{"@context":"https://schema.org","@type":"BlogPosting","headline":"Estimating WebMCP Implementation for Hotels: A France Test (2026)","description":"WebMCP is a proposed browser API (document.modelContext, renamed from navigator.modelContext on July 21, 2026) that lets a page declare callable tools — name, description, typed input — for an AI agent to call directly, instead of scraping the DOM. It borrows only the \"tool\" vocabulary from Anthropic's Model Context Protocol, not its JSON-RPC wire format. Incubating as a Draft Community Group Report in the W3C Web Machine Learning CG (dated Sept 17, 2026), not on the Standards Track. Chrome's origin trial now has a defined window (Chrome 149 through 156). Mozilla closed its standards-position issue neutral (Aug 5, 2026, citing cross-origin security concerns); WebKit closed its own issue opposing the proposal outright (June 11, 2026). Neither has shipped. Google has publicly committed that Gemini in Chrome will be the first mainstream consumer, targeted for late 2026, but it has not shipped as of this writing — no AI platform is a confirmed live consumer of arbitrary-site WebMCP tools. METHODOLOGY: pulled 18,744 French hotels from this site's own hotel index (operational, has a website, not spam-flagged, excluding gîtes/chambres d'hôtes/apart-hotels), checked robots.txt, llms.txt, Schema.org Hotel/LodgingBusiness JSON-LD, and WebMCP registration (verified by reading served JavaScript, not a keyword match) on each. Cleaned to 13,808 unique real-business domains after catching two artifacts: 1,420 records shared one Accor group booking-portal URL as their \"website\" (which would have let one llms.txt file get counted 1,420 times), and 203 records listed a Facebook/Instagram/Google-Sites/Linktree page instead of a real site (Facebook's own blanket-block robots.txt policy was about to be misattributed to 136 individual hotels). HEADLINE: of 13,808 real hotel domains, 25 (26 property listings, since one domain serves two properties) have a verified document.modelContext.registerTool() call running. Three unrelated sources account for all of it: 13 run byte-identical hand-written JS from an unidentified booking-widget vendor (its French comments now reference Chrome's July 21 rename, showing active maintenance); 10 run a Shopify-pattern adapter on wine estates/domaines with a Shopify storefront (Armand Heitz, Château L'Hospitalet, Château de la Lande) — Shopify's own Aug 5, 2026 changelog confirms this ships by default on every Liquid storefront and Hydrogen preview with a broader tool set (catalog search, cart, checkout) than found here; 2 newly-identified domains were caught only because this run also checked <script src> URLs, not just inline text — \"Dalmata Hospitality\" (a hotel-management company running B&B Hotel/Hotel F1-affiliated properties) serves a bespoke /webmcp.js, and Hôtel Aramis Saint-Germain (Best Western-affiliated) loads a genuine MCP-bridge library tagged \"[webmcp-interceptor]\". No hotel built its own integration from scratch — all adoption arrived through a vendor. READINESS BASELINE (per unique domain, n=13,808): robots.txt present 77.6%, blanket-blocks everything 0.4%, names a major AI crawler 5.7%, blocks one outright 2.0%, llms.txt present (soft-404 verified) 14.5%, correctly-typed Hotel/LodgingBusiness schema 12.6%, homepage unreachable 18.4%. RECONCILING LLMS.TXT: three measurements of France's llms.txt rate — 3.8% (March 2026, n=17,634, this site's hotel-llms-txt-adoption-study-2026), 11.2% (Sept 10, 2026 pilot, unpublished), 14.5% (this study, Sept 23, 2026) — track platform-driven rollout (Shopify defaulting llms.txt on since May 2026, WordPress SEO plugins adding llms.txt generation in late 2025/early 2026), not individual hoteliers acting on their own.","datePublished":"2026-09-23","dateModified":"2026-09-23","url":"https://nicolassitter.com/research/webmcp-hotels-france-test-2026","category":"research","keywords":["WebMCP","WebMCP hotels","document.modelContext","AI agent tool calling","WebMCP vs MCP","hotel AI readiness","agentic web hotels"],"articleSection":"Research","wordCount":2600,"readTime":"11 