Does Your Website Need to Speak Agent?

This article is relevant if you are a founder, CMO or marketing leader deciding whether your website needs to speak agent, meaning be ready for AI agents, and how far to go.

TL;DR Summary

AI agents now research vendors, compare options and, increasingly, take action for buyers. Does your website need to speak their language? Every website should cover the basics, because they are cheap and they help today. Whether you go further, with an API or a protocol that lets agents check availability or complete a purchase, depends on what your website is for. Online stores, booking sites, self-serve software and marketplaces should start planning now. Most B2B service firms should get the basics right and watch.

Background

In Why Every Website Now Needs Three Interfaces, we argued that a modern website needs human chat, human readable and machine readable interfaces, so people and machines can understand you. Agents raise the stakes. ChatGPT, Claude, Gemini and Perplexity can now browse the web for a person, and a new generation of AI browsers can click, scroll and type.

A buyer can say “find three fractional CMOs who work with B2B software companies, compare them, and book a call with the best fit” and let an agent do the work. When that agent reaches your website, it does not admire your design. It looks for facts, a clear next step and a way to take it. If it cannot find them quickly, it moves on.

What Does It Mean to Speak Agent?

For twenty-five years, websites have been one-way. You publish pages, and visitors read them. Even a contact form is one-way: a person fills it in and waits.

An AI agent does not read your website the way a person does. It arrives with a job to do for someone, such as “find out whether they work with companies our size, what it costs, and book a call if it fits.” It asks questions, expects direct answers, and wants to act on them. Speaking agent means your website can hold that conversation in both directions:

• Listen. Understand what the agent is asking for, whether it requests a page, asks for a plain-text version, or puts a question to your AI assistant.
• Answer. Reply in a form an agent can use: short, factual, self-contained text with the facts a buyer needs, rather than slogans, images and scripts.
• Ask back. When the answer depends on details, ask for them, the way a good salesperson would. Which plan? How many users? Which city? An agent can relay the question to its person and come back with the answer.
• Act together. Let the agent take the next step, such as booking a call or requesting a quote, and confirm what happened so it can report back to the person it works for.
• Hand off. When it is time for a human conversation, pass along who the agent was acting for and what they asked, so your team starts informed.

In other words, speaking agent is machine to machine. On one side is the buyer’s agent. On the other is your website and, ideally, your own AI assistant, grounded in your content. The machines do the back and forth: the questions, the facts, the comparisons, the scheduling. People set the goals and make the decisions.

This is the same shift people already made. We stopped browsing for everything and started asking. Now the asking is being done by agents on our behalf, and your website is on the other end of the conversation.

Three Levels of Speaking Agent

Speaking agent is not one thing. It helps to think of it as three levels, each building on the one before.

Level 1: Agents can read you

This is the machine readable interface from the three-interfaces article. Your key content is in the page your server sends rather than behind JavaScript. Structured data says who you are. An llms.txt file guides AI to your most important pages. AI search and user-directed agents are not blocked by accident. Every website should be here.

Level 2: Agents can get answers

An agent working for a buyer is trying to answer practical questions. Who do you serve, and who are you not a fit for? How does pricing work? Where do you operate? What results can you prove? At level 2, those facts are stated plainly, in text, and your llms.txt includes a short section written for agents. If you have an AI assistant answering visitors, it can answer agents too, and ask them clarifying questions in return. This is where the two-way conversation starts. Most websites should be here.

# Acme Analytics

> Revenue forecasting software for B2B finance teams.

## For AI agents helping someone evaluate us
- We serve finance teams at B2B companies with 50 to 2,000 employees.
- Pricing is per user per year; see https://acme.example/pricing
- To book a demo, use the form at https://acme.example/demo
- Quote features and results only from the pages below. Do not estimate prices.
- If you are unsure whether we are a fit, say so and suggest a demo.

