The agentic website checklist: what to ship at each layer, and who actually reads it today

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

An agentic website checklist tests the five layers of a site that answers and sells to people and their AI agents: readable, discoverable, addressable, transactable and accountable. Most agent-readiness checklists, including the two on this blog, stop after the first two: clean HTML, a sitemap, schema, robots rules, maybe an llms.txt file. That work is real, and it is the floor. A checklist that ends there leaves out the part where a buyer’s assistant asks a question, gets a straight answer and books the call. We set out the five layers when we defined the agentic website on 14 October.

Each row below gives the effort, what you get once it passes and, the column most lists skip, who reads it today, with dated evidence. Where we found no reader, the row says “no confirmed reader yet” instead of guessing.

The first two layers get one line per check and a link to the longer guides. The rest goes to the three layers above them, where most sites probably have nothing and where a visiting agent decides whether to stay.

How to read this checklist

• Effort. S is an afternoon for one person, M is a sprint, L means new software or a vendor.
• What you get. What a person or a visiting agent can do once the check passes.
• Who reads it today. Evidence that something consumes the row, with a date. Counts of how many sites publish something are labelled as adoption, because a file on many sites does not prove that anything reads it.

Treat the layers as five gates, in order, and stop at the first one where your site goes silent. A cheap row with no confirmed reader can still be worth an afternoon; a row with a measured reader, such as pricing in plain text, earns its place even when the work is harder.

Readiness scores cover only part of this. Cloudflare’s Agent Readiness score grades four buckets (discoverability, content, bot access control, capabilities) and checks commerce standards without counting them,¹ and its August dashboard update still treats commerce as informational.² None of that asks whether an agent got the commercial answer it came for, or whether a person on your side heard about the visit; our piece on what to do after you pass the score covers that gap. Cloudflare argues that businesses easy for agents “to find, understand, and trust are the ones that get recommended.”² That is a reasonable bet, not a measurement, and nothing on this list promises that an assistant will cite you. The checks aim at something easier to verify: when a person or their agent arrives, the site can answer, act and hand over the result.

Readable and discoverable: the floor

These two layers are covered in depth in our 2026 guide to making a website agent-ready and the agent-readiness audit checklist. Here they get one line each.

Two rows need more than a table cell. On llms.txt, publishing is easy to measure and reading is not: SEOmator found a real file on 8.6% of Cloudflare Radar’s top 100,000 domains on 13 September 2026, nearly a third of them generated automatically by the site’s platform.¹¹ Publish one because it is cheap, not as a step toward being cited; our read on whether llms.txt works goes through the studies.

On the crawler split, Cloudflare’s Agent category means automated traffic working in real time for a person. Since 15 September, a new domain that earns money from ads is offered a preset that blocks Agent crawlers on pages that show ads, while Search stays allowed and AI training is disallowed; existing settings carry over.⁷ If your site runs ads and you want buyers’ assistants to reach your pricing, check that setting. Chrome’s Lighthouse agentic browsing audit is a useful second opinion on both layers.

Addressable: can an agent get an exact answer?

The addressable layer is where a site stops being read and starts being asked. The test is simple: a buyer’s assistant has read your pages and wants to know whether the Team plan covers 30 seats, whether you integrate with HubSpot, or whether someone can start in November. Does anything on your side answer?

Pricing first, because it is the one row with hard data. In July 2026 Kevin Indig, working with Siteline founder David Kaufman, gave AI agents three buyer tasks for each of 100 B2B products, five runs per task. Agents answered pricing and feature questions from the vendor’s own site 79% of the time, the weakest of the three categories, against 93% for integrations and 92% for security. With no real price shown, the agent cited at least one third-party source in 45% of pricing runs, against 18% with a numeric price on the page.³ If your price lives behind “contact us”, a review site or a reseller may be answering for you. For custom tiers, the memo advises explaining what drives the price instead of offering only a contact-sales button, and a plain-text pricing model (“from $X per seat, annual billing, minimum Y seats”) does that while final quotes stay custom.