min","articleBody":"AI infrastructure study · September 2026Hotels · WebMCP · Agentic Web\n\n# Estimating WebMCP Implementation for Hotels:a France test\n\n**TL;DR:** WebMCP — a proposed browser API (`document.modelContext`) that lets a page declare callable tools for an AI agent — is still pre-standard, Chrome-led, and has no confirmed shipped consumer. And yet, across 13,808 real French hotel websites we checked directly, **26 already have it running**, via three separate vendors, none of them the hotel itself. Adoption is real and small at the same time, and it is entirely happening one layer up from the property.\n\nNS\n\nNicolas Sitter\n\nPublished September 2026\n\n26\n\nLive WebMCP registrations found\n\n13,808\n\nReal hotel websites checked\n\n3\n\nIndependent vendor sources, 0 built in-house\n\n0\n\nAI platforms confirmed consuming it — yet\n\n[Read the Report](#intro)\n\n[Overview](#intro)[What WebMCP requires](#spec)[Methodology](#methodology)[Readiness baseline](#readiness)[Who already has it](#webmcp-finding)[Reconciling the llms.txt jump](#reconciling)[Estimating the lift](#lift)[Open questions](#open-questions)[What this means](#recommendations)[FAQ](#faq)\n\n## Why a tool-calling standard matters more than the next citation study\n\nMost of the AI-search research on this site is about _retrieval_: what shows up when someone asks ChatGPT, Google AI Mode, or Perplexity for a hotel recommendation, and which sources those answers cite. That’s the world as it exists today — an agent reads the open web and decides what’s relevant, sometimes linking back to the hotel’s own site and sometimes not.\n\nWebMCP is a bet that this gets a second layer bolted on top. Instead of an agent inferring what a page “does” by scraping its DOM — parsing a booking widget, guessing which button submits a date range — the page declares a set of callable functions (`search_availability`, whatever it wants to expose), each with a typed input and a description the agent’s model can read. The agent calls the function directly instead of pretending to be a human clicking through a UI.\n\nWhether that’s close to real for an independent hotel in 2026 — technically, and in terms of who’s actually consuming it — is the question this piece tries to answer, tested against 13,808 real French hotel sites rather than the spec’s own examples.\n\nMethodology, part 1\n\n## What WebMCP actually requires\n\nWebMCP is a genuine, single, traceable proposal — not one of several competing drafts — but considerably earlier-stage than its name suggests.\n\n-   **It is not Anthropic’s Model Context Protocol.** The spec does not implement MCP’s wire protocol (JSON-RPC over stdio/HTTP); it borrows only the “tool” vocabulary — a name, description, and typed input, callable by an agent.\n-   **It’s an incubating browser API, with no manifest file involved.** The mechanism is `document.modelContext` (renamed from `navigator.modelContext` on July 21, 2026), exposed at runtime in JavaScript — there is no static file analogous to `robots.txt` a crawler could just fetch. Discovery happens by calling `getTools()` against a live page.\n-   **Governance and maturity.** Incubating as a Draft Community Group Report in the W3C Web Machine Learning Community Group (repo `webmachinelearning/webmcp`, dated September 17, 2026) — explicitly not on the W3C Standards Track. A joint Google Chrome / Microsoft Edge effort. Chrome’s public origin trial now has a defined window, Chrome 149 through 156, not the open-ended trial it read as two weeks earlier.\n-   **Two of three non-Chromium engines have already made a call, rather than leaving it pending.** Mozilla closed its standards-position issue neutral on Aug 5, 2026, citing cross-origin tool-sharing security concerns. WebKit closed its own issue opposing the proposal outright on June 11, 2026. Neither has shipped an implementation.\n-   **The security model is unfinished.** Only a coarse Permissions-Policy `tools` allowlist exists — no per-tool user consent flow. The declarative path (mapping an existing HTML `<form>` to a tool automatically) remains an explicit TODO; exposing a tool today means writing imperative JavaScript.