---
id: e1
slug: agent-discovery-866-domains
title: "Your Next Customer Won’t Read Your Website. Their AI Will."
date: October 2026
tags: [Agent Discovery, A2A & MCP, Field Study, Explorable]
canonical: https://amyote.com/#essay-agent-discovery-866-domains
---

# Your Next Customer Won’t Read Your Website. Their AI Will.

> I scanned 866 domains across 11 industries with the same protocol probes an autonomous AI agent uses to find an official business endpoint.

**Date:** October 2026 · **Reading time:** 4 min read + Interactive Explorer  
**Tags:** Agent Discovery, A2A & MCP, Field Study, Explorable  

Every AI lab now promises that your personal agent will soon book your hotel, haggle with your dentist and reserve your table, leaving you free for something more dignified, like staring at a wall. The standards exist: DNS-AID, the A2A Agent Card (/.well-known/agent-card.json), Anthropic’s MCP (v=MCPv1), Meta and Sierra’s PAP. Splendid acronyms, all of them. The only slight snag is that somebody on the receiving end has to switch them on. [1]

In developer tools and Web3 the future has already arrived, and it is rather smug about it. Our deterministic Python scanner finds Ed25519 keys in DNS TXT records, working MCP servers and even live HTTP 402 stablecoin handshakes. Run the recorded trace below on an API-first incumbent, then on the one pioneer hotel in Austria, and see how the other half lives:

> **Pivot Question:** What happens if an agent tries to book any of 166 real hotels in Rome or San Francisco — or checks local restaurants, dental clinics, and auto shops?

Across 103 verified hotel domains in Rome and 63 in San Francisco (sampled deterministically from OpenStreetMap with seed=42), official agent readiness is 0.0%. Not a single hotel published a DNS-AID record, an A2A Agent Card, an MCP server manifest, or a structured /llms.txt index. Worse, scanning 25 mainstream hotel property management systems (Sabre SynXis, Amadeus, Mews, Cloudbeds, SiteMinder, SimpleBooking, Blastness) also yielded 0.0% public A2A/MCP gateways, refuting the assumption that legacy PMS vendors will automatically bridge independent hotels into the agentic web. [2]

Yet when we ran the exact same probe across 300 restaurant and food-delivery targets in San Francisco, Rome, and London, 10.2% of independent restaurants (26 of 255) returned verified machine-readable endpoints. Why did neighbourhood pizzerias and trattorias beat luxury hotels? Platform federation. The moment Wix enabled auto-generated /llms.txt files pointing to a Site MCP Endpoint, 13 independent restaurants in our sample became agent-accessible overnight without writing a line of code, just as Salumeria Roscioli in Rome and top D2C brands (Allbirds, Glossier, Brooklinen) inherited Universal Commerce Protocol (UCP) and WebMCP via Shopify. [3]

Finally, inspecting the raw payloads across all 866 domains exposed five classes of security and discovery anomalies: Indonesian "Slot Gacor" casino SEO spam injected into public /llms.txt files to poison LLM context windows; internal /admin-bookings routes accidentally listed in machine sitemaps; multi-tenant cloud collisions on shared vercel.app and lovable.app suffixes; and "phantom fleets" where a single root DNS key signed 125 regional MCP servers (wishpool.app) that were completely dead on arrival.

## What, me worry? Three engineering & business takeaways

1. Walmart has engineers; the marina and the boutique hotel do not. Small and mid-sized businesses do not own their IT stack, they rent templates and legacy booking widgets. They will not configure DNSSEC or RFC 8615 manifests by hand.
2. Discovery is stochastic; transactions must be deterministic. Relying on organic LLM web scraping exposes both the customer agent and the business to prompt injection and hallucinated policies. Businesses need a bounded Agent Front Door (READ, ASK, REQUEST, GUARD) where the owner approves every fact and quote.
3. Distribution wins at the platform and connector layer. Just as Wix and Shopify lit up dozens of merchants overnight, onboarding the long tail requires zero-effort, consent-based installation that wraps existing calendars, PMS widgets, or a lightweight micro-admin.

