What is an agentic website? A working definition and a five-layer model

Published October 14, 2026 · Last reviewed October 16, 2026

An agentic website is a website that answers and sells to people and their AI agents. In practice it is readable, so a machine can parse it; addressable, so a person or a visiting agent can ask it a question and get a straight answer; and able to transact, so the conversation can end in a booked demo, a paid order or a trial. Search for the term today and you get at least three different meanings, from “built for agents” to “run by AI” to “built by AI”. This piece settles on one working definition and breaks it into five layers: readable, discoverable, addressable, transactable and accountable.

The layers matter because they are measured very differently. Cloudflare’s April 2026 scan of a filtered set of the 200,000 most visited domains found markdown content negotiation on 3.9% of sites, and MCP Server Cards and API Catalogs together on fewer than 15 sites in the whole dataset.¹ Most sites have not finished the first layer, and we know of no public count for the third. Below, each layer is mapped to the standards that exist for it, with an honest note on who reads it today.

Three meanings of “agentic website” in search today

Nobody owns this term yet. Read the top results side by side and three overlapping camps appear, each starting from a different place.

Built for agents. Search Atlas defines an agentic website as one “designed for autonomous AI agents rather than only human visitors.”² Its version goes well past markup: the site exposes data, actions and workflows through tool-callable endpoints such as APIs and MCP servers, so an agent can submit a request or complete a transaction. It is the closest to what most people mean, and it describes a machine interface: it says little about the question a visiting agent brings when no tool covers it, or about who on the business side learns what happened.

Run by AI. Automaton Agency’s June 2026 explainer describes a site that “runs its own growth and lets AI agents operate it directly”: wired to its own data, improving itself with a person on the judgment calls, and open to outside agents that read and act on it.³ Lazarev Agency sells a service built on the production half of that idea: a site that AI agents “draft, structure, and update”, with a person approving changes before they go live.⁴ Here the weight falls on how the site is produced and maintained. That changes how marketing teams work, but it says less about what a buyer’s assistant experiences on arrival.

Built by AI. Then there are AI website builders that generate a site from a prompt, some of them now sold as agentic; Hostinger’s AI Builder, for one, has an “agentic mode” in which the AI leads.⁵ Automaton draws this line itself, noting that a builder generates a site once and the result stays static.³ A site made by an AI is not, by that fact, a site an AI can talk to.

Each camp describes one property of a website. None starts from the question a site owner has: when a buyer sends ChatGPT, Claude or Gemini to check out my business, what happens next?

That question matters more in 2026 than it did a year ago. In March, OpenAI said the first version of Instant Checkout in ChatGPT lacked the flexibility it wanted, let merchants use their own checkout experiences, and turned its focus to product discovery.⁶ Adobe data reported by Digital Commerce 360 in August showed AI-referred visits to US retail sites converting 60% better than other traffic, with 53% more revenue per visit.⁷ Those are retail numbers, and B2B is not retail, but the direction is the same: the assistant finds you, and the decision lands back on your own site, which then has to do more than be read.

A working definition and the five-layer model

The definition this blog uses, and the one the rest of this series builds on: an agentic website answers and sells to people and their AI agents. It is readable, so software can parse it. It is addressable, so a person or a visiting agent has something to ask. And it can transact, so a good conversation ends in a result instead of a transcript.

The definition is about the visitor’s experience, not the production method, and it covers both audiences at once.

The outside reference point is Cloudflare. In August 2026 it described its picture of the agentic internet in four words: “readable, discoverable, callable, and payable.”⁸ That framing is written from the network’s side, about the infrastructure a site needs so agents can consume it, call it and pay it. The model below keeps readable and discoverable, folds callable and payable into one transactable layer, and adds two layers that an infrastructure view leaves out: addressable, a site that answers questions, and accountable, what reaches the human who owns the business.

The last column is our reading of public data where it exists (Cloudflare, Ahrefs) and a plain “early” where it does not. The layers are cumulative but not automatic: a site can have a perfect sitemap and nothing to ask, or a capable chatbot that no agent can find. The “who reads it today” column thins as you move down. That shows how few sites have reached the lower layers, not that they matter less.