\n\n**The clearest sign of real momentum:** Google has publicly committed that Gemini in Chrome will be the first mainstream AI agent to consume WebMCP tools, targeted for late 2026. As of this writing it has not shipped — the 26 registrations this study found are tools sitting on a shelf with a named future shopper, not yet a confirmed one.\n\nPut plainly: WebMCP today is an experimental, Chrome-led, origin-trial-stage browser primitive with no cross-browser agreement and an open consent model — not a standard a hotel can implement once and expect any AI platform to reliably consume. Everything below is “what it would take _if_ this became real,” not “what you should build this quarter.”\n\nIt’s also worth separating WebMCP from the surrounding agentic-commerce landscape, which is more mature and has nothing to do with it. Booking.com already runs a production MCP server exposing an `accommodations_search` tool (gated to partners, unchanged since first checked) — Anthropic’s actual MCP wire format, predating WebMCP entirely. Google’s Universal Commerce Protocol for Lodging remains a US-only AI Mode test as of its Aug 27 launch, though its named brand-partner list has grown (Hilton, IHG, Marriott, Wyndham, Choice, Priceline and Trip.com added since). None of that runs on WebMCP; it runs on server-side MCP servers and Google’s own commerce rails. WebMCP is a different, browser-native answer to a similar problem, aimed at letting any website — not just a partner with a bearer token — expose a tool.\n\nMethodology, part 2\n\n## 18,744 hotels, cleaned to 13,808 real business domains\n\nWe pulled every French hotel from this site’s own hotel index (operational, has a website, not a spam-flagged domain, excluding gîtes/chambres d’hôtes/apart-hotels — the same “kill crap” hygiene filter used across every other hotel study on this site) and checked each site directly: `robots.txt`, `llms.txt`, correctly-typed Schema.org `Hotel`/`LodgingBusiness` JSON-LD, and — the part that makes this a WebMCP study rather than a fourth repeat of the robots.txt/llms.txt/schema series — a real WebMCP registration, verified by reading the actual served JavaScript, not a keyword match.\n\n**Two data-quality artifacts, caught and corrected before publishing.** First: 1,420 of 18,744 hotel records list a shared Accor group booking-portal URL (`all.accor.com`) as their “website” — one llms.txt file at that single domain would otherwise get counted as a hit for every one of those 1,420 separate hotel listings. Second: 203 records list a Facebook, Instagram, Google Sites or Linktree page instead of a real business domain — Facebook’s own blanket-block robots.txt policy was on track to be misattributed to 136 individual hotels as if each one had chosen to block AI crawlers. Every number below is computed **per unique real-business domain** (13,808 of them), not per hotel listing, specifically to avoid both distortions.\n\nThe llms.txt figure also went through the same soft-404 verification this site’s March 2026 llms.txt study used: a naive HTTP-200-at-the-path check found 30.1% before verification — most of that was fallback pages that return 200 for literally any URL. After excluding those (and the two artifacts above), the real figure is 14.5%, detailed in the next section.\n\n18,744\n\nFrench hotels pulled\n\n13,808\n\nUnique real-business domains\n\nafter cleaning\n\n1,420\n\nRecords sharing one Accor URL\n\n203\n\nRecords with a social-media page as their 'website'\n\nFindings\n\n## The readiness baseline\n\nBefore asking whether a hotel has a callable tool, it’s worth checking whether it has the far more basic machine-readable groundwork WebMCP would sit on top of.