Ask your AI assistant to book a table for four tonight. It finds a restaurant, reads the menu, and confirms there's outdoor seating. Then it hits the reservation page and hands you back a link. You finish in ninety seconds what it couldn't do in five.
That gap between knowing and doing is the problem WebMCP exists to close. It's a proposed web standard that gives websites a structured way to tell AI agents what they can do there and not just what they can read. The agent stops guessing. It asks for the thing you wanted and acts.
It's tempting to blame the model. But there's a duller problem in front of the smart one. We are asking a machine to operate an internet built for human eyes and hands.
Which is why I want to talk about a pack of gum.
A pack of gum in Ohio
On June 26, 1974, a cashier at Marsh Supermarket in Troy, Ohio scanned a ten-pack of Wrigley's Juicy Fruit. Sixty-seven cents. First product ever checked out using the Universal Product Code.
Getting to that beep took twenty-five years. The first barcode patent was filed in 1949, and the grocery industry didn't agree on one universal code until 1973.
The breakthrough wasn't the pattern of black lines, which is the part everyone pictures. It was the agreement behind them. A manufacturer and a supermarket could finally give the same product the same name a machine could read, and once the register could look that name up, checkout and inventory and reordering and eventually the whole logistics chain could run off one shared identifier.
Adoption was slow. By the middle of 1975, roughly half the products in American supermarkets carried UPC symbols and exactly thirty stores could scan them. As late as 1984, only about eight percent of American grocery stores had checkout scanners.
Nobody was confused about the value. The problem was that collecting it meant rebuilding the physical world. Manufacturers changed packaging. Stores bought scanners and the computers behind them. That costs money, and more to the point, it costs years.
Once the machinery was in, the quiet little code became unavoidable. Scanning registers were nearly universal in grocery stores by the early 1990s, and today GS1 barcodes get scanned more than ten billion times a day.
The stripes never changed what the cereal box looked like to you. They changed what the store could do with it.
The stripes never changed what the cereal box looked like to you. They changed what the store could do with it.
This time, the scanner is already installed
Told that way, the barcode sounds like a warning about how slowly any of this will arrive. I read it the other direction.
The barcode needed new hardware in every store. WebMCP arrives through machinery the world already installed. Six billion people are online. Browsers update themselves in the background. Websites deploy new code overnight. And the audience is already assembled: hundreds of millions of people actively learning to hand work to an assistant.
Installing scanners took decades. The web can pick up a machine-readable layer with a browser update and a deploy.
That compresses the timeline. It also compresses how long businesses have to get ready, which is a less fun way of saying the same thing.
Instructions, not clues
When your assistant "uses" a website today, it reads the page the way you would. It works out the layout, guesses which button does what, and hopes nothing shifts before the task is finished.
If you have ever talked a parent through a form over the phone, you know how this goes. "Okay, there should be a blue button near the bottom. No, the other one. Did a new screen come up? Can you read the first line, so I know what page this is?" It mostly works, until it doesn't. It is also slow, fragile, and weirdly tiring, and nobody in the conversation is senseless.
Every popup, dropdown, and form change is another chance to misread the site. That's usually what happened when your assistant stalls and hands you back a link.
Now picture the website meeting the agent halfway.
Instead of making the AI inspect a page to work out what's possible, the business publishes a list of things you can do there. A restaurant might offer "check availability" and "book a table." A store might offer "search products," "start a return," or "track an order." The business writes that list itself, which means it's current and stops at whatever the business decided to offer.
Think of it as the agreement behind the actions. The same way the UPC was the agreement behind the lines. Instead of an agent guessing what a button means by reading a page built for humans, it calls a capability the business chose to expose.
Cloudflare has a preview that can put a WebMCP interface on a site without touching the site's own code. OpenAI is currently hosting a public ten-day hackathon on WebMCP-enabled websites. This is not a research paper waiting for someone to pick it up. It is moving.
The demand underneath it is already here. People are asking their assistants to finish tasks, and the assistants keep failing at the last step.
The machine audience is already here
Agents are not making most of the purchases on the internet yet. But websites are already serving an audience that isn’t entirely human.
