llms.txt Is Identity. WebMCP Is Capability. Most Sites Still Stop at Identity.

Robauto Team · Published Sep 21, 2026 · Last updated Sep 21, 2026

Sep 21 Robauto scans: leaderboard sites pass llms.txt but fail WebMCP. Chrome’s capability layer + Claude shopping agents explain why builders need both.

September 2026 made the split obvious: llms.txt tells an agent who you are; WebMCP tells an agent what it can do on your page. Chrome’s docs describe WebMCP as a proposed web standard that registers structured tools — search, filter, book, checkout — so browser-resident agents call your site instead of guessing at DOM clicks. Robauto’s Agent Scan already treats WebMCP as one of the Agent Integration checks. Live scans on September 21, 2026 show the gap in plain numbers: high agent-traffic sites often nail Signal Basics (including llms.txt) and still fail WebMCP.

That is not an insult. Identity files remain useful. The constructive point is that traffic without callable tools is only half a product — especially as shopping agents arrive with a “brain” and expect merchants to own the transaction surface.

What shipped in public this month

Chrome’s own framing is blunt: actuation (simulated clicks) is fragile; declared tools with JSON Schemas reduce hallucination. Sensitive actions can still require a human confirmation dialog. Limitations are real too — WebMCP is designed primarily for local browser workflows with a human in the loop, and clients must visit the page to discover tools.

Robauto network pulse (Sep 21, 2026)

While vendors debate identity vs capability, agents are already in the traffic mix. Robauto’s public AI-search summary (updated 2026-09-21) shows AI agent views at 20.84% of page views over 90 days. Within classified non-human visits, llm_crawler is 52.4% — the only class Robauto treats as AI search alongside ai_referral (currently 0).

{
  "as_of": "2026-09-21T16:00:10.683+00:00",
  "ai_search_90d": {
    "total_page_views": 1255468,
    "ai_agent_views": 261606,
    "ai_traffic_percentage": 20.84,
    "taxonomy_note": "Only llm_crawler + ai_referral count as AI search; search_bot and headless do not.",
    "llm_crawler": {"visits": 137005, "share_of_bot_traffic": 52.4},
    "search_bot": {"visits": 78844, "share_of_bot_traffic": 30.1},
    "source": "https://robauto.ai/ai-search-data"
  },
  "agent_readiness_index": {
    "attempted": 99,
    "scored": 79,
    "blocked": 20,
    "median": 18,
    "mean": 17,
    "min": 2,
    "max": 29,
    "last_run": "2026-09-02T17:55:37.973493+00:00",
    "source": "https://robauto.ai/agent-readiness-index"
  },
  "webmcp_live_scans_2026_09_21": [
    {"domain": "marketontology.com", "score": 24, "llms_txt": "pass", "webmcp": "fail", "scan": "https://robauto.ai/scan?url=marketontology.com"},
    {"domain": "shopify.com", "score": 24, "llms_txt": "pass", "webmcp": "fail", "scan": "https://robauto.ai/scan?url=shopify.com"},
    {"domain": "topleftdesign.com", "score": 30, "llms_txt": "fail", "webmcp": "fail", "scan": "https://robauto.ai/scan?url=topleftdesign.com"},
    {"domain": "robauto.ai", "score": 98, "llms_txt": "pass", "webmcp": "pass", "scan": "https://robauto.ai/scan?url=robauto.ai"}
  ]
}

Live Agent Scans: identity without WebMCP

We re-ran Robauto agent_scan (25 read-only checks) on September 21, 2026 MDT. The #1 Brand Leaderboard site by 7-day agent visits — marketontology.com with 8,484 visits — now completes Signal Basics 6/6, including a published llms.txt. It still scores 24/100: Technical Groundwork 0/4, Agent Integration 0/10 (WebMCP fail: “No WebMCP tools detected”), Agentic Commerce 0/5.

Read this constructively. A corporate marketing homepage (shopify.com) is not the same surface as merchant storefronts that may already ship agent files by platform default. A publisher that publishes llms.txt has done real work for discovery. The missing WebMCP check simply means: an assistant already in the user’s browser still has nothing structured to call.

Why Anthropic’s “brain without wallet” fits here

Coverage of Anthropic’s Claude Commerce Agents blueprint (webinar Sep 10; analysis Sep 19) describes customer-facing agents for catalog search, comparison, and cart building — and merchant agents for inventory and pricing — while excluding payment processing and checkout from the blueprint itself. That architecture assumes merchants (and payment rails) own the transaction. WebMCP, MCP server cards, ACP/UCP discovery, and related checks are how a site becomes that transaction surface instead of hoping the model invents one.

A practical ladder: identity first, then capability

If agents are not welcome, say so in robots.txt and Content-Signals and stop. A low readiness score in that case is policy working. If agents are welcome to help a human complete a task on your site, llms.txt alone will not get them across the finish line.

What to do this week

1. Run robauto.ai/scan?url=yourdomain.com (or the Robauto MCP agent_scan tool) and find the WebMCP row under Agent Integration.
2. If llms.txt is missing, publish a short brief at /llms.txt — then immediately pick one high-value on-page action (search, book, support form) and expose it as a WebMCP tool using Chrome’s imperative or declarative API.
3. Use the Model Context Tool Inspector extension from Chrome’s WebMCP docs to verify registration before you chase commerce protocols.
4. Only then open ACP/UCP/x402 if selling is the goal — reasoning agents are arriving; callable merchant surfaces are still scarce on the public web we measure.

Bottom line

On Robauto’s network, AI agents already account for roughly one in five page views over 90 days, while the Agent Readiness Index median across the most-visited cohort remains 18/100. Fresh September 21 scans show leaderboard sites graduating Signal Basics — including llms.txt — without registering WebMCP tools. Chrome is standardizing the capability layer; model vendors are shipping shopping brains that leave checkout to merchants. Builders who stop at identity files will stay discoverable. Builders who add WebMCP (and then server-side MCP/A2A and commerce rails) become actionable.

Sources

• Chrome WebMCP docs
• Business Circle — Get your website ready for AI agents (Sep 17, 2026)
• Anthropic — Building Claude Commerce Agents webinar
• Forkast — Anthropic shopping brain without wallet (Sep 19, 2026)
• Robauto AI Search Data
• Robauto Agent Readiness Index
• Robauto Brand Leaderboard
• Agent Scan — marketontology.com
• Agent Scan — shopify.com
• Agent Scan — topleftdesign.com
• Agent Scan — robauto.ai
• llms.txt
• Model Context Protocol
• A2A Protocol
• Agentic Commerce Protocol
• UCP overview

Where can an agent get this data live?

Robauto MCP — Agent Intelligence & Brand Discovery Infrastructure. A public, stateless Streamable HTTP endpoint at https://mcp.robauto.ai/mcp, callable with no key. Setup and the full tool list are in the Robauto MCP directory.

All blog posts