When the shopper is an AI agent

Two announcements in the last three weeks changed what "online shopping" might mean in a year or two.

On 16 September, Google published AP2, the Agent Payments Protocol. It is an open standard for the moment an AI agent pays on someone's behalf, built with more than sixty partners including Mastercard, American Express, PayPal and Adyen. Then on 29 September, OpenAI switched on Instant Checkout in ChatGPT. A user in the US can ask for, say, a ceramic vase under $100, see options, and buy one from an Etsy seller without leaving the chat. OpenAI and Stripe also published the technology behind it, the Agentic Commerce Protocol, as an open standard.

Instant Checkout is not available to shoppers in the UAE or India yet, and AP2 is a specification rather than something shoppers use anywhere. Both are worth understanding now, because they change what your store needs to be good at.

How it works for the merchant

The design is more conservative than the headlines suggest. In ChatGPT's version, the merchant stays the merchant. The order goes into your existing system, you take the payment through your existing processor, you ship it and you handle returns. ChatGPT acts as the shop window and the checkout counter. Everything after that is still yours.

For that to work, the store shares a product feed with OpenAI: titles, descriptions, prices, stock, images, shipping and return terms. ChatGPT reads that feed to decide what to show. When the user taps buy, it passes the order and a secure payment token to the merchant, who accepts or declines it like any other order. The merchant pays a fee on completed sales. Shoppers pay nothing extra.

Google's AP2 tackles a different problem: proving that a person really authorised an agent to spend money. It uses signed "mandates", a digital record of what the user asked for and agreed to, so that when an agent pays, everyone involved can check who asked for what. That matters most for disputes and chargebacks, which get messy fast when the buyer was software.

What this changes for a store

On a normal website, you control the whole visit: the layout, the photos, the upsell at checkout, the newsletter pop-up. Inside a chat, you control almost none of that. What you control is your data.

So the product feed becomes the storefront. When an agent compares five vases, it compares what the feeds say. Vague titles, missing dimensions, stock that is wrong half the time and return policies hidden in a PDF all count against you, in a way a human shopper might have forgiven.

A few other consequences are worth thinking through:

• Price comparison gets easier for the shopper and harder for you. An agent will happily check ten stores.
• Upsells mostly disappear. Single-item purchases are the starting point, and the agent buys what was asked for.
• The customer relationship gets thinner. You still get the order and the customer's details, but not the browsing, the email sign-up or the second visit.
• Reviews and clear policies matter more, because they are the main signals an agent can read.

What a store outside the US can do now

There is nothing to switch on yet in the Gulf or India, and that is fine. The work that prepares you for agent shopping is the same work that improves your Google Shopping listings and your own site search today.

1. Clean up your product data. Specific titles, every attribute filled in (size, material, colour, dimensions), real stock levels, and barcodes where you have them.
2. Add product structured data to your pages, so any machine reading them gets price, availability and reviews in a standard format.
3. Write shipping, delivery areas and return terms in plain text on the product page, not only in a policy page or a PDF.
4. Keep a single source of truth for price and stock. If your website, your marketplace listings and your feed disagree, an agent will notice before a customer does.
5. Check that a guest can buy without creating an account. Forced sign-ups are where people and software alike give up.

Should you worry?

Not yet. OpenAI started with a narrow pilot: US users, US Etsy sellers, single items, with Shopify merchants promised soon. Google's protocol is a specification, not a product you can sell through. Most of your customers will keep buying the way they do today for a while.

Still, more product discovery is moving into AI assistants, and an assistant can only recommend what your data describes well. Getting your catalogue into shape is cheap, and it already helps your Google Shopping listings and your own site search.