This extends the split in our explainer on agent-readable vs agent-addressable: readable and discoverable are the format side; addressable, transactable and accountable are the conversation side.

Readable and discoverable: the layers most sites already started

Readable is the oldest layer and the best documented. An agent should be able to get your content without rendering a page built for human eyes. Semantic HTML does most of the work. Markdown content negotiation, where a server returns a lighter Markdown version when an agent asks for it, saves the agent tokens and saves you bandwidth: in Cloudflare’s own example, one of its blog posts drops from 16,180 tokens as HTML to 3,150 as Markdown.⁹ Cloudflare counted the feature on 3.9% of the sites in its April scan.¹

llms.txt belongs here too, as a curated map of the pages that matter, proposed by Jeremy Howard in September 2024 as a convention rather than a standard.¹⁰ It is cheap to publish and worth keeping honest about. Ahrefs looked at 137,210 domains in its analytics in May 2026: 28% published an llms.txt file, a share Ahrefs treats as an upper bound, and 97% of those files received zero requests that month.¹¹ Google’s own Search guidance, updated in July, says Google Search ignores such files, so creating one “will neither harm nor help” your visibility there.¹² Publishing one is fine. Expecting it to get you cited is not supported by any data we have seen. Our longer read on whether llms.txt actually works goes through the studies.

Discoverable is about being found and being clear about the rules. Sitemaps and schema.org markup are the established part, because search and AI crawlers already read them. robots.txt is nearly universal (78% of the sites in Cloudflare’s scan), but only about 4% declare AI usage preferences through Content Signals.¹ Newer pieces sit on top. An A2A agent card, a JSON file at /.well-known/agent-card.json, tells other agents what an agent can do and how to reach it; it matters once you run an agent worth describing, and our explainer on how A2A agent cards work covers the format. Google’s Agentic Resource Discovery proposes a catalog at /.well-known/ard.json that points agents to such cards and to MCP servers. Web Bot Auth lets well-behaved agents sign their requests so a site can tell them apart from scrapers, which is the subject of a separate piece.

Our 2026 guide to making a website agent-ready is the practical checklist for these two layers.

The limit of both layers is the same. They describe what is on your site. They do not respond. A visiting agent that reads your pricing page and still has a question, about a plan, a seat count, an integration or a contract term, has nowhere to put it. It may guess, check a software directory, or move on to a competitor whose page happened to state the answer. In Kevin Indig and Siteline’s July 2026 test of 100 B2B products, agents asked about pricing cited a third-party source in 45% of runs when the vendor showed no real price, against 18% when a price was on the page.¹³ Readable and discoverable get you into the comparison. They do not let you take part in it.

Addressable: something on the site an agent can ask

The third layer is where a site stops being read and starts being asked. An addressable website has a live endpoint that takes a question and returns an answer: what does the Business plan include, do you integrate with HubSpot, can you onboard 30 seats before the end of the quarter, is there someone in Lisbon I can talk to. The asker can be a person typing into a chat window or an assistant acting for that person. Either way, the site responds instead of waiting to be read.

It takes three shapes today.

• A site agent. An AI agent on the site that answers from the company’s own materials and converses with whoever arrives. Many website chat products were built for human visitors first.
• A natural-language endpoint. Microsoft’s NLWeb, announced in May 2025, gives a site an /ask endpoint that accepts a question in plain language and returns structured answers built on schema.org types, and each instance can also act as an MCP server.¹⁴ Public adoption figures are thin, so treat it as an option to watch rather than a default.
• A structured guest contract. A machine-readable description of what a visiting agent may ask and in what format it gets the answer, so an assistant does not have to scrape a chat widget designed for humans.

What separates these from a classic chatbot is who they expect on the other side. A chatbot assumes a human with patience and a browser. An addressable endpoint assumes the visitor might be software that wants exact facts in its own format, will not click through a carousel, and will report back to a human who never saw your site. That is why “not a chatbot” sums up the layer, even when the product underneath also chats with people.

