Making your website ready for AI agents: llms.txt, WebMCP and Markdown explained

Key takeaways

• AI agents such as agent mode in ChatGPT open websites on their own, read them and prepare inquiries, bookings or purchases. Before important steps, they check with the user.
• The foundation of a website for AI agents is a clear structure, good accessibility, a stable layout and complete schema markup. People and machines benefit equally.
• An llms.txt file and Markdown versions of key pages give agents fast, compact access to your content. With WebMCP, your website can even offer its functions to agents directly as tools.
• WebMCP has been in a trial phase in Chrome since June 2026. If you prepare your website now, you are getting an early start.

As of October 2026

A homeowner wants to have his heating system serviced before winter. Instead of searching himself, he hands the job to his AI assistant: find three specialist companies nearby, compare their services and prepare an appointment request with the best fit. To do this, the assistant opens the companies' websites, reads their services and service area, and fills in the inquiry form. Ideally, it finds everything on your site right away.

What AI agents do on a website

An AI agent is an assistant that does not just answer but completes multi-step tasks on its own. OpenAI introduced ChatGPT agent in July 2025. It works on its own virtual computer and uses two browsers: a visual one that operates pages the way a person does, and a text-based one that reads large amounts of text quickly. According to the OpenAI Help Center, the agent clicks buttons, fills out forms and navigates websites.

Google is building similar features. Auto browse in Gemini in Chrome handles multi-step tasks such as reservations or product comparisons, and the agent Gemini Spark automates entire workflows. Both are initially available in selected countries. What all of these offerings have in common: before submitting a form or making a purchase, the agent asks for confirmation. The machine does the groundwork, and the person makes the decision.

How an agent reads your site

Alongside what is visible on screen, the accessibility tree plays a central role for agents. It is a structured description of the page that the browser automatically generates from the HTML: headings, text, links, buttons and form fields with their labels. Screen readers for blind and visually impaired people work with exactly this description.

In the documentation for its auditing tool Lighthouse, Google writes that agents evaluate the accessibility tree to identify interactive elements. That is why many accessibility principles help agents as well:

• Clear headings in a logical order, so that the structure and topics of the page are easy to recognize
• Real buttons and links with descriptive text such as “Request an appointment”
• Labeled form fields, so that it is clear where the name, phone number or preferred date belongs
• Alternative text for images that describes what they show

This makes your website understandable for agents and easier to use for everyone.

Fast and stable: why the layout matters

Agents work through a page step by step, so fast loading times benefit them at every step. A calm layout is just as important. According to Google, visual stability allows agents to interact reliably with buttons and input fields. Fixed dimensions for images and content that stays in place after loading create this stability. In our practice, it has also proven effective to deliver important content directly in the HTML so that every tool can read it right away.

Schema markup: facts without room for interpretation

An agent tasked with preparing an appointment needs reliable facts: services, service area, opening hours, contact options and prices. Structured data based on schema.org, schema markup for short, provides exactly this information in machine-readable form. Google describes it as explicit clues about the meaning of a page and recommends the JSON-LD format.

What matters is that the markup is complete and matches the visible information. With Schema & Snippets, our tool GeoUp generates suitable Schema.org markup and citable text snippets based on your actual visibility gaps and adds them to your website with a single click. For clients of our AI visibility service, complete schema markup is standard.

llms.txt: the signpost for agents

The llms.txt is a small text file in the root directory of your website, available at your domain followed by /llms.txt. Jeremy Howard proposed it in September 2024, and since August 2026 the proposal on llmstxt.org has been available in a second version. The file is written in Markdown, a simple text formatting syntax. It contains the name of the website, a short summary and lists of links to the most important pages, each with a brief explanation. Agents read it as needed when they have to find something out for a user.

According to llmstxt.org, thousands of websites now publish an llms.txt. OpenAI, Anthropic and Google also provide one for their developer documentation. And in its new category for agents, Lighthouse checks whether an llms.txt is available.

The llms.txt is a community proposal and complements a well-structured website. The effort is small: your name, one sentence about what you offer, and links to your services, frequently asked questions, “About us” and contact pages. Our own llms.txt on njudev.com shows what this can look like.

Markdown versions: your content without the clutter

Besides the actual text, a web page contains a lot of extras: navigation, footer, scripts and design. AI models process text in tokens, small units of text, and the less clutter there is, the faster they get to the content. Network services provider Cloudflare demonstrated this with one of its own blog posts: 16,180 tokens as HTML, 3,150 tokens as Markdown, around 80% fewer.

