The Death of the Web Scraper: How High-Street Shops Will Talk to AI Agents

Key takeaways

 

• Today an AI agent "uses" your site by screenshotting it and guessing where to click — slow and error-prone
• WebMCP lets a website simply tell the agent what it can do and how — no scraping, no guessing
• This is real: WebMCP is a W3C standard (Google + Microsoft), already in Chrome 146 Canary, built on the Model Context Protocol
• For a small business: declare your capabilities in machine-readable form, or be the shop the agent can't use

 

The Mundane: the allergy notice taped to the counter

A family bakery in Nottingham. On the glass counter, under a smear of flour, a hand-written card: "Please ask staff about allergens." It works fine for the person standing there. But a customer across town has just asked their phone's AI assistant to "find a nut-free birthday cake I can collect Saturday," and the assistant is now trying to read the bakery's website — taking a screenshot, sending it to a vision model, guessing which button orders a cake, failing on the allergen question because the answer is a card on a counter forty miles away. The bakery is real, excellent, and effectively invisible to the machine doing the buying.

The Machine: stop being scraped, start declaring

For years, software that wanted to use a website had to imitate a human badly — scrape the page, parse messy HTML, screenshot it, guess where to click. It's slow, brittle, and breaks the moment you redesign your homepage.

WebMCP ends that. It's a new web standard — a W3C Community Group specification co-authored by engineers at Google and Microsoft, already available experimentally in Chrome 146 Canary — that lets a website expose its features as clean, callable tools directly to an AI agent through a browser API (navigator.modelContext). Instead of an agent guessing, the site simply says: "Here are the things I can do, here are the parameters I need, here's how to use them." It builds on the Model Context Protocol (MCP), the open standard that lets AI systems plug into external tools and data. And crucially, the tools run inside the user's own logged-in browser session — so an agent acting for a signed-in customer can act safely, without a tangle of new logins. In plain terms: the website stops being a picture the machine squints at, and becomes a menu the machine can read.

The Digital Eye: a shop as a set of tools, not a poster

To a scraper, the bakery is a wall of pixels to decode. To a WebMCP-enabled agent it's a tidy list: check allergens, get collection slots, order cake. The agent doesn't admire the homepage; it reads the capabilities, confirms the cake is nut-free, books the Saturday slot, and moves on. The business that declared what it can do gets the order. The one that left it on a card under the flour doesn't.

The Mindset: make your value explicit

For your whole life as a business you've communicated value to humans — a friendly sign, a good reputation, a chat at the counter. In an automated economy you also have to communicate it to machines, explicitly and in structured form: what you do, what you need to do it, what you will and won't take on. The businesses that thrive won't necessarily be the best ones — they'll be the best ones that also made themselves legible. A brilliant service the agent can't parse loses to a mediocre one it can. Declaring your capabilities clearly isn't a technical chore; it's the modern version of putting your best goods in the window.

Try this, this week

Open your own website on your phone and pretend you're an AI with no eyes — you can only read text, not see layout. Can you find your prices, your service area, your booking method, and your key yes/no facts (do you do X? do you serve Y?) as plain, explicit text? Every answer that's trapped in an image, a PDF, or "ask us" is a tool the machine can't call. Write down the three biggest gaps.

Common questions

What is WebMCP?

A W3C web standard that lets websites expose their features as structured tools an AI agent can call directly (via the browser's navigator.modelContext API), instead of the agent scraping or guessing. It builds on the Model Context Protocol.

Is it actually real, or future talk?

Real and shipping early: co-authored by Google and Microsoft, accepted by a W3C Community Group in 2025, and available experimentally in Chrome 146 Canary as of early 2026.

What does a small business need to do?

Start by making your core facts and actions explicit and machine-readable; structured data and clear, declared capabilities are what let agents — and the customers behind them — actually use you.