Agent-Ready: When AI Assistants Don't Just Read Your Website, but Use It

ChatGPT, Claude, Gemini & co. have changed research – we already wrote about how your shop gets recommended in AI answers. The next stage is in sight: AI systems no longer just read, they act – they compare providers, prefill forms and book calls. This is called the agentic web, and its standards are emerging right now: MCP, WebMCP, Agent Skills, Content Signals, DNS-AID.

We made our own website fully agent-ready – not because hordes of agents will sign contracts tomorrow, but because the research phase already lives with AI assistants and the foundations are cheap today. Here is what that means in practice – with real measurements.

What makes a website agent-ready

Stage 1: be efficiently readable. An AI agent pays for every page with context window. A typical HTML page of our site costs it ~16,900 tokens – as markdown it is 5,470. That is why we serve every page as markdown on request (Accept: text/markdown): 68–73% fewer tokens per fetch. Add an llms.txt as a curated entry point, open AI-crawler rules in robots.txt including permissive content signals (search=yes, ai-input=yes, ai-train=yes), and Link headers that show agents the way.

Stage 2: offer defined actions. Readability is table stakes; the real difference comes with actions. We implemented three building blocks:

• Agent Skills (/.well-known/agent-skills/): markdown instructions that tell an agent how to use our content and how to prepare an account check for its user.
• WebMCP in the browser: on every page we register tools for browser agents – agency info, services, case-study numbers and a form prefiller. The agent may fill in fields, but consent and submission stay with the human.
• Our own MCP server (mcp.socialcooks.com): the official entrance for AI systems. Four tools, including request_account_check – the only writing tool, and it requires the user’s explicit consent to the privacy policy before anything happens.

Stage 3: be discoverable. An endpoint is useless if nobody finds it. That is what the server card at /.well-known/mcp/server-card.json and DNS-AID records are for – DNS entries through which agents discover a domain’s MCP server before the first request flows.

The principle behind it: agents assist, humans decide

Every building block followed the same line: an agent never gets more power than the human gave it. Our contact form stays protected by bot detection – agents use the purpose-built, limited paths instead. The form prefiller never touches the consent checkbox. The MCP lead tool requires documented consent. Exactly this separation – machine-friendly AND controlled – is, in our view, the standard e-commerce brands should aim for.

What does this mean for your shop?

Honest framing first: organic LLM traffic is still below one percent of visits today, and autonomous purchases by agents are the exception. But the sequence is the same as in every platform shift: readability comes first (and already pays into AI recommendations today), actions come second. For a start, stage 1 is enough:

1. Crawlability: does your robots.txt or bot protection lock out AI crawlers? Then you do not exist in AI answers.
2. Structured data & feeds: clean product data is to agents what shelf placement is to a supermarket – how to make your product feed AI-proof.
3. llms.txt: a curated overview of which pages matter and which facts are correct.
4. Markdown for agents: on Cloudflare, a single toggle.

And if you want to know where your setup really stands – from crawlability to whether your Meta account scales profitably: that is exactly what our free Meta account check is for. You get the results not as a PDF email but in an honest 30-minute call.