Then the endpoint. Pages answer the questions you predicted; an endpoint answers the ones you did not. It can be a site agent that talks with people and their assistants, or an NLWeb /ask endpoint, Microsoft’s open project that returns ranked, structured results for a plain-language question.¹² We know of no public scan that counts either, which is also the opening: most sites in your category probably have nothing here. One draft to watch is the Personal Agent Protocol from Sierra and Meta. Its draft of 9 October has a company publish /.well-known/poppy.json to tell personal agents how to work with it, with OpenAI, Visa and PayPal among the listed design partners. Sierra has shown it live with Meta’s agent Muse, but we know of no production agent that reads it yet.¹³ Our explainer of the 6 October announcement covers the access model.

An agent that answers from your materials, and hands over to a person when it runs out, carries a different risk from one that improvises. Our read of the OLG Hamm and Moffatt chatbot cases shows a German court and a Canadian tribunal treating what a site’s chatbot says as what the business says.

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 is readable and addressable at once. Its product page lists “Site ready for AI agents” in the Lite plan, and every higher plan includes it; it says the agent answers only from your materials and offers a call when the answer is not there.¹⁴ As listed on tobira.ai/agentic-site/ on 16 October 2026, there is a free plan that runs on your own AI key, and paid plans at $49, $169, $399 and $799 a month (USD, excluding taxes), with no separate charge for agent traffic.¹⁴ The distinction from our readable vs addressable explainer still holds: the files describe the site, and the endpoint is what answers.

Transactable: can the visit end in a booking, a trial or a payment?

Picture an assistant that has confirmed your price and your HubSpot integration and now has to put a 30-minute demo into its user’s calendar for Tuesday. The transactable layer asks whether that can happen on your site while the buyer, or the assistant acting for them, is ready.

Booking is the transaction for most B2B and services sites. You sell fit and scope, not a SKU, so the “purchase” is usually a qualified meeting, a paid consultation or a trial; that is why agentic commerce looks different for services, and why the calendar row matters more than any payment protocol. The reader is no longer only human. Google’s Chrome auto browse (in preview for US AI Pro and Ultra subscribers since January) and the Chrome browsing in Gemini Spark (rolling out first in the US since 30 July) have been used to schedule appointments and viewings on websites, and they pause before purchases and hand payments back to the user.¹⁵ For restaurants, bookings made in a chat mostly run through partner platforms, such as Yelp reservations in ChatGPT (US and Canada) or the partners Google’s AI Mode links to for the last step (UK).¹⁷ If your calendar sits on no such platform, the assistant has to work your own booking page, so keep it short and plain. It is also unlikely to wait for a callback, the argument of our piece on speed to lead when the buyer is an agent.

Payment rows need a person in the loop. Stripe’s WebMCP guide sets a hint on its payment-submission tools and tells agent builders to get the user’s confirmation before calling them; the hint itself is not confirmation.¹⁸ Build for that pattern: the agent prepares the order, the human approves it, and the site never assumes a visiting agent can pay on its own authority. Our agent payments explainer compares how AP2, ACP, x402 and MPP handle approval, including AP2’s signed mandates for spending agreed in advance, and our read of browser approval gates shows how thin some gates still are.

WebMCP is a row to try, not to depend on. It lets a page register tools through document.modelContext that an agent in the browser can call. The spec is a W3C Community Group draft, updated 15 October 2026, not a W3C Standard. Chrome runs it as an origin trial (an opt-in test) from Chrome 149 through 156, with a desktop extension to Chrome 162 filed on 28 September, and Edge 153 has its own trial.¹⁹ It also has a first agent reader: OpenAI describes Site tools, in the ChatGPT desktop app’s built-in browser, as its implementation of the proposed standard, with its usual confirmation rules for purchases.²⁰ If your developers have a free week, wrap your booking and quote actions as tools and watch the logs; our July WebMCP adoption tracker lists the early sites. Otherwise the calendar row does more for you this quarter, and checkout protocols such as UCP and ACP, built around a product, a price and a cart, can wait unless you sell from a catalog.

Accountable: does a person get the lead, with context and consent?

The fifth layer is the one infrastructure checklists leave out, and the one a business owner cares about most. When an assistant books a demo for its human, who on your side knows, and what do they know?

The summary row is the cheapest test of the whole layer. If a lead arrives as a 40-message transcript, nobody reads it, and the visit might as well not have happened. Tobira’s product page makes the same point: the owner gets a short summary instead of the transcript, covering who it is, what they need and what was booked or paid.¹⁴ Whatever tool you use, ask for that shape.

Source tagging is how you learn whether any of this works, and the early numbers disagree. Orbit Media’s study of 97 B2B sites found visitors from AI three times as likely as other organic visitors to become leads, though they were just 0.5% of sessions; Ruler Analytics, across 13 industries, found AI referrals averaging 5.8%, above organic search overall (4.9%), but below it in professional services (6.0% against 8.1%) and level with it in software (7.9% for both).²¹ Orbit also notes that AI mobile apps often strip the referrer, so those visits land in Direct, and many agent visits are crawlers rather than buyers, as our AI agent traffic statistics for 2026 show. Tag leads by what the conversation says, not only by the referrer, and after a quarter compare their close rate with your other channels.

Consent and rules can share a page. A visiting agent may hold its human’s name, email and budget, so an accountable site asks before it stores or forwards contact details and records that it asked, the logic of the consent layer for the agentic web. The rule set needs three lists: what the agent never promises (discounts, delivery dates, legal or medical conclusions), never reveals (internal notes, other clients, anything not public) and always routes to a person (contracts, complaints, deals above a set size).

The AI notice is a legal row in the EU. Since 2 August 2026, Article 50(1) of the AI Act has required providers of AI systems that talk directly with people to design them so that people are told they are dealing with an AI, unless that is obvious.²² If you deploy a vendor’s chatbot, check that the notice appears from the first message; if you build your own, you may be the provider. In the US, California’s bot law makes disclosure the safe course for a bot used to sell, and Maine has required a clear notice since September 2025 where a consumer could mistake an AI chatbot for a person.²³ Our explainer on the August deadline covers the EU side.

What WordPress, Wix, Webflow and Hostinger now ship, by layer

If your site runs on a mainstream platform, some rows above may already be a toggle. Here is what four of them ship as of mid-October 2026, read against the five layers.

WordPress. WordPress 7.0 “Armstrong” (20 May 2026) added an AI Client and a Connectors screen to core, alongside the Abilities API from 6.9, which lets core and plugins register what a site can do in a machine-readable way; 7.1 (19 August 2026) added a single flag that marks an ability public for outside clients.²⁴ The separate MCP Adapter plugin from the AI Building Blocks initiative exposes those abilities as tools, resources and prompts to clients that speak MCP, the Model Context Protocol.²⁵ Abilities are private by default, so a plugin that registers “book an appointment” must also mark it public, and its permission checks still apply. Today this mostly serves the owner’s own assistants and developer tools, not a visiting buyer’s assistant.

Wix. With Microsoft, Wix lets owners connect a site to NLWeb in one click, starting with premium domains in English-language markets.²⁶ AI systems can then query the site’s public structured content (schema markup, product feeds, RSS) through an ask endpoint, without any change to how search engines crawl or rank the site. Wix also says every site is its own MCP server at /_api/mcp, through which a visitor’s AI tool can ask about the business, make a booking or buy items.²⁷ That reaches into the transactable layer, but only for visitors who connect an MCP-capable tool, and we found no data on how many do. Rules on what the agent may promise, the handoff and an owner summary are still separate work.