## Margin Notes & Caveats

- **[1] Sample & Independence:** Total N = 866 targets across 11 cohorts (Oct 15–18, 2026): Rome Hotels (103), SF Hotels (63), DevTools (50), Travel/PMS (50), Web3/x402 (50), E-Commerce D2C (50), Restaurants in SF/Rome/London (300), Dental (100), Auto Repair (100). Stratified cohorts separate mainstream incumbents (Strata A) from self-declared registry pioneers (Strata B).
- **[2] The Hospitality Gap:** In San Francisco, 200 OSM hotel nodes collapsed into 63 unique root domains due to global chains (Marriott, Hilton, Hyatt, IHG) protected by Akamai/Cloudflare bot managers. Only one independent hotel globally in our pioneer stratum — Sporthotel Ideal in Austria (sporthotelideal.at) — operated a full live MCP booking server.
- **[3] Accidental Readiness vs. Intent:** A published /llms.txt or Wix MCP endpoint proves technical discoverability, not end-to-end transactional reliability or organic agent traffic. In our Agent Front Door roadmap, real third-party agent requests are tracked separately from synthetic audit probes under a strict 90-day kill rule.

## Appendix: Empirical Evidence

### Where do autonomous agents find a real endpoint, and where do they hit a brick wall?

**Takeaway:** API-first DevTools (83.3%) and D2C brands on Shopify (56.7%) already expose machine-readable files, whereas independent hotels (0.0%) and legacy hotel PMS vendors (0.0% A2A/MCP) remain completely dark.

*Strata A measures organic industry penetration among mainstream entities; Strata B measures verification and survival among self-declared A2A/MCP registry pioneers. All runs executed Oct 15–18, 2026.*

| Cohort / Vertical | Domains (N) | DNS-AID / v=MCPv1 | A2A / MCP JSON | Valid /llms.txt | Any Verified Endpoint |
| --- | --- | --- | --- | --- | --- |
| Exp 0A: Rome Hotels (OSM Seed 42)* | 103 | 0 (0.0%) | 0 (0.0%) | 0 (0.0%) | 0 / 103 (0.0%) |
| Exp 0B: San Francisco Hotels (OSM)* | 63 | 0 (0.0%) | 0 (0.0%) | 0 (0.0%) | 0 / 63 (0.0%) |
| Exp 1A: DevTools Incumbents | 30 | 7 (23.3%) | 10 (33.3%) | 24 (80.0%) | 25 / 30 (83.3%) |
| Exp 2A: Legacy Hotel PMS / CRS Vendors | 25 | 0 (0.0%) | 0 (0.0%) | 8 (32.0%) | 8 / 25 (32.0%) |
| Exp 3A: Web3 & Crypto Incumbents | 20 | 1 (5.0%) | 2 (10.0%) | 15 (75.0%) | 13 / 20 (65.0%) |
| Exp 4A: E-Commerce & D2C Incumbents | 30 | 1 (3.3%) | 0 (0.0%) | 17 (56.7%) | 17 / 30 (56.7%) |
| Exp 5A: Local Restaurants (SF, Rome, London) | 255 | 0 (0.0%) | 15 (Wix/UCP) | 26 (10.2%) | 26 / 255 (10.2%) |
| Exp 1B–4B: Registry Pioneers (Pooled) | 95 | 29 (30.5%) | 59 A2A / 31 MCP | 66 (69.5%) | 87 / 95 (91.6%) |

## Sources & Methods

- Scanner architecture: Python 3.12 deterministic auditor using httpx and dnspython with strict content-type, JSON schema, and first-party eTLD+1 ownership checks.
- Geographic sampling: OpenStreetMap relations for Rome (41485), San Francisco (111968), and Greater London (65606), deduplicated within 150m radius using fixed PRNG seed=42.
- Epistemic boundary: strata A (mainstream penetration) and strata B (pioneer survival) are reported separately and never pooled into a headline adoption rate.
