The Internet Is Getting a Separate Entrance for AI
For most of ecommerce history, businesses built one digital entrance: the website.
Humans visited it. Search engines indexed it. Apps connected to it. Everyone ultimately worked around the same commercial environment.
AI agents are forcing a separation.
This week, Google expanded infrastructure that lets merchants explicitly participate in agentic checkout. Developers spent five days building agents capable of buying from merchants without any integration at all. Shopify changed how it measures bot traffic. Then Amazon blocked Meta's Muse from shopping on its marketplace, calling it an unauthorized AI agent.
These are not isolated developments.
Commerce is beginning to distinguish between being accessible on the internet and granting a machine permission to act inside a business.
The web was built around pages that almost anyone could visit. Agentic commerce is creating something different: controlled entrances designed specifically for machines.
Google Is Turning Agent Access Into Merchant Infrastructure
On September 16, Google expanded the Universal Commerce Protocol integration hub inside Merchant Center. Google says UCP-powered direct checkout already works across hundreds of thousands of brands and retailers. Merchants using the hub can now enable cart transfer to their own sites and test checkout flows more extensively, with analytics planned.
The important development is not another checkout feature. It is where control is being placed.
Instead of an AI agent independently navigating whatever a merchant happens to expose on its website, the merchant gets an explicit infrastructure layer for deciding how machine-led commerce should work.
The distinction becomes clearer with cart transfer. An AI experience can help build the purchase, but the merchant can still receive the shopper on its own site for completion. Machine participation does not have to mean surrendering the entire transaction.
Imagine a retailer that wants its products available inside AI shopping but wants loyalty redemption, delivery selection or certain purchases handled on its own website. A structured agent channel allows those boundaries to exist.
The website remains. A second commercial entrance is being built beside it.
Mental Model — Machine Entrance: A website is designed primarily for people to navigate. An agent interface exposes selected commercial functions directly to machines, under rules the merchant can define.
Being visible to an AI no longer has to mean giving that AI unrestricted access to the merchant experience.
Source: Google agentic commerce updates
Developers Are Proving Agents Can Avoid the Special Entrance
Between September 15 and 20, the Agentic Commerce Pioneers Shopping Agent Challenge asked developers to solve a very different problem: build an agent that could take a real shopping request through to a confirmed order on almost any merchant's existing checkout, with no merchant integration required. Demo Day followed on September 21.
That approach matters because it exposes a tension at the centre of agentic commerce.
Protocols such as UCP create a structured route into a merchant. Browser-based agents can attempt to use the route that already exists for humans.
Consider a small furniture retailer with no agent API, no AI partnership and no new commerce protocol. A sufficiently capable browser agent could still search the site, select a chair, fill the checkout form and place the order.
From the customer's perspective, that sounds useful.
From the merchant's perspective, something important has changed. Software is operating inside an environment that was designed for a person, without the business necessarily creating a separate machine relationship.
This creates two competing models for agentic commerce: invited agents using structured infrastructure, and autonomous agents using the existing web.
The eventual balance between them will determine how much control merchants retain over machine customers.
Mental Model — Front Door vs Side Door: Protocols create an entrance designed for agents. Browser automation lets agents attempt to use the same entrance humans already use.
A merchant does not necessarily need to integrate with an AI agent for that agent to attempt to become its customer.
Source: The Shopping Agent Challenge
Amazon Drew the Boundary
On September 21, Amazon blocked Meta's Muse from accessing its marketplace.
Users attempting to shop through Muse began receiving a message stating that continued access by an unauthorized AI agent violated Amazon's Conditions of Use. Amazon said third-party applications purchasing on behalf of customers should operate openly and respect a service provider's decision about participation.
This may prove more important than the individual companies involved.
A customer can authorize an agent to act for them. But that does not automatically mean the merchant has authorized the agent to operate inside its environment.
Those are two separate permissions.
Imagine giving a personal assistant permission to buy something for you. Your permission establishes what the assistant can do on your behalf. It does not automatically require every private business to accept whatever method the assistant chooses to use.
Agentic commerce is discovering the digital version of that distinction.
This creates a new strategic question for retailers: Which agents are customers, which are partners, and which are unauthorized automation?
Until now, much of the agentic commerce discussion has focused on whether consumers trust agents. Amazon's decision introduces the other side of the transaction more clearly.
Machines need acceptance from businesses too.
Mental Model — Two-Sided Permission: The user authorizes the agent to act. The merchant separately decides whether and how that agent may participate.
Customer permission does not automatically create merchant permission.
Source: Amazon blocks Meta's Muse
Shopify Is Starting to Separate Humans From Bots in the Numbers
On September 21, Shopify began rolling out a change to how sessions are measured in Shopify Analytics. Among the changes, identified bot sessions are now filtered from session-related reports by default. Shopify also provides a human-or-bot filter in supported reports.
At first glance, this looks like an analytics update. It points to a larger operational problem.
As software increasingly browses stores, searches products and interacts with commercial systems, traffic stops being synonymous with potential human demand.
A store could receive 100,000 visits, but some may come from shoppers, some from search crawlers, some from AI research tools and eventually some from agents preparing or executing purchases.
Those machine visits have different economic meaning.
An AI agent researching a product for a customer may be commercially valuable even if it never behaves like a conventional website visitor. Meanwhile, automated traffic unrelated to purchasing may distort the same metrics.
Businesses therefore need to identify machine activity before they can decide how to value it.
The analytics layer is beginning to acknowledge that humans and machines are no longer interchangeable participants in digital commerce.
Mental Model — Dual Traffic: Human traffic measures people interacting with commerce. Machine traffic measures software interacting with commerce. The two can create value differently and need different interpretation.
A website session can no longer be assumed to represent a person moving through a buying journey.
Source: Shopify session and bot measurement update
The System That Is Emerging
A new machine participation layer is forming between the open web and the commerce systems underneath it.
The old internet made a simple assumption: if a resource was reachable, software could generally request it. Ecommerce added authentication, payments and account controls, but the primary experience remained built around a human visitor.
Agents make that model unstable because they do more than retrieve information. They can interpret, decide, fill forms, use accounts and initiate transactions.
Businesses therefore need something between open access and complete exclusion.
Google's UCP infrastructure shows one version: merchants deliberately expose commerce capabilities to agents. Amazon's Muse block shows the opposite boundary: a platform can reject an agent that attempts to operate without permission. Browser agents demonstrate why the distinction matters: increasingly capable software may not wait for formal integrations. Shopify's analytics change shows the operational consequence: businesses must first recognise machines before they can decide what those machines should be allowed to do.
The emerging architecture looks less like one universal storefront and more like access-controlled commerce for machines.
For leaders, the practical question changes. “Can AI access our business?” is too broad.
Businesses will need to decide what machines can discover, what they can retrieve, which actions require approved channels, when an agent can transact, what gets measured and when access should be refused.
The website was built to attract visitors.
The next commerce layer will also need to decide which visitors are allowed to act.
Core Truth
AI agents are turning access from a technical fact into a commercial permission.