\n\nReadiness baseline, per unique real-business domain (n=13,808)\n\nSignal\n\nDomains\n\nRate\n\nrobots.txt present\n\n10,721 / 13,808\n\n77.6%\n\nrobots.txt blanket-blocks everything\n\n50 / 13,808\n\n0.4%\n\nrobots.txt names a major AI crawler\n\n782 / 13,808\n\n5.7%\n\nrobots.txt blocks a named AI crawler outright\n\n282 / 13,808\n\n2%\n\nllms.txt present (soft-404 verified)\n\n2,006 / 13,808\n\n14.5%\n\nCorrectly-typed Hotel/LodgingBusiness JSON-LD\n\n1,743 / 13,808\n\n12.6%\n\nHomepage unreachable/erroring outright\n\n2,540 / 13,808\n\n18.4%\n\n77.6% of hotel domains have a robots.txt at all, but almost none of it is AI-specific: only 5.7% name a major AI crawler one way or the other, and just 2.0% actually block one. Blanket-blocking everything is rarer still at 0.4% — and even that small number is now real, not the Facebook-policy artifact from the raw crawl.\n\nThe actual finding\n\n## WebMCP adoption is not zero — and it’s not the hotels doing it\n\nOf 13,808 real hotel domains, **25 have a real, verified `document.modelContext.registerTool()` call running** — 26 hotel listings, since one domain serves two separate properties. Confirmed by reading the actual served JavaScript, not a keyword match. Three distinct, unrelated sources account for all of it, and not one is the hotel’s own engineering team.\n\n### Unidentified booking-widget vendor\n\n13 domains\n\nByte-identical, hand-written JavaScript with French comments, reading a booking-search widget and exposing it as a tool. The comment text has been updated since the first pass 13 days earlier — it now reads \"navigator.\\* est déprécié depuis Chrome 150,\" tracking Chrome’s July 21 rename to document.modelContext. Small independent hotels and gîtes; no shared CDN or generator tag found, so this reads as one vendor’s shared template reaching 13 unrelated hoteliers, rather than 13 independent implementations of a W3C draft.\n\n### Shopify-pattern adapter\n\n10 domains\n\nMinified inline JavaScript checking both document.modelContext and the older navigator.modelContext, on wine estates and domaines running a Shopify storefront alongside their hotel site (Armand Heitz, Château L’Hospitalet, Château de la Lande and others). Shopify’s own Aug 5, 2026 changelog confirms this ships by default on every Liquid storefront and Hydrogen preview, with a broader tool set than what appears here (catalog search, cart, checkout) — not hotel-specific, a commerce platform shipping to its whole merchant base.\n\n### Two newly identified, real vendor products\n\n2 domains\n\nFound only because this run also checked <script src> URLs, not just inline text — the original pass would have missed both. \"Dalmata Hospitality,\" a hotel-management company running several budget-chain-affiliated properties (B&B Hotel, Hotel F1), serves a bespoke /webmcp.js registering search\\_dalmata\\_hotels and two other tools across its portfolio. Hôtel Aramis Saint-Germain (Best Western-affiliated) loads a genuine bridge library at /.webmcp/bridge.js — its own code carries a \"\\[webmcp-interceptor\\]\" console tag and a tool-pack/MCP-result architecture consistent with the informal @mcp-b/WebMCP-org community tooling that has grown up around the spec, proxying an existing MCP-shaped tool server into the browser API.\n\n**Adoption happens at the platform/vendor layer, not the individual property.** No hotel in this sample built its own WebMCP integration from scratch. The two newly-found sites in this run only surfaced because this pass also checked `<script src>` URLs, not just inline page text — a raw keyword search alone would have kept missing them, which is itself evidence that a text-only WebMCP census will always undercount.\n\n## Reconciling three llms.txt numbers\n\nThree measurements of the same country’s llms.txt adoption, six months apart at the widest gap, tell a consistent growth story once the sample and verification method are held constant.\n\nFrance llms.txt adoption across three measurements\n\nWhen\n\nSample (n)\n\nllms.txt rate\n\nSource\n\nMarch 2026\n\n17,634\n\n3.8%\n\nhotel-llms-txt-adoption-study-2026 (this site)\n\nSept 10, 2026\n\n18,810\n\n11.2%\n\nfirst WebMCP France pilot (unpublished)\n\nSept 23, 2026\n\n13,808\n\n14.5%\n\nthis study (per-domain, cleaned)\n\nThe jump from 3.8% in March to double digits by September is best explained by platform-driven rollout rather than individual hoteliers acting: Shopify began shipping llms.txt by default in May 2026, and WordPress SEO plugins (AIOSEO, Yoast, Rank Math) added llms.txt generation as a feature in late 2025/early 2026 — the same plugin-and-platform mechanism this site’s own llms.txt study already documented driving roughly a third of global adoption. The smaller move from 11.2% to 14.5% over the following two weeks sits inside that same trend, not a separate event.