Short answer: AI assistants such as ChatGPT, Gemini, Claude and Perplexity now research, compare and shortlist on behalf of people, and some can start a purchase. An agent-ready website gives them a clean, machine-readable way to read your content and, for a shop, to check products and prices, using open standards such as MCP. Cloudflare’s Agent Readiness score measures how far along a site is. Most sites score low, so getting ahead costs little today. We treat it as one more sales channel next to search, ads and marketplaces, not a replacement for any of them.
This article explains what the score checks, what it does for a company website and for an online shop, what we expect over the next one to five years, and why we started building it into our projects.
What “agent-ready” means
An AI agent is an assistant that does a task for someone: “find three agencies in Europe that build headless WordPress sites”, or “find running shoes under €150 that ship to Sweden by Friday”. Most agents still read websites the hard way. They load the HTML, guess where the menu and the price are, and hope they picked the right number.
MCP (Model Context Protocol) is an open standard that lets a website or app offer tools an agent can call directly, such as “search products”, “get price and stock” or “request a quote”. The agent asks a question and gets a structured answer instead of scraping a page. Agent readiness is the wider set of signals around it: discovery files, Markdown versions of pages, rules for AI bots, MCP and, for shops, commerce protocols.
How the Agent Readiness score works
Cloudflare launched the score in April 2026 at isitagentready.com. You enter a URL and it runs a set of checks grouped into categories:
| Category | What it checks | Who needs it |
|---|---|---|
| Discoverability | robots.txt, sitemap, Link headers, DNS records for AI discovery | Every site |
| Content | Markdown versions of pages served on request | Every site |
| Bot access control | AI bot rules and Content Signals in robots.txt, Web Bot Auth | Every site |
| API, auth, MCP and skills | API catalog, MCP server card, Agent Skills, WebMCP, OAuth discovery | Sites where an agent can do something, not just read |
| Commerce | UCP, ACP, x402 and similar payment protocols | Online shops (shown separately) |
When Cloudflare scanned the 200,000 most visited domains, 78% had a robots.txt, but most were written for classic search crawlers. Only 4% declared how AI may use their content, 3.9% served Markdown to agents, and fewer than 15 sites in the whole set published an MCP server card or an API catalog. That gap is the opportunity.
Where things stand in October 2026
- AI traffic converts now. Adobe’s data shows AI-referred traffic to US retail sites up 393% year on year in Q1 2026. By March it converted 42% better than other traffic and brought 37% more revenue per visit. A year earlier it converted 38% worse. Adobe also found about a third of product pages could not be read properly by AI.
- Google is building checkout into AI Mode and Gemini. Its Universal Commerce Protocol (UCP), developed with Shopify, Etsy, Wayfair, Target and Walmart, launched in January 2026. It is open to US merchants through Merchant Center early access, with wider rollout planned.
- OpenAI stepped back from its own checkout. In March 2026 it deprioritised Instant Checkout in ChatGPT and moved the focus to product research and comparison, with merchants keeping their own checkout.
- WooCommerce ships native MCP support, still labelled a developer preview. It covers store management (products and orders), not shopper-facing agents.
- WebMCP, which lets a page register tools for an agent inside the browser, is still experimental in Chrome, and its API changed name during 2026.
The direction is clear, the details are not. Discovery and comparison already happen inside AI assistants. Checkout completed by an agent is early and uneven. That shapes what we recommend.
What it does for a company or service website
- Agents describe you correctly. Services, starting prices, locations and contact details come from your own pages instead of an outdated directory listing.
- You get onto shortlists. When someone asks an assistant for three options, the sites that are easy to read and cite are the ones that get named.
- Agents can act. An MCP tool or a WebMCP form can let an assistant request a quote or book a call with the details already filled in.
In practice this means an llms.txt file, Markdown versions of key pages, clear AI rules in robots.txt and an MCP server card that lists a few read-only tools. On a well-built site that is days of work, not weeks.
What it does for an online shop
- Product facts agents can trust. Price, stock, variants, shipping and returns in structured form. When the price on the page, in the feed and in the schema disagree, agents skip the shop.
- A place in AI shopping results. A clean Merchant Center feed plus a UCP profile is how a shop becomes eligible for Google’s AI Mode and Gemini checkout as it rolls out.
- Agent checkout, later. UCP and OpenAI’s Agentic Commerce Protocol let an agent hand a ready cart to your checkout, with payment staying under your control.