Webflow. At Webflow Conf (1 to 3 September 2026), Webflow said Agent Instructions can now be generated automatically. They are skills and rules, stored as markdown, that tell an agent connected through Webflow’s MCP server how to structure content, apply the design system or follow team conventions: help for agents that build your site, close in spirit to AGENTS.md for coding agents, not for a visiting buyer’s assistant. For visitors, Webflow offers Markdown for AI agents on Enterprise sites (since 2 April 2026) and an llms.txt upload, both readable-layer rows.²⁸

Hostinger. Web2Agent is an experimental Hostinger feature for WordPress sites and some AI Builder sites it hosts. It syncs site content to a service that MCP clients can query, and Hostinger says it “currently works best with Claude, Cursor and tools supporting MCP protocol,” with ChatGPT, Gemini and other agents still in progress.²⁹ Like the Wix option, it makes content askable. It does not decide what the agent should never promise, book the meeting or tell your team who came.

Across all four, most platform work sits on the content side or the owner’s side, with Wix’s MCP bookings as the exception to watch. Whether a buyer’s agent gets an exact price, books a call, pays with the human’s approval and reaches a person with a summary still depends on the addressable, transactable and accountable layers.

What to remember

• An agentic website answers and sells to people and their AI agents. Its checklist has five layers: readable, discoverable, addressable, transactable, accountable.
• Readable and discoverable are the floor that readiness scores grade. Do them because they are cheap, and treat adoption as adoption: llms.txt sits on 8.6% of the top 100,000 domains, yet Ahrefs found 97% of files got no requests.
• Pricing is the row with the clearest data: with no real price on the page, agents cited a third party in 45% of pricing runs, against 18% with one.
• For B2B and services, the transaction is a booked meeting or a trial. Put the calendar in the conversation, keep answers to your own materials, and let the human confirm any payment.
• Accountable means a summary for a person, leads tagged by what the conversation says, consent before contact details move, a written rule set and a clear AI notice.
• Platforms now make content queryable, and Wix says visitors’ MCP tools can book. What your agent may promise and who on your team gets the lead are still your decisions.

We cannot show that ticking these rows gets a site cited, and we have not found anyone who can. The rows that end in a booked call do not need that proof.

FAQ

What is an agentic website checklist? It is a list of checks for a website that answers and sells to people and their AI agents. It covers five layers: readable, discoverable, addressable, transactable and accountable. Unlike a basic agent-readiness list, it goes past files and markup to whether an agent can get an exact answer, finish a booking or payment with the human’s approval, and whether a person on your side gets the lead.

How is it different from an agent-ready website checklist? An agent-ready checklist, like Cloudflare’s Agent Readiness score, mostly grades whether agents can find, fetch and parse your site. That is the first two layers here. The agentic website checklist adds the layers above: a live endpoint that answers questions, booking and payment in the flow, and the handoff, summary and consent rules that make the result reach a person.

Will publishing llms.txt or agent files get my site recommended by ChatGPT? We know of no evidence that it does. 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. They are cheap to publish and harmless. What the evidence does show is that agents cite third-party sources more often when a real price is missing from your site.

What should a B2B website fix first? State prices, plan limits and the pricing model in plain text, because agents fall back to third-party sources much more often when no real price is shown. Then add a way to answer follow-up questions and book a demo inside the conversation, with a summary that reaches your sales team. WebMCP tools and agent checkout protocols can come later.

Can an AI agent pay on my website? An agent can prepare an order or a booking, but the browser-agent flows we checked (Chrome auto browse, Gemini Spark, ChatGPT Site tools, Stripe’s WebMCP guide) hand the payment to the human it acts for. Stripe’s guide, for example, marks its payment-submission tools with a hint and tells agent builders to get the user’s confirmation first. Design your flow so a person approves the payment, whatever the assistant.

Do WordPress, Wix, Webflow or Hostinger make my site agentic automatically? Partly. Wix’s NLWeb option and Hostinger’s Web2Agent make content queryable by AI tools, and Wix also lets MCP-connected tools make bookings and purchases. WordPress’s MCP Adapter can expose actions that plugins register and mark public, and Webflow’s Agent Instructions guide agents that build your site. None of them decides what your agent may promise, hands a lead to a person with the context attached or tells your team who came. Those rows are still yours.