\n\n## Estimating the lift\n\n_Everything here is a reasoned estimate, not a measured cost — no dataset, ours or otherwise, contains actual WebMCP implementation-cost data. This framework is argued from first principles plus the readiness baseline above._\n\n### Tier 0 — no structured data, generic CMS template\n\nRoughly a third of hotels globally have no schema at all (36.3%, per this site’s schema study), and even after six months of platform-driven growth, 85.5% of French hotel domains still have no llms.txt. For this tier, WebMCP isn’t a marginal add — it requires first building or exposing some machine-queryable representation of rooms, rates and availability before there’s anything to wrap in a `registerTool()` call. This is the largest tier.\n\n### Tier 1 — existing structured data or a live API, no agent-facing declaration\n\nFor sites with a working internal REST API already, the lift is closer to a thin adapter: a script that calls the existing endpoint and registers it as a tool with a typed description. The infrastructure exists; only the declaration layer is missing.\n\n### Tier 2 — already publishing llms.txt and correctly-typed schema\n\nThis site’s llms.txt study found hotels with llms.txt present average a schema score of 22.4/100 versus 13.8/100 without it — 62% higher, globally. Not proof of WebMCP readiness, but evidence the same operators who invest in one machine-readable declaration tend to invest in more than one — a head start in operational habit more than in engineering.\n\nFor an independent hotel with none of this in place, implementation is gated less by the WebMCP code itself — a modest scope for one tool, comparable to documenting a single internal API endpoint — and more by what its booking-engine vendor supports. Most independent hotels embed a third-party booking widget rather than running their own reservation system. Until that vendor ships a WebMCP tool, an individual hotel’s realistic path runs through its PMS/booking-engine provider, not through its own developer writing `registerTool()` calls against a black-box iframe — exactly the pattern this study found: three vendors, zero in-house implementations.\n\nOpen questions\n\n## What would need to be true for this to matter\n\nRegistration existing, as this test shows it does at a small base rate, is not the same as anything consuming it.\n\n-   Gemini in Chrome is the first named, committed mainstream consumer, targeted for late 2026 — a concrete step up from “no platform found consuming this” two weeks earlier, but still unshipped as of this writing. The 26 registrations this study found are tools with a named future shopper, not a confirmed one.\n-   Firefox (neutral, Aug 5) and WebKit (opposed, June 11) have both closed their standards-position discussions rather than leaving them pending — a one-browser-family standard is a materially weaker signal than a cross-browser one, and neither has moved toward implementing it.\n-   The consent/security model — arguably the precondition for a hotel exposing a booking-capable, not just search, tool — is still limited to a coarse Permissions-Policy allowlist, with no per-tool user consent flow specified.\n-   Separately, and more consequentially for hotels near-term: Booking.com, Google (via UCP for Lodging) and OpenAI’s ChatGPT apps already give agents a server-side tool-calling path into hotel inventory that has nothing to do with WebMCP and is materially more mature. A hotel deciding where to spend limited engineering time in 2026 arguably gets more from being reachable through those channels than from an experimental browser API with a named but unshipped consumer.\n\n## What this means\n\n-   **Check your own booking-engine and Shopify vendors before writing any code.