Can AI Agents Actually Buy From Your Website?

It is increasingly possible to ask an AI assistant to find a product, compare alternatives and, in certain supported experiences, complete a purchase. That does not mean every website with a shopping cart is ready for an AI agent to buy from it.

The distinction is operational. A recommendation requires useful information. A transaction requires valid inventory, pricing, customer choices, payment authority and a reliable record of the order.

The central question for merchants is not simply whether an agent can click a button. It is whether an agent can help a customer buy under the same commercial rules the business already operates.

Discovery and purchasing are separate capabilities

An assistant may discover a product through a search result, a merchant feed, a catalog service or a public product page. It may then compare ingredients, materials, sizes, availability or price.

At that point, a customer still needs an offer that can be purchased. A variant may be out of stock. The quoted price may depend on country or currency. A discount may be available only to eligible customers. Shipping may change the total.

These decisions belong to the merchant's commerce systems. An agent should not infer them from old text or invent an exception to a published policy.

OpenAI's shopping documentation describes product discovery and purchase links, with checkout inside ChatGPT available in certain eligible experiences. Eligibility and product coverage vary. Visibility in an assistant does not guarantee access to that checkout experience.

What a legitimate transaction path looks like

There are several practical ways an agent assisted purchase can reach completion.

A customer can be referred to the merchant's website, where the existing checkout handles the order. An agent can prepare a cart and pass the customer to a checkout with the relevant items already selected. In compatible systems, an agent may use documented interfaces to create or update a checkout session, subject to authorization and confirmation requirements.

These models are not interchangeable. A handoff is not an autonomous purchase. A checkout session is not necessarily a completed payment. And browsing a site as an automated user is not the same thing as a supported commerce integration.

The Universal Commerce Protocol, introduced by Google with commerce partners, is one attempt to give agents and merchant systems common transactional interfaces. Shopify's agent checkout documentation describes flows that can return a buyer to the merchant's checkout or support other eligible completion methods. In its implementation, the merchant remains the merchant of record.

This last point is fundamental. Participating in a new interface should not require a business to surrender its control of pricing, fulfillment, fraud decisions or contractual obligations.

The merchant still owns the hard decisions

A typical checkout resolves more than the item itself. It establishes the correct variant and quantity, applies applicable taxes and shipping, checks restrictions, calculates the final amount, and obtains customer authorization.

For a beauty brand, this may involve product bundles, loyalty pricing and region specific shipping limits. For a fashion retailer, it could involve size availability and returns. Some transactions may need identity checks or human confirmation.

A credible agent integration must respect those conditions instead of constructing its own version of the checkout. It must also handle errors. If a product sells out, the customer needs an accurate explanation and an alternative path, not a false order confirmation.

How to tell whether your store can support the journey

Begin with a specific purchase request rather than a protocol checklist. For example: “Find the 250 mL version of this product, confirm it is available in New York, and help me buy one.”

Test whether the product can be identified unambiguously. Confirm that the price and availability match the commerce backend. Follow the action route and observe whether customer decisions and product selections survive. Finally, verify that the transaction can be completed only through the authorized payment and order process.

Where agent interfaces are available, test with them directly. Where they are not, do not pretend that a successful manual checkout demonstrates agent purchasing support.

The larger business opportunity

The most promising model is not one in which every merchant abandons its storefront or replaces its operations with an assistant. It is one in which customers can discover and act through new interfaces while the business's own systems remain authoritative.

That creates a second challenge: understanding what happened. A merchant may receive a referral or an order without knowing the customer's earlier comparison journey. Instrumentation can improve visibility into identifiable handoffs and downstream outcomes, but private assistant conversations generally remain outside the merchant's view.

For businesses, the sensible first move is to audit access to their commercial information and the reliability of existing transaction paths.

Check the public readiness of your website with Arch. The scan examines public signals and visible action paths. It does not claim to place orders or certify live agent checkout integrations.