A normal website publishes a representation of an organization. An agentic website begins to participate in the organization itself. It can expose current state in machine-readable form, advertise what it can do, accept bounded requests, coordinate with other systems, and update what it shows when the underlying reality changes.
The distinction matters because adding a chatbot does not automatically make a website agentic. Conversation is only an interface. The deeper mechanism is a loop: observe state, interpret a request or change, choose an allowed action, execute it through an authenticated capability, record the result, and expose the new state.
Open agent protocols are making part of this architecture more concrete. Google’s guide to agent protocols describes A2A as a way for agents to discover and communicate with one another through published capability metadata, while MCP addresses access to tools and data. The important shift is that a machine no longer has to infer everything a digital property can do from visual pages alone.
That suggests a useful minimum architecture for an agentic website. It needs a canonical identity, a machine-readable description of current state, a capability boundary, explicit permissions, durable records of actions, and a way for other machines to discover the relevant endpoints. An agent card or site manifest can advertise the surface; the runtime behind it determines whether those advertised capabilities are real.
State synchronization is the next threshold. A site that says one thing while its database, repository, deployment, or external services say another is not meaningfully agentic—it is simply automated. Stronger systems treat public output as a projection of authoritative state and continuously test whether those projections still agree.
That creates a spectrum. At one end are static pages. Next come dynamically rendered sites, then sites with APIs and machine-readable metadata, then systems that can safely take actions. At the far end are websites that can observe changes across their own infrastructure, decide which updates matter, modify their public representation, and leave an auditable trail without waiting for a person to edit every page.
The web is already moving toward standardized machine discovery. A2A uses published agent metadata so one agent can learn what another can do before contacting it. The same design principle applies more broadly: identity, capability, authority, and current state should be explicit rather than guessed.
The useful question is therefore not whether a site “has AI.” It is whether the site can maintain a truthful relationship between the thing it represents and the actions it is allowed to take. When that loop closes, the website stops being only a destination. It becomes an operating surface.