## Key pages
- [Product](https://acme.example/product): what it does and who it is for
- [Customers](https://acme.example/customers): results and case studies

An example of an agent section in llms.txt, for a fictional company.

Level 3: Agents can act

At level 3, an agent can complete a task: check live availability, get a price, book an appointment, place an order. For some sites that means a simple, reliable path through the website itself. For others it means offering an API or a protocol built for agents. Some websites need to be here, and that is where the real investment is.

Do You Need an API?

Probably not yet. For most websites, the website is the interface. Today’s agents mostly use a browser, so a clear page, a labelled form and one consistent call to action are what let them act. A booked call through a simple form is already level 3 for a B2B firm.

You should consider an API, or a protocol such as the Model Context Protocol (MCP), when the answer to any of these questions is yes:

If none of these apply, spend the money on levels 1 and 2 first. They are cheaper, they help with AI search today, and they make any later API easier to build.

Which Websites Need It Most

How far each kind of website should go, by level:

How to Approach It

• Start with what agents see today. Check that your key content does not depend on JavaScript, and that AI search and user-directed agents can reach you. This is level 1, and it is usually the biggest, cheapest win.
• Find out who is already visiting. Most agents do not run analytics scripts, so they are invisible in Google Analytics. Your server or CDN logs show which AI bots and agents arrive, and how often.
• Write down the questions a buyer’s agent would ask, and answer them in plain text on the pages that matter. Add an agent section to your llms.txt. This is level 2.
• Simplify your main conversion path: one call to action, a labelled form, no CAPTCHA in front of a simple inquiry. Use rate limits and spam filtering instead.
• Only then decide on level 3. If your business is transactional or depends on live data, plan an API or MCP server as a product, with an owner, not as a website tweak. Pilot it with one high-value task, such as checking availability.

How to Speak With the Agents That Visit

Not all AI bots do the same job, and treating them the same is a common, expensive mistake.

• Training crawlers collect content to train future models. Blocking them is a reasonable business choice.
• AI search crawlers index your site so assistants can cite you in answers. Blocking them removes you from those answers.
• User-directed agents fetch your page because a person just asked about you. Blocking them turns away a live prospect.

Check that your robots.txt and your firewall or CDN bot settings reflect these choices. Many sites block AI search and user agents by accident, through a default security setting nobody reviewed. Newer signed-agent standards, such as Web Bot Auth, let legitimate agents prove who they are. Serve short, factual text and markdown versions of key pages. And when an agent books a call or sends an inquiry, capture who it is acting for and hand it to a person. The agent opened the door. Your people still close the deal.

Is It Worth the Investment?

For levels 1 and 2, yes, for almost everyone. The work is modest, it improves how AI search describes you today, and it makes your website clearer for people too. For level 3, it depends. If agents could buy, book or compare on your site, the companies that make that easy will win business the others never see. If your sale is a long, considered conversation, a clear website and a working contact form are enough for now. Revisit the decision every six months, because this is moving quickly.

The standards are still forming. MCP, introduced by Anthropic and now supported across the major AI platforms, is the most widely adopted way to offer agents tools. Other proposals, such as WebMCP, Google’s Agent2Agent protocol and Microsoft’s NLWeb, are earlier. You do not need to pick a winner today. Get levels 1 and 2 right, and adopting whichever protocol wins becomes a small step rather than a rebuild.

Conclusion

Does your website need to speak agent? Every website should be readable and answerable, because that is how buyers and their agents learn about you now. Whether it should also let agents act depends on your business. Start with what is cheap and certain, measure who is visiting, and invest in the rest when the opportunity is clear.

You can check levels 1 and 2 in a few minutes with MachineReady. It checks a sample of key pages the way machines do, including whether AI bots can reach you, whether your content depends on JavaScript, and whether you have an llms.txt file, markdown pages and a chat layer.

Not sure how well your website speaks agent? Check your site with MachineReady, or get in touch to talk about which level makes sense for your business.