Online shops have spent years making pages easier for people to navigate. A newer kind of visitor now needs attention: the AI agent trying to find a product, compare the options and prepare a basket on somebody's behalf.
Cloudflare and Commerce.com have published details of a WebMCP rollout across almost 100,000 storefront domains. Their work gives shopping agents a set of defined tools, such as searching the catalogue and adding a selected item to a basket, instead of making the agent inspect a page and guess which buttons to press. For UK retailers, the important point is not that autonomous shopping has suddenly arrived in finished form. It is that the technical route is becoming much clearer, and businesses can begin testing it without surrendering control of the customer journey.
Why the Current Web Is Awkward for Agents
A conventional ecommerce site communicates through layout. People understand that a product image links to a detail page, a dropdown selects a size and a button adds the item to a basket. An AI agent working through the browser often has to interpret the same visual clues, cope with cookie notices and distinguish a genuine action from decorative content.
That approach is brittle. A theme update can move an element or change its label. A variant selector can be misunderstood. Bot protection may block the agent even though the customer asked it to act. The agent can also take several slow steps to complete a task that the shop's own system already understands precisely.
Cloudflare's announcement describes WebMCP as an experimental proposed browser standard that exposes structured tools directly on a page. Instead of simulating a series of clicks, a compatible agent can discover an approved search_products or add_to_cart action and supply the required information in a predictable format.
That distinction matters. A website designed only for visual navigation leaves an agent to infer what is possible. A site with a small, explicit toolset tells the agent what it is allowed to do and how to request it.
What Commerce.com and Cloudflare Have Built
Commerce.com already uses Cloudflare across almost 100,000 domains. The two companies have added an edge-delivered bridge that makes Commerce.com's existing commerce tools available to agents operating inside a shopper's browser session. According to the primary source, the visit remains on the merchant's site and credentials are not handed to a third party.
The implementation also keeps the merchant's backend in charge of the commerce rules. The agent can search available products and update the live basket, while the shopper completes checkout. That is a sensible boundary for an early deployment because it allows useful assistance before the most sensitive action.
There are practical controls for a platform operating many shops on shared infrastructure. Commerce.com can enable the capability across its service while allowing an individual merchant to opt out. Custom paths can point the browser bridge towards the correct MCP endpoint. If the supporting server is unavailable, or the browser does not support WebMCP, the script exits and the normal storefront continues to work.
Those details make this more useful than a polished demonstration. They address the ordinary deployment problems that appear when one experimental feature meets thousands of different shops: consent, tenant-level control, existing infrastructure and a safe fallback.
The Practical Consequence for UK Retailers
Retailers do not need to rebuild their websites around agents today. WebMCP is still in an early preview, available through experimental Chrome settings, origin trials and selected development environments. It would be premature to treat it as a settled route used by most customers.
It is reasonable, however, to review whether the current product and stock systems could support an agent safely. Many shops have a good-looking front end sitting on top of inconsistent catalogue data, unclear variant names or stock information split across systems. Giving an agent structured access will expose those weaknesses quickly. If the business cannot reliably answer which size is available, which delivery rule applies or which product identifier belongs in the basket, a new browser interface will not fix the underlying data.
The best first trial is narrow. Product discovery is a more useful starting point than an open-ended instruction to operate the whole site. A retailer could expose catalogue search, product details and basket preparation for one category, then compare the agent's result with the website and backend records. The test should include out-of-stock products, similar names, regional delivery restrictions, discontinued lines and unusual variants rather than only the cleanest examples.
Control needs to be designed into the tool definitions. Each action should have a clear purpose, restricted inputs and an expected result. The system should reject an invalid identifier rather than guess, return current information from the source system and record enough detail to diagnose a failed request. An agent should not acquire broader account access simply because one useful shopping action needs to read the catalogue.
Checkout deserves a separate decision. Keeping payment and final confirmation with the customer reduces the consequence of a mistaken selection and preserves a clear moment of consent. Businesses can still remove friction by preparing an accurate basket and handing the customer a normal checkout page. Full purchasing authority can wait until identity, returns, fraud, accessibility and customer-support implications have been tested properly.
Prepare the Operation, Not Just the Interface
Agent-ready commerce will affect more than web development. Customer service needs to recognise orders or baskets initiated with assistance. Analytics teams need to distinguish agent activity from malicious automation without treating every automated request as a sale. Merchandising teams need product descriptions and attributes that are accurate enough for structured search, not just persuasive enough on a page.
There is also a measurement problem. A successful test is not merely an agent managing to call a tool. The result must be correct, the normal site must remain reliable, and staff must be able to explain what happened when something fails. Useful measures include product-match accuracy, basket errors, fallbacks to human navigation and the number of cases needing customer-service correction.
Cloudflare and Commerce.com's work is a credible signal because it combines an emerging standard with a large, multi-merchant deployment. It does not prove that every UK retailer needs WebMCP this quarter. It does show that AI-mediated shopping is moving from browser improvisation towards controlled software interfaces.
The businesses best placed to use that shift will be those with clean product data, dependable integrations and firm boundaries around what software may do. Start by making one useful action reliable. Keep the customer in control of the final purchase. That creates a practical learning path without betting the storefront on an experimental standard.
