Deutsch
Agent readiness & browser agents

Agent readiness

Agent readiness describes how well AI agents can discover, read and use a website. The term also appears for how prepared a company is to run AI agents in production. This entry covers the website meaning.

At a glance

Type
how suitable a website is for AI agents
Used in
Cloudflare Agent Readiness Score; Lighthouse "Agentic Browsing" (from Chrome 150, experimental)
Status
scoring model from individual vendors, not a standard
Docs
Cloudflare: Agent Readiness Score

Last checked:

How is agent readiness measured?

Cloudflare published an Agent Readiness Score on April 17, 2026, and compares the idea to what Lighthouse does for performance. Its checks fall into four categories:

  • Discoverability: robots.txt (RFC 9309), sitemap.xml and Link headers (RFC 8288).
  • Content: serving Markdown to agents through content negotiation (Accept: text/markdown).
  • Bot access control: Content Signals, AI bot rules in robots.txt and Web Bot Auth.
  • Capabilities: Agent Skills discovery, an API catalog (RFC 9727), OAuth server discovery (RFC 8414, RFC 9728), an MCP Server Card and WebMCP.

Cloudflare’s post lists commerce standards such as UCP, ACP and x402 separately from the score, and the isitagentready.com tool now shows them as a fifth category, Commerce. The score is a model from one vendor and has no standard status (as of October 2026; the post was last modified on July 15, 2026). Cloudflare’s own conclusion was that most sites are not very agent ready yet. The checks can be run with isitagentready.com.

Starting with Chrome 150, Lighthouse includes an experimental “Agentic Browsing” category. It checks WebMCP tool registration, a subset of accessibility audits, layout stability (CLS) and whether an llms.txt file exists, and it reports the result as a fraction of passed checks, such as 2/3.

Which protocol covers what? Agent protocols as of 2026-10-08Map of protocols for AI agents in four layers, as of 8 October 2026. Agent to tools and data: MCP, by Anthropic, at the Agentic AI Foundation since December 2025, version 2026-07-28. Agent to agent: A2A, donated by Google to the Linux Foundation, specification 1.0; it includes the Agent Card, a JSON description at /.well-known/agent-card.json. Agent to website: WebMCP, a draft of the W3C Web Machine Learning Community Group and not a W3C standard, in which a web page provides tools in the browser; NLWeb by Microsoft, open source, where every instance is also an MCP server. Agent to merchant, checkout and payment: ACP by OpenAI and Stripe (beta) with a checkout API and Delegate Payment; UCP by Google with commerce partners (stable) with checkout and payment handlers, plus AP2 mandates as an optional extension; AP2 by Google (v0.2) with a checkout mandate and a payment mandate. Cross-links: ACP and UCP can also be used via MCP, UCP also via A2A; AP2 is an extension for A2A and MCP.PROTOCOL · PUBLISHER · STATUSAs of 2026-10-08Agent ↔ tools, dataMCPModel Context ProtocolAnthropic; at theAgentic AI Foundation(Linux Foundation) since2025-12-09Version 2026-07-28Based on specifications andpublisher sites, checkedon 2026-10-08.↳ = link to otherprotocols.Grey: draft or beta.Agent ↔ agentA2A · Agent2AgentGoogle, donated to the LinuxFoundation; joining the AgenticAI Foundation announced on2026-08-27Specification 1.0(v1.0.1 of 2026-05-28)Agent CardJSON description of an agent:skills, endpoint, auth/.well-known/agent-card.jsonPart of the A2A specificationAgent ↔ websiteWebMCPW3C Web Machine Learning CommunityGroup (editors: Microsoft, Google)Draft of 2026-10-02, not a W3Cstandard; Chrome origin trial from 149↳ the web page provides tools in thebrowser, much like an MCP serverNLWebMicrosoft (announced 2025-05-19),open source under the MIT licence,no published version number;uses the site’s Schema.org and RSS↳ every NLWeb instance is alsoan MCP serverAgent ↔ merchant: checkout, payment, linksACPbeta, 2026-04-17Agentic Commerce ProtocolOpenAI and StripeAgentic Checkout: create,update and complete sessionsDelegate Payment: paymenttokens passed via the agent↳ REST; MCP binding sinceversion 2026-04-17UCPstable, 2026-08-25Universal Commerce ProtocolGoogle with commercepartnersCheckout capability: create,update, completePayment handlers; AP2 mandatesas an optional extension↳ REST, MCP, A2A,Embedded ProtocolAP2v0.2 (2026-04-28)Agent Payments ProtocolGoogle; standardisation inFIDO working groupsCheckout mandate: recordsitems and purchase detailsPayment mandate: authorises apayment with an instrument↳ extension for A2A and MCP;usable in UCP
Map of agent protocols as of 2026-10-08: which connection each protocol covers, who publishes it and what its status is. In commerce, ACP, UCP and AP2 overlap, and several protocols build on MCP or A2A. Versions change quickly. Sources: Model Context Protocol: Specification 2026-07-28; A2A Protocol: Specification (Agent Card, relationship to MCP); W3C Web Machine Learning CG: WebMCP (Draft Community Group Report); NLWeb (GitHub, Microsoft); Agentic Commerce Protocol (GitHub, OpenAI and Stripe); Universal Commerce Protocol: Specification overview; AP2: Agent Payments Protocol.

Example

An online shop publishes a robots.txt with explicit rules for AI bots and an XML sitemap. It serves product pages as Markdown on request. At /.well-known/api-catalog it lists its APIs as RFC 9727 describes. That covers parts of discoverability, content and capabilities.

Checklist

  • robots.txt, sitemap.xml and Link headers (RFC 8288) are in place (the discoverability category of Cloudflare’s score).
  • Key pages can be retrieved as Markdown (Accept: text/markdown), optionally complemented by an llms.txt file.
  • Rules for AI bots are in robots.txt, and Content Signals are set if wanted.
  • Existing APIs are listed at /.well-known/api-catalog as RFC 9727 describes.
  • Buttons and links are real <button> and <a> elements, and labels are tied to their fields with for (web.dev, as of 2026-04-01).
  • The layout does not shift while loading, because agents that use screenshots would otherwise miss their targets (web.dev).
  • No transparent overlays cover clickable elements (web.dev).
  • Optionally, WebMCP tools are registered (a W3C Community Group draft).

Common mistakes

According to web.dev (as of 2026-04-01), these patterns make a site harder for agents to use:

  • Transparent overlays over clickable elements. Agents may discard covered nodes.
  • Divs and spans instead of <button> and <a>. Without a fitting role, agents recognize the elements less reliably.
  • Labels without a for attribute. The purpose of the field stays unclear.
  • A layout that keeps shifting. Agents that analyze screenshots lose track of elements.

FAQ

Is agent readiness a ranking factor? No evidence supports that. The score measures how usable a site is for agents.

How does it differ from the business meaning? In companies the term covers internal requirements such as infrastructure, logging and security when deploying agents. The website meaning covers what a site offers to agents from the outside.

More on the searchVIU blog

Sources

  1. Introducing the Agent Readiness Score – Cloudflare Blog (Apr 17, 2026)blog.cloudflare.com
  2. RFC 9727: api-catalog – IETFrfc-editor.org
  3. Is it agent ready? – Cloudflareisitagentready.com
  4. Optimizing for AI Agents — Common Pitfalls and How to Fix Them – searchVIU blog (2026)searchviu.com
  5. Lighthouse agentic browsing scoring – Chrome for Developersdeveloper.chrome.com
  6. Build agent-friendly websites – web.dev (updated 2026-04-01)web.dev

Last updated: