E-commercepurchases by AI agents

AI agent purchasing: is your online shop really ready?

ChatGPT Wallet opens the door to purchases by AI agents. Here are the checks to prioritise on your catalogue, payments and security.

Mehdi B.

Co-founder · SEO & digital strategy

AI agent purchasing: is your online shop really ready?
In brief

Since 13 November 2026, ChatGPT Wallet lets an AI agent buy on a merchant site, for now in the United States only. The agent goes through your pages like a customer: for it to reach the order, you need a consistent catalogue, well-configured anti-bot protections and a payment flow whose refusals you can understand.

Purchases by AI agents are no longer a hypothesis. Since 13 November 2026, an OpenAI assistant can buy on a merchant site in place of its user. For now, only a few US users have access. But the question already arises for every online shop: if an agent visits your site tomorrow, will it be able to find the right product, at the right price, and go all the way to the order?

The good news is that the preparation looks a lot like what a shop should be doing anyway. There is no need to buy an integration in a rush.

Purchases by AI agents: what ChatGPT Wallet changes

ChatGPT Wallet is a payment wallet built into ChatGPT, designed with American Express, Mastercard and Visa. It lets “dots”, the OpenAI agents that carry out tasks for the user, pay online.

The principle is simple. The agent does the research, chooses the items and fills the basket. Before paying, it shows the user the contents of the order, the total and the card used. The user confirms with Face ID or their fingerprint. Payment then goes through a token supplied by the card network: the merchant never sees the real card number.

The launch remains very limited: ChatGPT Pro subscribers, adults, in the United States, with a US card. Other countries will have to wait.

What matters to you: the agent needs no special access to your shop. It uses a remote browser and goes through your pages, like a customer. That is where the problems begin, because an agent cannot guess what a human understands at a glance.

First sticking point: a catalogue that contradicts itself

Person paying for an online purchase with a bank card on a laptop

A human customer puts up with a small inconsistency. They see “in stock” on the product page, then “out of stock” in the basket: they sigh, pick another size and carry on. An agent, on the other hand, stops, or reports false information to its user.

Take a clothes shop. A dress comes in four sizes, has a 20% discount and size L is sold out. For an agent to buy it without error, four pieces of information must stay identical from the product page to the confirmation:

  • the variant: the size and colour chosen, without ambiguity;
  • the price: with the discount, in the right currency;
  • availability: a displayed item cannot necessarily be ordered;
  • delivery: countries served, fees and lead times.

OpenAI also accepts product feeds, which help ChatGPT discover your offers. But a feed does not replace verification at the time of purchase. If the feed shows one price and the page another, the page prevails, and the order goes wrong.

The real work therefore happens upstream: knowing who changes a price, and where that change carries through. If your catalogue is already well maintained for Google Shopping or price comparison sites, you have done a good part of the job. We explain this link between product data and visibility in our article on e-commerce visibility in AI answers.

Second blocking point: your anti-bot protections

Many shops filter bots to protect themselves from price scraping, fake orders or attacks. That is sensible. The problem is that an agent’s browser looks just like a bot. It can be blocked without anyone noticing.

OpenAI offers Web Bot Auth for this: requests from its browser carry a signature that your site can verify. You then know the visit really comes from OpenAI, and not from an imitator.

Recognising a visitor does not mean opening everything to them. The agent can browse the catalogue and place an order, but customer accounts, the administration and sensitive actions keep their protections. OpenAI itself recommends this, in fact.

In practice, we only touch the rules when a block has been observed. We identify the rule responsible, create a limited exception, and make sure it can be removed easily.

Third sticking point: payment and fraud prevention

According to OpenAI, ChatGPT Wallet works with your current payment pages, with no specific development. But the company warns: some anti-fraud tools consider the remote browser suspicious and refuse legitimate purchases.

When a payment fails, you therefore need to know where. Site access, form, bank or anti-fraud rule: each cause calls for a different fix. Without this diagnosis, you risk changing the wrong setting, or even weakening your protection for nothing.

Another trap on the numbers side: user validation is not an order. What counts is your shop’s confirmation. A journey that reached the payment step is not a sale. Every attempt must end in a clear state: order confirmed, payment declined or journey abandoned.

Plugins and WebMCP: not until you have a specific need

OpenAI also offers more advanced ways to connect a shop to its agents. There are plugins, and WebMCP, a standard in preparation that lets a site describe its actions directly to agents: search for a product, add to basket, and so on. The company points out, however, that browser access works without a plugin or a product feed.

A direct connection makes the agent less dependent on your site’s layout. It still has a cost, at creation and then at every update. And OpenAI itself advises measuring the full journey: a plugin is not automatically faster.

Our advice: test the current journey first. If a specific problem appears, for example a variant selection that fails every time, an integration may be justified. Otherwise, the money is better spent on catalogue quality.

Is your shop ready? Five questions to ask

To get a first idea, put these questions to your team or your provider:

  1. Do the price and stock of a product with variants stay the same from the product page through to order confirmation?
  2. Are a promotion, a stock shortage or an unserved country clearly displayed, without the customer having to guess?
  3. Do our anti-bot protections block legitimate visits, and would we be able to see it?
  4. When a payment is declined, do you know the reason?
  5. Do our statistics distinguish a confirmed order from an attempted payment?

If you answer “I don’t know” to more than two questions, be aware that these points are already costing you sales today, with very human customers.

These questions apply to WooCommerce as well as to Shopify or a custom-built shop. To test a real payment with ChatGPT Wallet, you need an eligible user. As long as the service remains limited to the United States, we focus on what can be checked directly on your site.

Where to start

AI agents will arrive at their own pace, and nobody yet knows when they will really weigh on sales. What helps them also helps your current customers: an accurate catalogue, a payment that goes through, an order that can be tracked. That is where to start, and it is what we look at first in our e-commerce support.

Sources

E-commerceAI agentsPaymentsChatGPT
Mehdi B.
About the authorMehdi B.

Co-founder of Smart Impact. Hooked on the web from the start, he launched his first project in 2006: an online music magazine still running today. With close to 20 years of SEO experience, a Swiss federal diploma in marketing and a solid geek streak, he turns clients' ideas, sometimes hazy ones, into concrete digital projects with his team.

LinkedIn
Further reading