AI Agents Now Operate Your Website for Customers. Can They Finish the Job?

An artificial intelligence (AI) agent can now open a business's website on a customer's behalf and try to finish the job itself: find the product, fill in the form, book the table. At the end of August 2026 OpenAI switched on what it calls site tools inside the built-in browser of the ChatGPT desktop app, so that instead of relying only on clicks and typing, ChatGPT can use tools a website provides ¹, a launch Search Engine Journal reported on 27 August ². Those tools run on WebMCP, a proposed web standard ¹ that Google is testing in Chrome at the same time ³. Whether a website can be operated this way is no longer a question for its developer alone. When an agent cannot complete a task on a site, the customer who delegated it never sees a clumsy page. They see a task that failed, and the business has no report that tells it a sale or an inquiry was lost there.

OpenAI's documentation lists what a supporting website can let ChatGPT do: search documentation or compare information, explore a dashboard, build a travel itinerary, update a shopping cart ¹. An arrow appears in the address bar when a site offers tools, and the person can see which of them only read information and which make changes ¹. Then comes the sentence that matters commercially: each website decides which tools to provide ¹. For the first time, a business defines in advance what an agent may do on its site, instead of hoping the agent guesses correctly. OpenAI framed it for developers in exactly those terms: "Instead of leaving agents to guess their way through your [user interface], you define exactly how they can use your app" ².

The launch did not arrive alone. OpenAI announced a ten-day WebMCP developer challenge on 25 August with Google Chrome, Cloudflare, Shopify, Vercel, Render and Netlify as partners ⁴. Shopify had already moved. Since 5 August 2026, every Shopify store built on its standard theme system, Liquid, exposes WebMCP tools that let an agent search the catalog, manage the shopper's cart and proceed to checkout, with nothing for the merchant to install or configure ⁵. Shopify's changelog states the case against the alternative plainly: without WebMCP, agents have to read every page's code and simulate clicks, which it calls slow and error-prone ⁵. A small retailer on Shopify became agent-callable by default this summer. A business on a custom-built or page-builder site did not, and the gap between the two opened without either of them doing anything.

Without published tools, an agent falls back on what Google's Chrome documentation calls actuation: simulating mouse clicks and text input as though it were the human user ³. Agents already run errands on ordinary websites this way: auto browse in Gemini in Chrome handles tasks from appointment booking to finding items in stock ⁶. To click anything, the agent first has to read the page, and Google's web.dev guidance lists the three ways it does so: screenshots, raw HTML and the accessibility tree ⁷. Screenshots are the fallback, slow and expensive in the tokens an AI model consumes ⁷. The same guidance is blunt about where most websites stand. Sites designed to be beautiful for humans, with complex hover states, shifting layouts and fluid motion, are in its words "functionally broken for agents" ⁷. That describes the design brief of a great many business websites, and none of it was a mistake while the only visitor was a person.

Google has also started measuring it. Lighthouse, the auditing engine behind PageSpeed Insights and the Chrome DevTools reports, added an Agentic Browsing category to its default configuration on 7 May 2026, with the release notes expecting it to reach PageSpeed Insights within two weeks ⁸. It does not produce the familiar score out of 100. It reports how many agent-readiness checks a page passes, and Google gives the reason: the standards for the agentic web are still emerging, so the focus is to gather data and provide actionable signals rather than a definitive ranking ⁹. The checks cover whether WebMCP tools are registered, a subset of accessibility audits that are critical for machine interaction, layout stability measured as Cumulative Layout Shift (CLS), and the presence of an llms.txt file, a short machine-readable summary of the site ⁹. Half of that list is not new technology at all. Accessibility and layout stability were already expected of any competent build, and they now carry a second consequence.