** Every confirmed WebMCP registration in this sample arrived through a vendor, not a hotel’s own developer — if a property runs Shopify anywhere (a gift shop, a wine club) or a shared booking widget, it may already have a live tool without anyone knowing.\n-   **The realistic near-term move is llms.txt and clean schema.** Both are simple, static, and already growing fast platform-by-platform; WebMCP can wait — it has no confirmed consumer yet and an unresolved consent model.\n-   **Server-side tool-calling paths are more mature right now.** Booking.com’s MCP server, Google’s UCP for Lodging, and ChatGPT apps all give agents a path into hotel inventory today, independent of whether WebMCP ever gets a real consumer.\n-   **Revisit this once Gemini in Chrome actually ships.** That is the single event most likely to convert this from “experimental primitive” to “worth a hotel’s engineering time.”\n\n## FAQ\n\nYes: of 13,808 real French hotel websites checked directly, **25 domains (26 property listings)** have a verified `document.modelContext.registerTool()` call running, confirmed by reading the served JavaScript. All of it arrived via three third-party vendors — none of the hotels built it themselves.\n\nSources\n\nWebMCP spec\n\n-   [github.com/webmachinelearning/webmcp](https://github.com/webmachinelearning/webmcp)\n-   [developer.chrome.com/docs/ai/webmcp](https://developer.chrome.com/docs/ai/webmcp)\n-   [developer.chrome.com/blog/ai-webmcp-origin-trial](https://developer.chrome.com/blog/ai-webmcp-origin-trial)\n-   [Mozilla standards-position (neutral, closed Aug 5 2026)](https://github.com/mozilla/standards-positions/issues/1412)\n-   [WebKit standards-position (oppose, closed June 11 2026)](https://github.com/WebKit/standards-positions/issues/670)\n-   [Spronta, \"State of WebMCP\" (July 2026)](https://www.spronta.com/blog/state-of-webmcp-july-2026/)\n-   [dev.to — Gemini in Chrome commitment (edited Sept 7 2026)](https://dev.to/r0bertini/gemini-in-chrome-is-about-to-call-webmcp-the-no-agent-uses-it-yet-excuse-just-got-an-expiry-date-51be)\n-   [Shopify changelog — WebMCP on Liquid/Hydrogen (Aug 5 2026)](https://shopify.dev/changelog/webmcp-liquid-hydrogen)\n\nIndustry precedent\n\n-   [developers.booking.com/mcp-server/docs/about](https://developers.booking.com/mcp-server/docs/about)\n-   [Skift — Google's UCP for Lodging launch](https://skift.com/2026/08/27/googles-agentic-hotel-booking-tool-comes-to-ai-mode/)\n-   [ppc.land — UCP expansion (non-lodging)](https://ppc.land/google-expands-ucp-to-hotels-food-delivery-and-three-new-countries/)\n-   [OpenAI — Apps in ChatGPT](https://openai.com/index/introducing-apps-in-chatgpt/)\n-   [github.com/openbnb-org/mcp-server-airbnb](https://github.com/openbnb-org/mcp-server-airbnb)\n\n### Summarize with AI\n\n### More AI-search infrastructure research\n\n[Hotel llms.txt Adoption Study](/research/hotel-llms-txt-adoption-study-2026)[Hotel robots.txt & AI Blocking Study](/research/hotel-robots-ai-blocking-study-2026)[Hotel Schema.org Adoption Study](/research/hotel-schema-adoption-study-2026)[All research](/research)","author":{"@type":"Person","name":"Nicolas Sitter","url":"https://nicolassitter.com/about","sameAs":["https://www.linkedin.com/in/nicolassitternolleau/","https://github.com/Nicositter88","https://hotelrank.ai"]},"publisher":{"@type":"Person","name":"Nicolas Sitter","url":"https://nicolassitter.com"},"image":"https://nicolassitter.com/api/og/webmcp-hotels-france-test-2026","mainEntityOfPage":{"@type":"WebPage","@id":"https://nicolassitter.com/research/webmcp-hotels-france-test-2026"},"tags":["AI Search","Hotels","WebMCP","Agentic Web","GEO"],"sameAs":["https://hotelrank.ai/research/webmcp-hotels-france-test-2026"],"alternateFormat":{"html":"https://nicolassitter.com/research/webmcp-hotels-france-test-2026","json":"https://nicolassitter.com/api/post/webmcp-hotels-france-test-2026","rss":"https://nicolassitter.com/rss.xml"},"datasets":[{"name":"summary","contentUrl":"https://nicolassitter.com/data/webmcp-hotels-france-test-2026/summary.csv","encodingFormat":"text/csv"}]}