The buyer side is starting to describe the same choice. When Meta and Sierra announced the Personal Agent Protocol on 6 October 2026, they listed three ways a personal agent works with a company: through its regular web pages, through its APIs, or through its own agent for “tasks that need conversation.”¹⁵ That third route is the addressable layer; our read of the protocol covers what it settles and what it leaves open.

Who reads it today: early. We have not seen independent data on how often assistants call a site’s endpoint. The structural point holds anyway. A buyer’s agent that hits a question your pages do not answer has two choices on a site with no endpoint: guess, or leave. On an addressable site it has a third: ask.

This is the layer Tobira works on. Tobira Site Agent turns your website into an agentic site: it publishes the machine-readable files AI agents look for and gives them a live endpoint to ask and act, so the site becomes readable and addressable at once. The page’s FAQ names those files as “llms.txt, agent.json and AGENTS.md” (as listed on tobira.ai/agentic-site/ on 14 October 2026).¹⁶ The files describe the site; the endpoint is what answers. On their own, as the previous section shows, files are not a route to being cited. The Tobira network behind the product, which counted 729 agents as of September 2026 per Tobira internal data, is in early access.

Transactable and accountable: where the conversation ends

An answer is useful. An answer that ends in a booked call is a lead. The fourth layer, transactable, is about letting the visit finish: a calendar slot, a checkout, a trial sign-up, a quote request with the context already filled in.

For human visitors this is ordinary conversion work. For visiting agents it is where the standards get busy and status matters.

• WebMCP lets a page register tools, such as “book a demo” or “get a quote”, that an agent in the browser can call instead of guessing at buttons. The page exposes them through document.modelContext. It is a Draft Community Group Report from the W3C Web Machine Learning Community Group, not a W3C Standard. Chrome opened an origin trial in Chrome 149, scheduled through Chrome 156, and on 28 September 2026 its team filed to extend the trial to Chrome 162 after developer feedback suggested changes to the API. Edge 153 lists its own origin trial, and OpenAI’s ChatGPT desktop app implements WebMCP in its built-in browser as Site tools, subject to rollout. Neither Chrome nor Edge has shipped it to all users.¹⁷ Build against it as an experiment, not as settled infrastructure.
• Agent payment protocols each handle approval differently. Google’s AP2 builds a payment from cryptographically signed mandates¹⁸ and was donated to the FIDO Alliance in April 2026;¹⁹ ACP, developed by Stripe and OpenAI,²⁰ and x402, now stewarded by the x402 Foundation under the Linux Foundation,²¹ take other routes. Our agent payments explainer compares them. Approval stays with the buyer: Stripe’s own WebMCP guide flags its payment-submission tools so an agent gets the user’s confirmation before calling them.²² A site should never assume an agent can pay without its human signing off.
• Most live rails are retail. Product feeds and agent checkouts were built for a SKU, a price and a cart. Services and B2B sell fit and scope, which is why agentic commerce looks different for services: the transaction is often a qualified meeting, not a basket.

The fifth layer, accountable, is the one infrastructure framings leave out, and the one a business owner cares about most. When an agent books a demo for its human, who on your side knows, and what do they know? An accountable site hands off to a person at the right moment, passes the lead with the context of the conversation, only shares what the visitor agreed to share, and tells the owner what happened. The phrase on Tobira’s product page is a good test for any vendor: you get the result, not the transcript.

Accountability also covers the rules. A site that answers on your behalf needs a short list of things it must never promise, never reveal, and always route to a human. Without that, the transactable layer is a liability. With it, the agent becomes an extension of your sales team rather than a guess generator with your logo on it. Our piece on a website for humans and their agents looks at the same handoff from the buyer’s end.

Which layer to fix first, by site type

Nobody needs all five layers on day one. The right order depends on how your business closes a deal.

B2B SaaS. The questions that lose deals are about plans, seats, integrations and security. Make pricing and plan limits explicit on the page first (readable), because an agent that cannot find a number may take one from a directory instead, as the Siteline test above found. Then add an addressable layer that can answer the follow-ups and book a demo with the context attached. A demo booked by an assistant is still a demo; what matters is that your sales team knows what was asked before the call.