Across Cloudflare’s network, automated requests outnumbered human requests to HTML pages for much of 2025. Identifiable AI bots averaged 4.2 percent of those requests, and Googlebot, which feeds both search and AI, another 4.5 percent. The number I keep coming back to is a different one.
21×
Surge in human-prompted AI web browsing in a single year
AI crawling triggered directly by a real person asking an assistant to inspect a website.
Most of that traffic is still crawlers and scrapers rather than the helpful personal agent we picture. But twenty-one-fold isn’t noise. Machines already read the web at enormous scale, and they’ve started acting on it for people.
What changes for you
Run the restaurant again. Booking a table: Your assistant asks the site for tonight’s availability and gets a real answer back: an open table at 7:30 on the patio. It checks with you: “7:30 on the patio, table for four, and I noted one vegetarian. Book it?” You say yes. The confirmation lands in your inbox and it’s real.
Starting a return: Or the return you’ve been putting off for two weeks. “Return the shoes I ordered last week.” Your assistant doesn’t need you to dig up an order number or find the returns portal, because the store publishes “start a return” as something an assistant can do. It starts one and tells you where to drop the box.
Rescheduling travel: Or you’re traveling, the meeting moved, and you text your assistant to push your flight to the evening. It comes back with two options and the fare difference, you pick one, and that’s the whole interaction.
The shift underneath all three is the same. Today your assistant is very good at telling you about businesses. This is where it starts transacting with them.
If you run a website
Picture an assistant choosing between two restaurants. One gives it a dependable way to check live availability and make a reservation. The other makes it inspect pages, dismiss popups, and guess its way through a form. Which one do you think it completes the booking with?
The same logic runs through products, appointments, returns, applications, and support.
If an agent can understand what your site does, it can use your site. If it can't, your site is a maze it will route around.
If an agent can understand what your site does, it can use your site. If it can't, your site is a maze it will route around.
Starting now doesn't mean betting the company on an unfinished standard. It means working out which journeys your customers will hand off, deciding what an assistant should and shouldn't be allowed to do on your behalf, and getting some practice expressing those actions clearly while the ecosystem is still soft enough to shape. That's a quarter of work, not a replatform.
And when agents do start looking for these capabilities, the worst way to find out is watching a competitor complete the task while your site hands the customer back a link.
What doesn't change
The web you see stays the web you see. Sites still look the same. You can browse them, compare things yourself, and buy the way you always have. Nothing gets taken away from the human internet.
That's the barcode again. The stripes went onto the package; they didn't replace it. WebMCP adds a layer for machines without touching the experience built for people. If you never use an AI assistant, there's a decent chance you never notice any of this exists.
Who’s in charge
There’s a fear here worth taking seriously, and it’s the one I get asked about most. An AI with a payment method, doing things while you aren’t looking.
What’s being built points the other way. Your assistant acts with your permission, as you, and only in the places you’ve connected it to. Anything involving real money or a real commitment gets a check-in first. “This is $84 total including fees. Book it?” Nothing irreversible happens without you saying yes.
Control 1 • The User
You control your assistant
You determine which permissions your assistant has, which services it connects to, and explicitly approve any financial or irreversible action.
Control 2 • The Business
The business controls its actions
The business decides what actions to expose through its published schema. An agent cannot invent or execute an action that was never placed on that list.
Those are two real controls, and neither one depends on the other.
How you'll notice
Infrastructure doesn't get a launch day. Nobody threw a parade for the barcode.
You'll notice this the way you noticed phones getting good cameras: gradually, and then it's just how things are. The signs are small and easy to miss. Tasks that finish instead of stalling. Fewer apologies from your assistant. Businesses eventually putting "works with your AI assistant" on a page the way they once put "find us on the App Store."
Underneath the small signs, though, a split is forming between the websites agents can reliably use and the websites they have to fight.
The best infrastructure is invisible
You never watched the barcode work either. It has saved you time in every checkout line of your life and you've almost certainly never thought about it once.
That's what good infrastructure does: it disappears into the result.
So you probably won't see the agentic web arrive. One day you'll just notice that the thing you asked for actually got done. And if you run a website, the question worth sitting with is whether it got done with you.
Optimizely is currently running a closed beta for WebMCP for Agent Ready Experiences.
Join the beta
