[email protected] 0330 122 6522
Abstract dark editorial graphic of an AI agent passing a basket of products through a checkout gate, illustrating agentic commerce and AI agents buying for people

Agentic Commerce: When AI Agents Do the Buying

Agentic commerce is when an AI agent shortlists, compares and buys on someone’s behalf, within limits they set. The shopper states a goal; the agent does the legwork. For retailers it means being chosen by software that reads product data, not persuaded by a page design. The work that matters now is accurate, structured, crawlable product information.

What is agentic commerce?

It is shopping with a delegate. Instead of opening eight tabs, the buyer tells an AI agent what they want, what they will spend and what matters, and the agent comes back with an answer or just gets on with it.

The word doing the work is “agent”. A chatbot answers a question. An agent acts: it searches, reads, compares, fills in forms and, if allowed, pays. That last step is what separates agentic commerce from the product recommendations I covered in our piece on ChatGPT shopping. Recommendations help a person decide. Agentic commerce removes the person from the middle of the process, at least for the dull purchases.

Think of reordering printer toner, booking a train, replacing a kettle, buying the same trousers in the next size. Nobody enjoys those. They are the first things people will hand over.

I should be straight about timing. This is early. The protocols are new, the volumes are small and anyone giving you a confident market-size forecast is selling something. But the direction is clear enough that the groundwork is worth doing, and most of it pays off today anyway.

How do AI agents actually buy things?

Through a handful of open protocols that let an agent prove it has the buyer’s permission and hand a merchant a payment without exposing card details. Two have been published so far by the big names, and both date from September 2025.

The Agentic Commerce Protocol

Stripe and OpenAI co-developed the Agentic Commerce Protocol (ACP), and Stripe describes it as open source and Apache 2.0 licensed. The detail worth knowing is who stays in charge: the business remains the merchant of record. It keeps the customer relationship and controls what is sold, how it is presented, how payments are processed and how orders are fulfilled.

The practical shape is that the agent sends the merchant a checkout request and a payment credential, and the merchant accepts or declines it using its usual payment and fraud signals. Instant Checkout in ChatGPT, which launched alongside it, is the first consumer version.

AP2

Google’s Agent Payments Protocol, announced on 16 September 2025, comes at the same problem from a different angle. Payment systems assume a human clicks “buy” on a page they trust. An autonomous agent breaks that assumption, so AP2 is built around three questions: was the agent authorised, does the request reflect what the buyer really wanted, and who is responsible if it goes wrong?

Its answer is signed digital records. An intent mandate captures what the buyer asked for and any limits, such as a price ceiling. A cart mandate records the exact items and price the buyer approved. Google says it worked with more than 60 organisations on it, including Adyen, American Express, Mastercard and PayPal.

What the two have in common

Neither asks the merchant to build a robot-friendly storefront from scratch. Both assume the merchant keeps control of fulfilment, support and returns. And both treat trust as the main problem to solve, not convenience. That tells you where the friction will be: not whether an agent can click a button, but whether your data and policies are clear enough for it to be allowed to.

How will an agent shortlist your products?

By filtering first and persuading never. An agent works from a brief with hard constraints, so the first pass is elimination on facts it can read.

Take a buyer who asks for waterproof walking boots, wide fit, delivered by Friday, under a set budget, with free returns. An agent can cross off anything where it cannot confirm each of those from the data in front of it. Not “probably waterproof”. Confirmed.

That has three consequences:

  • Missing data is a no. If delivery time is buried in a PDF or a tab behind JavaScript, the agent cannot verify it, so it moves on. A human might click through. An agent working through forty options will not.
  • Conflicting data is worse. If your feed says in stock and the page says sold out, you have given the agent a reason to distrust everything else you say.
  • Marketing language does nothing. “Our most loved boot” is not a filter criterion. “Waterproof membrane, EE width, 1.1 kg per pair” is.

Ranking after that point is not documented anywhere. OpenAI does not publish how it orders products, and neither does anyone else, so treat any claim about the secret formula with suspicion. What you can control is whether you survive the elimination round.

What does this change about discoverability?

It moves discovery from pages to data, and from your site to wherever the agent reads. You still need a good site. But the first reader of it may not be a person.

Three shifts are worth noticing.

The shortlist is formed before anyone visits you. In classic search, a weak page still got a click and a chance to win the buyer back. When an agent is doing the comparing, you are either in the answer or you are not. This is the same logic behind AI search generally, and it is why AI visibility is now a commercial question rather than a vanity metric.

Crawlers decide whether you exist. OpenAI’s crawler documentation explains that OAI-SearchBot is used to surface sites in ChatGPT search features, and that sites which opt out will not be shown in those answers. If you block the bot, or your content only appears after JavaScript runs, you have opted out without meaning to. We see it often.

