For roughly twenty years, a small business website has had two audiences. People, and Google’s crawler. You wrote for the first, formatted for the second, and most of the craft of the web lived in the gap between them.
In late August, OpenAI quietly added a third audience. It does not read your page.
The short version: ChatGPT’s built-in browser now supports site tools, OpenAI’s implementation of a proposed web standard called WebMCP. Instead of an AI agent squinting at your layout and simulating clicks, your website declares a short, explicit list of actions it will accept, things like check availability, search the catalogue, add to cart, and the agent calls those directly. The meaningful shift for AI agents for small business is not speed. It is control: you decide what an agent is permitted to do on your site, rather than hoping it guesses your buttons correctly. What it will not do, by OpenAI’s own account, is help you rank, get cited, or get found.
What are ChatGPT site tools, and what actually changed?
OpenAI announced site tools in its developer community on August 25, and the feature was widely covered a couple of days later. The mechanic is simple enough to explain at a counter.
Today, when an AI agent wants to do something on a website, it behaves like a very fast, very literal customer. It looks at the rendered page, finds what it believes is the booking button, and clicks. If you moved the button last Tuesday, it may fail silently. We wrote about that model two weeks ago, when AI tools for small business learned to click. Clicking works, but it is guesswork dressed as automation.
WebMCP inverts it. The page publishes its own capabilities as typed functions with defined inputs and outputs, and the agent calls one of them. Nobody is interpreting pixels. The human interface stays exactly as it is, and the person and the agent see the same page update.
The specifics are worth knowing before anyone gets excited. Site tools currently work in the ChatGPT desktop app’s built-in browser, with GPT-5.6 Sol and Terra; on Luna the feature is switched off, and it is not available in Enterprise or Edu workspaces at all. Tools belong to the page that offers them, so navigating away can make them disappear. OpenAI’s browser does not yet support the declarative HTML version of the standard or tools registered inside iframes, which means a site has to register them with JavaScript in the top level page.
WebMCP itself is not OpenAI’s. It is an experimental proposal in the W3C Web Machine Learning Community Group, developed jointly by the Chrome and Edge teams, with an early preview shipped behind a flag in Chrome 146 back in February. It is not yet on the W3C Standards Track. That is a real caveat, not a footnote.
Why does this matter for AI agents for small business?
Because the failure mode it fixes is one that costs small businesses money in a very specific place: the moment somebody decides to buy.
Adoption so far has clustered exactly there. Reporting from VKTR notes that millions of Shopify storefronts are already WebMCP enabled, with agents able to explore catalogues and build carts, and that Expedia, Instacart and Target have been experimenting since Google surfaced the idea in May. Progress Software shipped support in its Telerik and Kendo UI toolkits earlier in August. Notice what those have in common. Not one of them is a small business that hand coded anything. They are platforms and component libraries, which is how this will actually reach the average ten person company.
That is the genuinely non-obvious part, and it cuts against most of the advice circulating this week. The instruction “add WebMCP to your website” is close to meaningless for an owner-operator who has a Squarespace site and a booking widget. The realistic path is that your platform ships it, the way your platform shipped mobile responsiveness and structured data, and one day a toggle appears in your settings. The useful question is not “how do I implement this,” it is “when is my platform shipping it, and what will I be asked to expose when it does.”
Does WebMCP help you rank or get cited in AI answers?
No, and this is the claim most likely to be sold to you incorrectly over the next six months.
Search Engine Journal is explicit that WebMCP does not affect search rankings, citations, or discoverability. Site tools only come into play once an agent is already on your page, in a browser, on behalf of somebody who chose to go there. It is a conversion mechanism, not an acquisition one.
Those are two different problems with two different fixes, and conflating them is how businesses end up paying for the wrong thing. Getting found by an answer engine is still about consistent, machine-readable facts about your business, which we covered in AI search for small business. WebMCP is what happens after that works.
What should an owner actually do this month?
Four things, in ascending order of effort.