On WooCommerce, the native MCP support is for running the store. Endpoints for shoppers’ agents still need to be built. A headless WooCommerce shop makes this easier, because the product data already sits behind an API.
The medium and long term
| Horizon | What we expect | What to do now |
|---|---|---|
| Next 12 months | AI assistants become a research and shortlist channel. More referrals from ChatGPT, Gemini, Perplexity and Claude. Purchases still finish on your site. | Markdown and llms.txt, AI bot rules, consistent product data, AI referrals tracked separately in GA4 |
| 1 to 3 years | Agent checkout spreads beyond the US through UCP and similar protocols. Browsers ship WebMCP. Assistants prefer sites with tools over scraping. | An MCP server with read tools, a UCP profile for shops, quote and booking tools for service sites |
| 3 to 5 years | A share of enquiries and orders starts and ends inside an assistant. Your website becomes one of several front doors. | Maintain the agent interface like the mobile site: tested, versioned, kept in sync |
These are our expectations, not forecasts with numbers. The standards will change; the habit of asking an assistant first is unlikely to.
Why we do it
In our view this is one more sales channel. Not a replacement for Google or for your website, but another door customers use, the way shopping feeds and marketplaces were. Each of those rewarded the businesses that showed up early, while it was cheap and uncrowded. Today fewer than 4% of large sites even serve Markdown to agents.
It also helps what you already do. Clean structure, accurate data and fast pages are the same things classic SEO and AI search reward. And it is cheapest during a build: on headless WordPress the content already lives behind an API, so adding Markdown, MCP and agent endpoints is an extension, not a rebuild.
Our own numbers, from Cloudflare’s scanner today: pavels.dev scores 73 (level 5, “Agent-Native”). It passes Markdown, bot access rules, MCP server card, Agent Skills, API catalog and WebMCP. We skipped the OAuth checks on purpose, because the site has no private API for an agent to log in to. Our demo shop Pewes Energy scores 93, with a UCP profile and agent commerce endpoints. The lesson: treat the score as a checklist, not a target, and pass the checks that match what your site does.
It also shows up in Google. After we deployed the current version of pavels.dev, with these changes in place, impressions in Search Console’s Generative AI features report started growing visibly from the first days. The volume is still small, but the change lines up with the deploy and it clearly helped.
What we do not promise
- AI referrals are growing fast from a low base. For most small businesses they will be a minority of leads for a while.
- The standards are moving. Some checks will be renamed or dropped, so build it in a way that is cheap to update.
Where to start
- Decide which AI assistants may read your site, and say so in robots.txt with Content Signals.
- Serve a Markdown version of your key pages and publish an llms.txt.
- State services, starting prices and contact details plainly on the site.
- For shops: one source of truth for prices across the site, the feed and the schema, and an up-to-date Merchant Center account.
- Add an MCP server card with a few read-only tools: search, product or service details, request a quote.
- Track AI referrals separately in GA4 so you can see the channel grow.
- Shops selling in the US: apply for UCP early access through Merchant Center.
We started building this into our latest projects. It is not part of every build by default, because it takes extra time, but we add it to any WordPress or headless WordPress site we develop when the client asks for it. We only do it on sites we build ourselves. On a WordPress site held together by dozens of plugins, one for every small feature, the time goes into fighting the setup instead of into the agent layer. Tell us about your project and we will tell you which checks matter for your business and what it takes to pass them.
Frequently asked questions
A website that AI assistants can read and use reliably: it publishes clear rules for AI bots, Markdown versions of its pages and machine-readable tools, often through MCP, so an agent can get accurate answers or take an action without guessing from the HTML.
MCP (Model Context Protocol) is an open standard that lets a site or app offer tools an AI agent can call, such as searching products or requesting a quote, and get a structured answer back.
In our experience, yes. After we deployed these changes on pavels.dev, impressions in the Generative AI features report in Google Search Console started growing. Clean structure, Markdown pages and accurate data help both classic SEO and AI search.
The basics, yes: AI bot rules, Markdown pages, llms.txt and accurate service or product data are cheap and useful today. Agent checkout and full MCP tooling can follow as the standards settle.
Yes. Classic WordPress can cover the basics with careful setup. A headless WordPress or WooCommerce build makes MCP tools and agent endpoints easier, because the content and product data already sit behind an API.