Brand still matters, just differently. An agent asked for “a reliable brand” leans on reviews, mentions and consistent information across the web. Reputation gets turned into data it can read. If you want the background on how engines decide who to trust, our explainer on what AEO is covers the basics.

What should you fix in your product data now?

Start with the boring things: complete, accurate, consistent product information on pages a crawler can read. None of it is glamorous, and all of it compounds.

Work through this in order:

  1. Audit your product pages as a machine would see them. View the page source, or switch JavaScript off, and check that the name, price, availability and key specifications are still there. If not, that is your first job.
  2. Add Product structured data. Google’s product structured data documentation sets out the properties for price, availability and reviews. Mark them up properly and keep them in sync with what the page shows.
  3. Write specifications as facts. Dimensions, materials, compatibility, capacity, warranty length. One fact per line, units included. Cut the adjectives.
  4. Make delivery and returns explicit. Cost, timescale and the returns window should be stated on the product page or one click away in plain text. These are exactly the constraints a buyer will hand an agent.
  5. Keep every source consistent. Your site, your feed, your marketplace listings and your Google Business Profile should agree on price and stock. Contradiction is the quickest route to being dropped.
  6. Check your robots.txt. Decide deliberately which AI crawlers you allow. Blocking training bots and allowing search bots are separate decisions.
  7. Check your platform. Some platforms pass product data to AI engines for you, others need a feed or an integration. Find out which yours is before you assume.

If you run a headless store, you have an advantage here, because structured product data is usually a first-class citizen of the build. Our write-up on headless commerce goes into why. On a heavily plugin-driven site, the same job takes more untangling.

How worried should you be?

Be curious, not frightened. Agentic buying is a real shift, but it will arrive unevenly, starting with repeat purchases and low-risk categories and reaching considered ones much later.

My honest view is that most retailers will not lose sales to an agent next quarter. What they will lose, quietly, is consideration: the slow drift of buyers who ask an AI first, get a shortlist without them in it, and never arrive. That has been happening in search already.

The reassuring part is that you are not choosing between two projects. Clean product data, structured markup and a crawlable site are what Google rewards, what ChatGPT and Perplexity read, and what an agent will need. Do the work once.

If you want to know where you stand, the quickest route is to see your site the way an AI engine sees it. That is what our audit does.

Free AI audit

See how AI engines, and soon AI agents, read your site

Our free AI audit checks whether ChatGPT, Perplexity, Gemini and Google can find, read and recommend you, and whether your product and service information is clear enough to survive an agent's first filter. You get a plain-English list of what to fix, in order. If your site is already in good shape, we will say so.

  • Whether AI engines mention you today, and what they get wrong
  • Crawler access, structured data and content gaps on your key pages
  • A short, prioritised fix list rather than a sales pitch
Get my free AI audit

Frequently asked questions

What is agentic commerce?

Agentic commerce is buying done by an AI agent on a person's behalf. The agent understands what the buyer wants, researches options, builds a shortlist and, with permission, completes the purchase. The human sets the goal and the limits. The comparison and often the checkout are handled by software working through your product data and your site.

How is agentic commerce different from ChatGPT shopping?

ChatGPT shopping is mostly about discovery: it shows products and helps a buyer decide. Agentic commerce goes further, covering the agent acting for the buyer, including paying. Shopping results are one early piece of it. The wider shift is software that can shortlist, negotiate terms within limits and transact without the buyer visiting your site.

What are the Agentic Commerce Protocol and AP2?

Both are open protocols for agents making purchases. The Agentic Commerce Protocol was co-developed by Stripe and OpenAI and announced in September 2025. AP2, the Agent Payments Protocol, came from Google that same month. Each tackles how a merchant can trust that an agent has genuine authority from the buyer to pay.

Do small businesses need to prepare for AI agents buying?

Prepare, but do not panic. Volumes are small today and the protocols are still settling. The groundwork is cheap and useful anyway: accurate product data, structured markup, pages a crawler can read without JavaScript, and clear policies on delivery and returns. That helps with Google, ChatGPT and Perplexity today, and with agents tomorrow.

What product data do AI agents need?

Agents need data they can compare without guessing: a clear product name, price, currency, availability, delivery cost and time, returns terms, and specifications written as facts rather than marketing. Published as structured markup and kept consistent between your site and any feed, it lets an agent rule you in or out quickly.

Written by
Sam Wright · Founder, Aeonix

Sam Wright is the founder of Aeonix, an AI-first UK marketing agency. He writes about AEO, GEO and SEO, and what it takes to get found and cited now that buyers ask AI before they search Google. Less theory, more of what actually works.

Want this working for your business?

Start with a free audit, or talk to us about where to begin.