for agents — so the problem is invisible.
5
MCP adoption is growing but most platforms still treat it as a future consideration rather than a current requirement.

The agent-discoverable checklist

1
**llms.txt** at root: plain-text overview of what your platform does, key pages, API documentation links, and capability summary.
2
**agent-policy.json** at /.well-known/: JSON declaring ALLOW, FRICTION, REVIEW, and BLOCK actions for agents, trust thresholds, and confirmation boundaries.
3
**mcp.json** at /.well-known/: MCP tool manifest listing available tools, their descriptions, input schemas, and endpoint URLs.
4
**JSON-LD** in HTML: complete Organization, Product, FAQPage, and BreadcrumbList schemas on every relevant page.
5
**OpenAPI spec**: machine-readable API documentation for any REST endpoints agents might use.

How to implement each discovery file

Start with llms.txt — it is the simplest to create and provides the highest impact. A well-structured llms.txt takes 30 minutes to write and immediately makes your platform visible to language models. Next, add agent-policy.json to declare your governance rules. This tells agents what they can do without asking and what requires human confirmation. Then implement JSON-LD on your key pages if you have not already. Finally, if you offer MCP tools, create a mcp.json manifest. Each file serves a different discovery need, and together they make your platform fully autonomous-agent-ready.

Frequently asked questions

What is llms.txt?

llms.txt is a plain-text file served at the root of your domain that tells language models and AI agents what your site offers, where key resources are located, and how to interact with your platform. It is the robots.txt equivalent for AI agents.

What is agent-policy.json?

agent-policy.json is a machine-readable JSON file in the .well-known directory that declares what actions agents are allowed to take, which actions require user confirmation, what trust thresholds apply, and what is explicitly blocked. It provides governance rules that agents can read before taking any action.

Do I need MCP to be discoverable by agents?

Not necessarily. llms.txt, agent-policy.json, and JSON-LD already make a significant difference. MCP adds a deeper interaction layer — letting agents use your tools natively — but the basic discovery files are the essential foundation.

How do I test my platform's discoverability?

Send an AI agent to your domain with the instruction to discover what your platform does and what actions are available. Check if it can answer those questions without scraping HTML. Alternatively, manually verify that /llms.txt, /.well-known/agent-policy.json, and your JSON-LD schemas return valid, complete responses.

Sources and references

Agentic Commerce