How AI Agents Buy Products Online (and How to Be Chosen)

The Purchase Is Moving Into the Chat

Recommendation was the first act. The second act is transaction. A growing set of AI agents no longer stop at "here's the product you want" — they proceed to buy it. Understanding the mechanics of that flow is now part of ecommerce strategy, because the store that's easiest for an agent to transact with has a real advantage over the store that isn't.

There are two distinct ways an agent completes a purchase, and they demand different things from your store.

Path One: Browser Automation

The oldest approach is the most literal: the agent drives a browser the way a human would. It navigates to your site, reads the page, clicks buttons, fills in fields, and submits the order. OpenAI's Operator and Perplexity's Comet browser are examples of this style of agent.

Browser-driven agents live or die on your front-end. If a price only appears after a heavy JavaScript render, if the "add to cart" control is buried behind an unlabeled element, or if your checkout throws unexpected steps, the agent stalls. Machine-readable, standard HTML — proper form labels, accessible controls, prices and availability present in the markup — is what lets an automated agent actually complete the task.

This path also has friction beyond the technical. Automated access to sites is contested territory: in early 2026 a US court granted Amazon a preliminary injunction blocking Perplexity's Comet agent from accessing password-protected sections of Amazon, an early sign that browser automation on third-party platforms will be governed by legal boundaries, not just technical ones.

Path Two: Protocol-Based Checkout

The newer and more scalable approach skips the browser entirely. Instead of pretending to be a human, the agent talks to your commerce backend through a standardized protocol.

The Agentic Commerce Protocol (ACP), co-developed by OpenAI and Stripe and open-sourced in September 2025, works like this: when a shopper confirms a purchase inside the agent, the agent sends the order details to the merchant's backend. The merchant accepts or declines, charges the customer through its existing payment provider, and handles fulfillment and support exactly as it does for any other order. The agent is a conduit; you remain the merchant of record. For merchants already on Stripe, enabling agentic payments can be as small as a configuration change rather than a rebuild.

Google's Agent Payments Protocol (AP2), announced in 2025 with a broad coalition of payment and platform partners, solves the authorization question. It uses signed "mandates" — carried as verifiable credentials — so a merchant or payment network can trust that a real person approved a specific cart at a specific price. Google has built shopping experiences on top of this, and card networks including Visa and Mastercard have launched their own agent-payment schemes that interoperate with the wider ecosystem.

A note on honesty: this landscape is moving fast, and individual features get launched, renamed, and restructured on short cycles. The protocols are live and real, but they are not yet how most purchases happen. Optimize for the direction, and don't over-index on any single feature that may look different next quarter.

What Actually Gets an Agent to Choose You

Whether the agent automates a browser or calls a protocol, it has to decide which product to buy first. That decision runs on the same signals that drive recommendation:

Structured specifications with standardized attribute values, not facts trapped in paragraphs
Reliable pricing and availability the agent can read without guessing
Review signals — how many, how recent, how specific
Product and Review schema that converts your page into extractable data
Feed completeness across your catalog, with accurate identifiers

The through-line is that agents reason over data, not design. A store with a complete, accurate catalog is legible to every agent; a store with a beautiful page and a thin data layer is legible to none. This is exactly why product optimization for AI channels focuses on attribute completeness and specificity rather than persuasive copy.

Make the Transaction Frictionless

Being chosen is necessary but not sufficient — the agent still has to complete the buy. Two things reduce the chance it fails at the finish line:

A clean technical layer. Readable markup, accessible controls, and pricing and stock in the HTML help browser-driven agents finish the job. This is the substance of a solid technical foundation.
Protocol readiness. If your platform supports an agentic commerce protocol, enabling it removes the automation guesswork entirely and lets qualified agents transact directly.

How to Prepare Without Chasing Every Announcement

The temptation is to react to each new headline. A steadier approach:

Fix the data first. Complete, structured, specific product data is the one investment that pays off no matter which agent or protocol wins. It improves both recommendation and checkout.
Harden the front-end. Even if you adopt a protocol later, a machine-readable storefront helps today's browser-driven agents and tomorrow's crawlers.
Watch platform terms, not just tech. As the Amazon–Perplexity dispute shows, who is allowed to transact on which surface is being decided in parallel with how. Know the rules for the platforms you sell on.
Measure what agents see. You can't optimize what you can't observe.

That last point is where most brands are flying blind. An AI visibility audit reveals whether agents can find, understand, and act on your products — and pinpoints the gaps between "recommended" and "actually purchasable."

The agents that buy on a shopper's behalf are still early, but the qualifying criteria are already set. The stores winning the recommendation are the same ones positioned to win the transaction: legible, specific, and ready to be acted on.