That is why the llms.txt proposal recommends also offering important pages as a clean Markdown version, at the same address with the extension .md. This is especially worthwhile for services, pricing, frequently asked questions and contact. It is important that the Markdown version matches the content of the visible page exactly.

WebMCP: the website offers its functions directly

Today, agents operate websites by imitating clicks and inputs. Google describes this as a process with many steps, each of which the agent has to interpret on its own. WebMCP makes this simpler: the website tells the agent which tools it offers, for example “Request an appointment” with the fields name, service and preferred date. The agent calls the tool directly and passes along the relevant information.

Technically, there are two approaches: existing HTML forms receive additional attributes, or developers define tools in JavaScript. The specification is being developed in the W3C's Web Machine Learning Community Group, with developers from Microsoft and Google listed as editors. It is a draft and not yet an official W3C standard. Since June 2026, Chrome has let websites work with WebMCP in an origin trial, and in May 2026 Google announced that Gemini in Chrome will support WebMCP. If you get started now, you gain experience while the field is still young.

According to the specification, a website with WebMCP can be thought of as a server for the Model Context Protocol (MCP) whose tools run directly in the browser. MCP is an open standard that AI applications such as Claude or ChatGPT use to access data and tools from external systems. The GeoUp article on MCP and agent-based search explains why this direct data access matters. You can take a first step today with the public GeoUp brand server: it provides company data, services, unique selling points and opening hours to any AI assistant, accessible without signing in, as soon as your listing in the GeoUp business directory is published. For online stores, the Universal Commerce Protocol is being developed in parallel. You can read more about it in our article on shopping in AI Mode.

How to proceed step by step

1. Structure and accessibility: clear headings, labeled forms, descriptive links
2. Speed and a stable layout: fast pages, fixed image dimensions, important content directly in the HTML
3. Schema markup: complete and consistent with the visible information
4. llms.txt: a short signpost to your most important pages
5. Markdown versions: for the pages agents need most often
6. WebMCP: prepare inquiry and booking forms during the trial phase

The experimental Lighthouse category Agentic Browsing provides a technical overview. Among other things, it checks accessibility for agents, layout stability, the llms.txt and WebMCP tools. We build websites that work equally well for people and agents, and we implement these steps as part of our AI visibility service. You can book our SEO and GEO expert team at any time in addition to GeoUp. Simply write to us at our email address.

GeoUp in practice: your website is already being read today

The answers of ChatGPT, Gemini and other AI systems already show that AI systems retrieve websites themselves. In our ongoing GeoUp measurement series since July 2026, with 19 client websites, 119 measurement runs and 17,790 AI answers, we checked for every mention whether the AI linked to the company's website as a source. For a B2B software provider, this was the case for 158 of 164 mentions, or 96%. A website that machines can read easily is the best foundation for this.

With GeoUp, you can see what this looks like for your website. With every measurement, our tool asks ChatGPT, Claude, Gemini, Perplexity and Google AI Overviews 30 typical customer questions from your industry and analyzes whether your company is mentioned, in which position and which sources the AI cites. The measurement runs regularly, up to weekly, and you can add your own questions. This shows you whether your website appears more often as a source after the changes.

Frequently asked questions

What is a website for AI agents?
A website that AI agents can easily read and operate: with a clear structure, labeled forms, a stable layout and complete schema markup. An llms.txt, Markdown versions of important pages and WebMCP help as additions. The same foundations also make the site more pleasant for people to use.

Do I need an llms.txt?
It is a useful addition that takes little effort. The file shows agents in compact form where your most important information is. Update it whenever your services or pages change.

Is WebMCP ready to use?
WebMCP is in a trial phase in Chrome, and the specification is a draft. Websites can already try out the technology and prepare their inquiry and booking forms for it.

Do I have to rebuild my website for this?
In our experience, usually not. Headings, form labels, schema markup, llms.txt and Markdown versions can be added to an existing website. For a relaunch, we plan for the requirements of agents from the start.

Conclusion

AI agents read websites, compare providers and prepare inquiries. Much of what helps them is proven craftsmanship: a clear structure, accessibility, a stable layout and complete schema markup. An llms.txt, Markdown versions and WebMCP build on this foundation. If you start now, you are getting an early start, and regular measurements show you what is working.