Agencies and consultancies. You sell fit, not a SKU. Your pages rarely answer “can you do this for a company like mine, in this timeframe, at roughly this budget”, and that is exactly what a buyer’s agent asks. Go to the addressable and accountable layers early: an endpoint that can qualify the request and route it to the right person with a summary beats any amount of markup.

Professional and local services. Lawyers, accountants, clinics and similar firms usually close with a booked consultation. The fastest gain is often transactable: a calendar inside the conversation, with clear rules about what the agent will and will not say. Keep the discoverable layer honest too, with accurate opening hours, locations and service areas, since an assistant that finds the wrong hours will pass them on.

Ecommerce. Retail has the most developed transactable rails, mostly handled by your commerce platform. The gap for many stores is the addressable layer for pre-purchase questions about sizing, compatibility and delivery, which feeds and checkouts do not answer.

Across all four, do the readable and discoverable work because it is cheap and expected, not because it will get you recommended. Then spend real effort on the layer where your site goes silent: for most B2B companies, the moment after an agent has read everything and still has a question.

If you are still deciding what “turning your website into an AI agent” should mean for your business, our earlier piece on the three meanings asks which one makes money. For how these layers fit with MCP and A2A at the protocol level, see the three-layer agent stack explainer.

What to remember

• An agentic website answers and sells to people and their AI agents: readable, addressable, and able to transact.
• The term has at least three meanings in search today. “Built by AI” describes how a site is made and “run by AI” mostly how it is maintained; neither starts from what a visiting agent can do with it.
• The five layers are readable, discoverable, addressable, transactable and accountable. The first two describe your site; the last three let it respond, act and report back.
• The readable layer is cheap and worth doing, but no data we have seen shows that files like llms.txt get a site cited on their own.
• Cloudflare’s April 2026 scan found markdown negotiation on 3.9% of sites and Content Signals on about 4%. No public scan we know of counts addressable endpoints, and that gap is the opening.
• Which layer comes first depends on how you close: plans and pricing for SaaS, qualification and routing for agencies, a calendar in the conversation for local services.

Pick one question your site cannot answer today, and start there.

FAQ

What is an agentic website? An agentic website is a website that answers and sells to people and their AI agents. It is readable, so software can parse it; addressable, so a person or an assistant can ask it questions; and it can transact, so the conversation ends in a booked demo, an order or a trial, with a person on the owner’s side getting the result.

Is an agentic website the same as a site made with an AI website builder? No. An AI website builder uses AI to generate the site, and the result can be as static as any other page. An agentic website is defined by what visitors can do with it, including visiting AI agents. A hand-built site can be agentic, and an AI-built one may not be.

Does publishing llms.txt make my website agentic? Not by itself. llms.txt is a readable-layer convention, a curated map of your important pages. Ahrefs found that 97% of the llms.txt files in its May 2026 sample received no requests at all, and Google’s Search guidance says Google Search ignores such files. It is cheap to publish, but an agentic website also needs something an agent can ask and a way to act.

What are the five layers of an agentic website? Readable (content an agent can parse cheaply), discoverable (sitemaps, schema, robots rules, agent cards), addressable (a live endpoint that answers questions), transactable (booking, checkout or tools an agent can call) and accountable (human handoff, consent and a summary for the owner). The model extends Cloudflare’s August 2026 framing of a readable, discoverable, callable and payable internet.

Is an agentic website the same as the agentic web? No. The agentic web is the wider shift: an internet that agents can read, find, call and pay, in Cloudflare’s framing. An agentic website is one business’s part of it, a site that people and their AI agents can read, ask and buy from, with the outcome reaching a person on the owner’s side.

How is an agentic website different from a website chatbot? A typical chatbot is built for a human with a browser and patience. The addressable layer of an agentic website also expects software on the other side: an assistant that wants exact facts in its own format and will report back to a person who never saw your site. Many products do both; the difference is who the endpoint is designed to serve.