The accessibility tree is the part that matters most and the part nobody sees. It is the browser's structured summary of a page, the roles, names and states of every interactive element, built for assistive technology such as screen readers ⁷. Google's Lighthouse documentation calls it the agents' primary data model ⁹, and states the consequence in one sentence: missing labels can block users with visual disabilities and agents from completing a task ¹⁰. The WebAIM Million, an annual accessibility audit of the top one million home pages, found in February 2026 that 51% of them had missing form input labels and 30.6% had empty buttons, and that one third of all form inputs detected, 33.1%, were not properly labeled ¹¹. Detected errors averaged 56.1 per home page, up 10.1% on the year before ¹¹. A contact form whose fields carry no programmatic label is a lead form a screen reader user cannot reliably complete, and now neither can the agent a customer sent to fill it in. Those are also the defects that carry legal exposure, covered in the three website laws most sites already break.

Layout stability is the second old requirement with a new consequence. Google's scoring notes state that layout shifts caused by ads, images without dimensions or injected content can move elements between the moment an agent identifies them and the moment it tries to interact ⁹. The web.dev guidance adds transparent overlays, which an agent's visual analysis may treat as covering the button underneath, and interactive elements too small to register at all ⁷. A person shrugs at a banner that pushes the page down and taps again. An agent that clicked where the button was a moment ago has clicked something else, on a customer's behalf.

WebMCP itself is lighter than it sounds on a well-built site and heavier on a poorly built one. Its declarative form adds annotations to standard HTML forms, and its imperative form registers tools in JavaScript ³. Google describes it as a progressive enhancement ³, meaning the site keeps working exactly as before for every visitor that does not use it. The same documentation names the limit: a highly complex site likely needs refactoring or additional JavaScript to handle application and interface state ³. A business whose booking form is a genuine HTML form is a set of annotations away from offering an agent a clean tool. A business whose booking flow is a chain of styled boxes that only look like buttons has a rebuild in front of it before any tool can sit on top.

None of this is settled, and it should be read that way. The WebMCP specification is a draft from the World Wide Web Consortium (W3C) Web Machine Learning Community Group and is not on the W3C Standards Track ². In Chrome it runs as an origin trial, a time-limited experiment, starting from Chrome 149 ³, and Google said in May that Gemini in Chrome would soon support it ⁶. ChatGPT's site tools work only when a person's account and chosen model support them ¹, and not in Enterprise or Edu workspaces ². Nor is this search optimization under a new name: OpenAI does not connect WebMCP with search rankings, citations or recommendations ². What agent readiness decides is narrower and closer to revenue: whether a task that starts on a business's website finishes there.

Inside most businesses, nobody owns this. The website was signed off when it looked right on the owner's phone, an accessibility review was never commissioned, and the analytics dashboard was built to count human sessions. An agent that abandons a booking halfway leaves no complaint and no support ticket. The loss accrues quietly, one delegated task at a time, while a competitor on a platform that switched agent tools on by default takes the booking without ever having planned to. The defects that cause it are already on the page, and the WebAIM figures say they are on most pages ¹¹.

So, plainly: should a business make its website agent-ready now? Yes for the foundations, not yet for the new standard. Give every button, link and form field a real name, and stop the layout from moving under the cursor, now, because every agent reads those signals whichever company built it, and Google's own guidance notes that everything that makes a site agent-ready also makes it better for humans ⁷. Treat WebMCP tools as the second step, worth building early only where revenue runs through a form or a cart that customers would plausibly hand to an agent: bookings, reservations, quote requests, reorders. The tie-break for everything in between: if a task on the site cannot be completed with a keyboard and a screen reader, assume an agent cannot complete it reliably either, and fix that before anything else.

Italian DesAIgns builds websites on those foundations from the first component: semantic HTML, full keyboard operability, screen-reader compatibility and the Web Content Accessibility Guidelines (WCAG) 2.1 AA, with Core Web Vitals, layout stability among them, prioritized in the build and accessibility treated as a structural requirement rather than a post-launch audit. The payment side of the same shift, the rails agents use to buy, is covered in AI agents now completing checkout directly, and the crawler side, where AI systems read a page without ever running it, in why AI crawlers can't read JavaScript. The insights archive follows WebMCP as it moves from trial toward a standard. For a first look at how legible a page is to AI crawlers, the free AI visibility check scores one page out of 100 on metadata, heading structure, structured data and crawler access, and returns up to three prioritized fixes. It reads the page the way a crawler receives it, not the way an agent operates it, and the second half of that question starts with how the site was built.