Ask your platform. One email to your website provider, your booking system, and your e-commerce platform: do you support WebMCP or agent tools, and when. The answers will tell you more about those vendors’ roadmaps than their marketing pages will.
Write down what you would expose. This costs nothing and it is the part that outlives the standard. If an agent could call three functions on your site, which three? Check availability is safe. Book the slot is a commitment. Apply a discount code is a decision you want a human near. That list is your policy, and you will need it whether the mechanism ends up being WebMCP or something that replaces it.
Read the security note honestly. OpenAI reviews each tool invocation before it runs and requires confirmation for higher stakes actions like purchases and permission changes. Its own documentation is refreshingly blunt about the limit: “These checks reduce risk; they don’t make a website or its output trustworthy.” Data exfiltration and prompt injection remain live concerns. An agent operating a signed-in session is holding your customer’s account, and that deserves the same seriousness you would give a new employee with till access.
Fix the thing WebMCP cannot. A declared tool that returns wrong availability is worse than no tool. The prerequisite is accurate underlying data, which is unglamorous and entirely within your control today.
The take
There is a reading of this story where the website becomes a vending machine and the people behind it fade out. We think that reading is backwards, and the design of the standard says so.
Scraping and clicking gave the agent free rein over an interface you never designed for it. WebMCP requires somebody to sit down and decide, deliberately, which actions a machine may take on the business’s behalf. That is judgment moving upstream, not disappearing. The owner who writes that list is doing the most valuable thing in the whole workflow, and no model does it for them.
The work it captures is mostly work that was never getting done anyway: the 11pm availability check, the customer comparing three suppliers at once, the booking that died because somebody could not find the button on a phone. Meanwhile the conversations that actually need a person, the awkward quote, the complaint, the repeat client who wants to talk, are exactly the ones no agent is calling a function for.
Worth keeping in perspective: this is an experimental proposal, running in one browser, on two models, unavailable to enterprise accounts. It may not survive in this form. OpenAI is running a WebMCP Challenge with submissions closing on September 3, which is what a company does when it is trying to find out whether an idea has legs, not when it has already won. Treat it as a strong signal about where the web is heading and a weak signal about what to buy this quarter.
The direction, though, looks durable. If agents keep growing as a way people transact, every small business site eventually needs an answer to a question it has never been asked before: what, exactly, are you willing to let a machine do here? The businesses that have thought about it before their platform ships the toggle will be the ones who use it well. For the tools that are ready today rather than experimental, our roundup of AI tools for small business by the job you need done is the more practical starting point.
Frequently Asked Questions
Do I need to change my website right now?
No. Site tools only work in ChatGPT’s built-in desktop browser, on two specific models, and the underlying standard is still an experimental proposal that is not on the W3C Standards Track. For almost every small business, the sensible move is to ask your website or booking platform whether WebMCP support is on their roadmap, and wait for them to ship it rather than commissioning custom development.
Will adding WebMCP help me show up in ChatGPT or Google AI answers?
No. WebMCP has no effect on search rankings, citations, or discoverability. It only applies once an AI agent is already on your page in a browser. Being found by answer engines is a separate problem solved by consistent, accurate, machine-readable information about your business across the places it is listed.
Is it safe to let an AI agent take actions on my site?
It depends on which actions you expose. OpenAI runs a safety review before each tool call and requires user confirmation for higher stakes actions such as purchases and permission changes, but its own documentation states plainly that these checks reduce risk without making a site or its output trustworthy. Prompt injection and data exfiltration are real concerns, so start with read-only actions like checking availability before exposing anything that commits money or changes an account.
How is this different from the MCP servers I keep reading about?
A regular MCP server is a separate service that an AI assistant connects to, usually to read data from a platform you have an account with. WebMCP works inside the browser, on the live page, in whatever session the person is already signed into, so the tools come from the website itself rather than from a server you connect separately. We covered the other kind in our piece on which ads MCP servers can spend your money.
Here is what we keep turning over: if you had to publish a list tomorrow of the three things an AI agent is allowed to do on your website, what would make the list, and more interestingly, what would you refuse to put on it?
