Your website is becoming an interface for agents, not just a page for people

Google does not appear to be saying “add an llms.txt file and win AI Search”.

Good. Because that would be stupid.

It would also be perfect bait for the next wave of AI SEO grifters. “Install this file and get cited by ChatGPT.” “New Google loophole.” “Rank in AI Mode overnight.” You can almost hear the webinar being written.

The more useful story is quieter.

Google has already published guidance saying its AI Search features still sit on familiar foundations: crawlable pages, useful content, technical clarity, quality systems, first-hand evidence, decent media, good product and local data where relevant. In other words, the AI layer has not magically removed the need for a proper website. Annoying, I know. Work remains undefeated.

That same guidance also warns against treating things like llms.txt, special AI markup, fake mentions or endless cloned question pages as ranking magic. So if someone is selling llms.txt as the new golden ticket for Google AI Search, keep hold of your wallet.

But here is where it gets interesting.

Chrome Lighthouse now has Agentic Browsing audits. One checks llms.txt. Others look at WebMCP-style metadata on forms and registered tools on a page. The docs describe llms.txt as an emerging convention for giving LLMs and AI agents a machine-readable summary of a site. Missing it is currently optional. A 404 is marked as not applicable, not a disaster.

That nuance matters.

Google is not saying, “Do this one little trick and rank in AI answers.” It is saying something more practical: agents are becoming a real class of web user, and your site may need to be legible to them.

That is a different job.

SEO is about being found, understood and trusted by search systems and humans. Agent-readiness is about whether software can actually use the site once it gets there.

Can it understand the contact form?

Can it tell which button books a demo and which one downloads a PDF?

Can it compare two service packages without inventing the missing bits?

Can it add the right product to a cart, check a policy, or stop and ask for approval before spending money?

Most websites are not ready for that. To be fair, plenty are barely ready for humans.

This matters because Google’s direction is fairly obvious now. AI Mode queries are getting longer. Google says users are searching with more voice and image input, asking more planning questions, and using AI Mode for decision-making. At I/O, Google announced Universal Cart, designed to follow shopping activity across Search, Gemini, YouTube and Gmail. It also talked about Gemini Spark, an agent that can take action under the user’s direction, with future plans around custom sub-agents and authorised payments with budgets and merchants.

Strip out the launch-event polish and the direction is simple: the web is moving from pages you browse to tasks you delegate.

That does not mean websites disappear. I have never understood that argument. AI systems need sources, pages, product data, policies, media, forms, feeds, APIs and transaction paths. The website does not vanish. It becomes infrastructure.

The question changes from “does our website look nice?” to “can this thing be understood, cited and operated?”

That is a much better audit.

Take a normal service business. A buyer asks an AI tool, “Who can help us implement AI safely across marketing and sales without turning the team into prompt jockeys?” The system goes looking for evidence. It needs clear positioning, named services, proof, examples, pricing signals if available, strong pages, useful articles and enough entity clarity to know what the business actually does.

That is answerability.

Then the buyer says, “Shortlist three, check whether they work with UK SMEs, and draft an enquiry.” Now the agent needs to operate. It has to find the contact route, understand the form fields, avoid weird JavaScript traps, include the right context and maybe attach a brief.

That is actionability.

Most companies have treated AI visibility as a content problem. Write more articles. Add FAQs. Chase mentions. Make the blog bigger. Some of that can help, if it is useful. A lot of it becomes sludge.

The better work is more joined-up.

Your pages need to say something specific. Your services need to be named clearly. Your proof needs to be inspection-proof. Your forms need to make sense. Your product data needs to be clean. Your policies need to be findable. Your site structure needs to stop hiding important information inside pretty but unreadable components. Your analytics need to show whether qualified people are actually moving, not just whether traffic went up.

And if agents are going to take actions, the controls matter. Budget caps. Confirmation steps. Approval gates. Logs. Clear language around what happens when a form is submitted or a payment is authorised.

That last bit is where a lot of AI people get carried away. They want the agent to do everything. Fine. But in the real world, “everything” includes spending money, submitting forms, changing data, booking calls and annoying customers. The boring controls are not admin. They are the product.

For business owners, the next step is not to panic-buy an AI SEO package.

Run a practical agent-readiness check instead.

Can a human understand the page in ten seconds?

Can Google crawl and index the important content?

Can an AI system answer sensible buyer questions from the page without making things up?

Can a browser agent identify the main action?

Are form fields labelled in a way software can understand?

Is the checkout or enquiry flow robust enough for delegated use?

Are the dangerous actions protected by confirmation, limits and logs?

That is the work. It is not glamorous. It is absolutely commercial.

This is also why the post-agency model keeps making more sense to me. The old agency habit was to sell output: pages, posts, campaigns, decks, assets. The new job is to build operating layers. Content still matters, but it sits inside a system with search, data, workflow, automation, measurement and governance.

A blog post can help you get found.

A clear service page can help you get shortlisted.

A well-structured form can help an agent send the right enquiry.

A decent CRM handoff can turn that enquiry into follow-up before it goes cold.

That is not “AI replacing marketing”. That is marketing becoming more operational.

The web is not going away. It is becoming more machine-readable, more task-oriented and less forgiving of vague brochureware.

If your website cannot explain what you do, prove it, and help software take the next step safely, it is going to feel older than it looks.

Pretty pages will not be enough. The agents need handles.

---