For most of the web’s history, the machines reading your website were search crawlers. They indexed your text and followed your links. That is changing. AI agents are starting to act on websites, not just read them: completing a booking, filling in a quote request, starting a trial. And Google has started measuring whether your site is ready for them.
There is a clean way to hold the two halves of this apart. Google Search says you do not need special files to be found. Google Chrome now audits whether agents can act once they have arrived. Discovery and functionality: the same company, two different jobs.
This page is about the second job. It explains what “agent ready” actually means, what Google now checks, who needs it now and who does not, and what is worth doing today regardless. The honest headline first: most sites do not need most of it yet.
Four questions an AI asks about your business
Before the detail, it helps to see where this sits. An AI moves through four questions about you, in order, and each one is a gate to the next.
Most SEO and AI-visibility work lives in the first three questions. Being chosen, the third, is the subject of our citation magnet work. This page is about the fourth and newest one: can an AI agent actually do something on your site. That four-question view is our AI Discovery Stack in plain English; the full model separates choosing and recommending into two steps.
What “agent ready” actually means
Agent ready means an AI agent can complete a real task on your site, a booking, a quote request, a trial sign-up, a purchase, and not just read your pages. Today, agents mostly read. Increasingly, they act. The clearest way to see why this matters is a before and after.
| Today | Soon |
|---|---|
| A customer asks ChatGPT for a supplier. | A customer asks ChatGPT for a supplier. |
| ChatGPT sends them to your website. | ChatGPT completes your enquiry form for them. |
| The customer fills in your enquiry form themselves. | The customer just reviews and confirms. |
That last step is the shift. When the agent does the form-filling, the site that is easy for an agent to operate has an advantage the site that is not does not. WebMCP is one way a site declares its forms and actions so an agent can complete them directly, rather than guessing from the pixels on the screen. It is a mechanism, not the point. The point is the task getting done.
Google has started measuring this
This is not us chasing a word. In May 2026 Google added an Agentic Browsing category to Lighthouse, the free audit tool built into Chrome that web developers already run every day. For the first time, Google is putting a readiness signal against whether AI agents can use your site. It is worth saying plainly that Google calls the category experimental and deliberately does not give it a 0 to 100 score, because the standards are still settling. The signal that Google built this at all matters more than the exact checks.
What the audit actually checks
- llms.txt: a short machine-readable summary at your domain root. A two-minute job.
- The accessibility tree: whether a machine can read your buttons, labels and structure. This helps screen readers and agents alike.
- Layout stability: whether elements move as the page loads. Agents click the wrong thing when they do.
- WebMCP: whether your forms and tools declare what they do, so an agent can call them directly. Three checks cover registered tools, form coverage and schema validity.
Where llms.txt fits, honestly
One small file attracts more noise than it deserves, so here is the plain position. llms.txt is useful. It is not a ranking factor. It is not an AI visibility strategy. It is a small capability enhancement. Google Search states you do not need it to be found, and that maintaining one will neither help nor harm your rankings. Google Chrome checks it so an agent can orient quickly once it arrives. Both are true at once, because they answer different questions.
Who needs this now, and who does not
Two questions decide it. Does your site have a transactional surface an agent could complete, a cart, a booking, a quote, a sign-up, a search? And is your audience plausibly arriving through an AI agent yet? Most businesses answer no to at least one, and that is fine.
| Your site | WebMCP now? | Worth doing today |
|---|---|---|
| SaaS, e-commerce or booking platforms (transactional, agent-likely) | A candidate now, on the key flow only | llms.txt, accessibility, layout stability |
| B2B software or quote-form services (transactional, low agent traffic) | Prepare; add at the next rebuild of that flow | llms.txt, accessibility, layout stability |
| Law firms, local services, content sites (no agent-completable transaction) | Not yet; the WebMCP checks correctly read Not Applicable | llms.txt, accessibility, layout stability |
| Any site not yet found, understood or chosen | No; those questions come first | Fix those first |
The pattern is clear. Accessibility, layout stability and a clean, honest llms.txt are worth doing now for almost everyone, because they help humans and agents at the same time. WebMCP is the one piece most sites can wait on.
What we would actually do, and in what order
- Ship a real llms.txt, written rather than auto-generated. Two minutes, and it passes the audit.
- Fix accessibility-tree gaps: unlabelled buttons, missing form labels, alt text. A triple-purpose fix for humans, screen readers and agents.
- Audit and fix layout shift on key pages. This is already a Core Web Vitals job.
- Park WebMCP until you next rebuild a transactional flow, then add it to that flow only.
What happens if you ignore this?
The honest answer is: probably nothing, for now. Most businesses do not need WebMCP today. Most should still fix accessibility and layout stability, because that helps real people right now, agents aside. Transactional businesses should watch the space and be ready to move when their customers do. The risk is not today. The risk is waking up in three years having ignored a shift that turned out to matter, while quieter competitors got ready. That is the balance worth holding: not “everything changes tomorrow”, and not “none of this matters”.
What we will not sell you
Two things are already being sold around this that we will not. The first is a cheap “pass the audit” product for a small fee, for a file you can write yourself in two minutes. The second is a five-figure “llms.txt implementation”, which is expensive work on a file Google Search has said it ignores. Our position is simpler: llms.txt is a small honest fix you should not overpay for, and WebMCP is real but early and can wait until it earns its place in a rebuild.
We pass our own audit
This site scores full marks in Lighthouse and passes the applicable agentic checks, with the WebMCP checks correctly showing Not Applicable, because a content page has no agent-callable tools and we do not fake the parts that do not apply to us. We run the audit we advise on, and we read it the same way for ourselves as for you.
Where this sits, and what to do next
Being chosen and being usable are two different stages. Our citation magnet work is about the third question, will an AI choose you. This page is about the fourth, can an agent then act. Citation magnet ends where agentic readiness begins.
If you want to know whether any of this matters for your site, including whether the honest answer is not yet, we will run Google’s audit and tell you straight. For the detail underneath, see the guides on WebMCP, agent-actionable forms and the agentic browsing audit.