How visible is your company in ChatGPT, Gemini and other AI systems?

With the free GeoUp quick check, you can see in moments, based on sample questions, whether ChatGPT, Perplexity, Claude, Gemini and Google AI Overviews mention your company. GeoUp is a product by njudev.

Would you rather have us prepare your website for AI agents? Then our team will take care of it. You can book our SEO and GEO expert team at any time in addition to GeoUp. Simply write to us at our email address or learn more about our AI visibility service.

Sources

1. OpenAI: “Introducing ChatGPT agent: bridging research and action”, July 17, 2025. https://openai.com/index/introducing-chatgpt-agent/
2. OpenAI Help Center: “ChatGPT agent”. https://help.openai.com/en/articles/11752874-chatgpt-agent
3. Google Chrome Help: “Ask Gemini in Chrome to complete tasks for you with auto browse”. https://support.google.com/chrome/answer/16821166?hl=en
4. Gemini Apps Help: “Use Gemini Spark to manage your tasks & workflows in Gemini Apps”. https://support.google.com/gemini/answer/16596215?hl=en
5. Chrome for Developers: “Accessibility for agents” (Lighthouse), as of May 5, 2026. https://developer.chrome.com/docs/lighthouse/agentic-browsing/accessibility-for-agents
6. Chrome for Developers: “Layout stability” (Lighthouse), as of May 5, 2026. https://developer.chrome.com/docs/lighthouse/agentic-browsing/layout-stability
7. Google Search Central: “Intro to How Structured Data Markup Works”, as of December 10, 2025. https://developers.google.com/search/docs/appearance/structured-data/intro-structured-data
8. schema.org: “LocalBusiness”. https://schema.org/LocalBusiness
9. GeoUp: “Schema & Snippets”. https://geoup.io/schema-snippets (in German)
10. Jeremy Howard: “The /llms.txt file, v2”, llmstxt.org, published September 3, 2024, updated August 10, 2026. https://llmstxt.org/
11. llms.txt files of the developer documentation from OpenAI, Anthropic and Google, retrieved October 12, 2026. https://developers.openai.com/llms.txt, https://platform.claude.com/llms.txt, https://ai.google.dev/gemini-api/docs/llms.txt
12. Chrome for Developers: “llms.txt” (Lighthouse), as of May 5, 2026. https://developer.chrome.com/docs/lighthouse/agentic-browsing/llms-txt
13. njudev: llms.txt, retrieved October 12, 2026. https://njudev.com/llms.txt
14. Cloudflare: “Introducing Markdown for Agents”, February 12, 2026. https://blog.cloudflare.com/markdown-for-agents/
15. Chrome for Developers: “WebMCP”, as of October 7, 2026. https://developer.chrome.com/docs/ai/webmcp
16. Web Machine Learning Community Group (W3C): “WebMCP”, Draft Community Group Report, October 9, 2026. https://webmachinelearning.github.io/webmcp/
17. Chrome for Developers: “Join the WebMCP origin trial”, June 9, 2026. https://developer.chrome.com/blog/ai-webmcp-origin-trial
18. Chrome for Developers: “15 updates from Google I/O 2026: Powering the agentic web with new capabilities, tools, and features in Chrome”, May 19, 2026. https://developer.chrome.com/blog/chrome-at-io26
19. Model Context Protocol: “What is the Model Context Protocol (MCP)?”. https://modelcontextprotocol.io/docs/getting-started/intro
20. GeoUp: “MCP und agentenbasierte Suche: Warum ein direkter Datenzugang zählt”, March 6, 2026. https://geoup.io/news/mcp-agentenbasierte-suche (in German)
21. GeoUp: “MCP-Server”. https://geoup.io/mcp (in German)
22. Google for Developers: “Universal Commerce Protocol: Frequently Asked Questions”, as of October 2, 2026. https://developers.google.com/merchant/ucp/faq
23. Chrome for Developers: “Lighthouse agentic browsing scoring”, as of May 5, 2026. https://developer.chrome.com/docs/lighthouse/agentic-browsing/scoring
24. GeoUp: “Wann zitiert die KI die eigene Website? Bei kleineren Anbietern in bis zu 96 Prozent der Nennungen”, September 10, 2026. https://geoup.io/news/eigene-website-als-ki-quelle (in German)