The Agentic Web Is Already Being Built — A Hospitality Operator's Read

George Roukas published a long-form analysis last month that has been circulating in hospitality circles. It is worth reading in full — Roukas examines the infrastructure being built beneath the AI consumer experience, and makes a compelling argument that most of the important work is happening out of public view.

I am sharing it here because I think it is one of the clearest explanations of what the "agentic web" actually means for distribution — and because his framing aligns closely with what I have been seeing from the inside of hospitality technology over the past year. I will add my own layer where I have direct experience, and flag clearly where I am reflecting his analysis rather than adding to it.

The core argument — and why it is credible

Roukas opens with a provocation that will resonate with anyone paying attention to hospitality AI: most operators think AI agents are stalled, because the consumer-facing demos have been underwhelming. He argues that is the wrong conclusion to draw.

This tracks with what I have observed. The conversation in hospitality technology has shifted significantly in the past 12 months — away from "will AI do anything useful?" and toward "how do we connect our systems to it properly?" That is an infrastructure conversation, not a product one.

The three infrastructure layers that matter

Roukas identifies three categories of infrastructure currently being built, and I think this framing is genuinely useful for operators trying to understand where to pay attention.

MCP — the connector standard. The Model Context Protocol, originally developed by Anthropic, has achieved remarkable cross-platform adoption in under a year. Roukas cites 97 million monthly SDK downloads and 5,800+ connectors, now governed under the Linux Foundation. I have been building MCP servers for hospitality clients, so I can confirm the practical reality: the tooling works, it is becoming easier to implement, and the major AI platforms are all speaking this language. What does not yet exist is a simple packaged solution a hospitality company can install without development work.

A2A — agent to agent coordination. Google's Agent2Agent protocol allows AI agents to communicate with each other directly. Roukas describes a practical scenario: a traveler's personal agent could negotiate with a hotel's agent to build a package, check availability, and confirm preferences — all without a human interface in the middle. This is further out, but the standard already exists and is gaining adoption.

Registries and governance. Less visible but important — the security and discovery infrastructure that allows organizations to trust and manage which AI connectors they expose. This is the kind of work that never makes headlines but determines whether any of this actually scales.

The demand signal is already there

Roukas cites Phocuswright's March 2026 report on AI travel adoption. These numbers are significant:

Importantly, Roukas distinguishes how travelers are currently using AI: primarily for discovery, research, and comparison — not yet for autonomous booking. The agent handles the cognitively heavy shopping, then hands the traveler off to complete the transaction. He calls this the metasearch model: "The shopping is the disruption, not the checkout button."

The control point — and why it sounds familiar

Roukas makes a point I find particularly sharp. On the human web, power sits with whoever controls the interface — OTAs, search engines, booking platforms. On the agentic web, power shifts to whoever controls the consumer-side agent and its "harness" — the tools, rules, policies, and data that shape what the agent recommends.

OTAs like Booking.com and Expedia have already built MCP connectors that make their inventory accessible to AI platforms. A supplier without structured, machine-readable data and an agent-accessible interface is, effectively, invisible to this channel — even if the channel is not yet generating significant bookings.

The practical implication for operators

Roukas makes an argument I find compelling and worth repeating directly: investments in structured data, offer management, and personalization pay off twice. They improve the human web now — better content, better conversion, better direct booking performance — and they position the same systems for the agentic channel the moment that infrastructure is ready.

For vacation rental operators and independent hotels, the immediate practical question is not "should we build an AI agent?" It is: how well structured and machine-accessible is our data?

Property information, availability, rates, policies, and guest-facing content — if these live in well-structured, accessible systems, they can be exposed to AI channels with manageable effort. If they live in PDFs, spreadsheets, or siloed systems with no API access, the technical gap becomes significant before any AI benefit is possible.

A note on what I am adding — and what is Roukas's work

The statistical data, infrastructure analysis, and distribution framework in this piece are drawn directly from Roukas's article. I have cited him throughout and encourage you to read the original — it is substantially longer and covers the OTA competitive dynamics, the A2A and payment protocols, and the broader economic implications in much greater depth than I have summarized here.

What I have added is the operational lens: what this means specifically for vacation rental operators and independent hotels, and where the practical starting points are based on the work I have been doing in this space. That layer is my own, and it reflects my experience building these systems — not a claim to independent research.