What agentic commerce is, which agent shopping and checkout programmes were really live by August 2026, what agents need from your catalog, stock and policies, and a six-area readiness scorecard.
Author
Anichur Rahaman
An AI assistant can check forty stores in a few seconds. On a Tuesday morning in August, a shopper types into one: "waterproof jacket under 150 dollars, arrives before the weekend." It shows three. The owner of a small outdoor-gear shop, with 400 products, a tidy storefront and steady traffic, never learns the search happened, and her best-selling jacket is not one of the three.
Nothing is wrong with the jacket. Its waterproof rating sits inside a marketing paragraph, its stock count refreshes once a night, and the returns page promises "easy returns" without a single number. She was not rejected. She was never shown, and her analytics will not record a thing. (This is an illustrative scene, not a real shop.)
When software does the shortlisting, a store wins on facts the software can verify, not on persuasion. That is the idea behind agentic commerce: software agents that discover products, compare them, check stock and prices, and in some cases complete the purchase on behalf of a person.
The programmes are young, some have already been reshaped, and the hype runs ahead of the numbers. But the direction is clear enough that every merchant should know what these agents need from a shop.
The rest of this article separates what was actually live by August 2026 from what was only announced, explains what an agent needs from your data and policies, and ends with a readiness scorecard you can run in an afternoon.
This is part 1 of a three-part series, "Selling to AI Agents". Part 2 covers getting your products recommended by AI assistants, and part 3 covers agent-ready checkout and payments.
A chatbot answers questions. An agent acts. In shopping terms, an agent can take a request such as "a waterproof jacket under 150 dollars that arrives before the weekend", search many stores, filter by what it can verify, put one item in a cart and, if the shopper has granted permission, pay.
It helps to split the journey into stages, because different stages are live at different speeds:
Discovery and comparison are the stages most assistants already offer. Buying inside the assistant is more complicated, as the next section shows.
The public record shows a mix of working features, retreats and announcements.
| Programme | What it is | Status in August 2026 |
|---|---|---|
| OpenAI Instant Checkout and the Agentic Commerce Protocol (with Stripe) | Buy inside ChatGPT; open protocol for merchants | Launched September 2025 with Etsy sellers in the US. Reported ended in March 2026; checkout moved to merchants' own systems and apps |
| Google Universal Commerce Protocol (UCP) and agentic checkout | Open standard for catalog and checkout, used in Google Search and Gemini | Announced January 2026 with Shopify, Etsy, Wayfair, Target and Walmart. Universal Cart set to launch in the US in summer 2026 |
| Google Agent Payments Protocol (AP2) | Verifiable permission for an agent to pay within limits the shopper sets | Announced September 2025; part of the Universal Cart plan |
| Perplexity with PayPal (Instant Buy) | Checkout inside the answer engine | Live for US users since November 2025 |
| Amazon Buy for Me | Amazon's agent buys from other brands' sites | Launched in April 2025 as a limited feature |
| Shopify Agentic Storefronts | Catalog syndicated to AI assistants | Introduced in the Winter '26 Edition |
On 29 September 2025, OpenAI announced Instant Checkout in ChatGPT and published the Agentic Commerce Protocol, built with Stripe, so that other merchants and developers could integrate. Etsy sellers in the US went first, with Shopify merchants to follow.
In March 2026, The Information and CNBC reported that OpenAI was ending Instant Checkout in its original form. Purchases would instead complete in merchants' own checkout or in retailer apps, and ChatGPT would concentrate on product discovery. Walmart's EVP of AI Acceleration, Daniel Danker, told Wired that conversion inside ChatGPT checkout was about one-third of Walmart's own website. Product data inaccuracy was one of the reasons given.
The lesson is not that agents fail. The shopper still wants the merchant's checkout, loyalty account and delivery promise, and bad product data makes a buying agent unreliable.
At the National Retail Federation show on 11 January 2026, Google and Shopify announced the Universal Commerce Protocol. Etsy, Wayfair, Target and Walmart signed on as co-developers, with more than twenty other partners including Visa, Mastercard and Stripe. The idea: a merchant publishes one machine-readable description of its catalog and checkout, and any compliant agent can use it.
On 19 May 2026 at Google I/O, Google introduced the Universal Cart, with agentic checkout through Google Pay and AP2 guardrails (shopper-set spending limits and product preferences). It was scheduled to launch in the US in summer 2026 across Search and the Gemini app. Treat anything beyond that, such as other countries, as a plan, not a feature.
Perplexity and PayPal launched Instant Buy for US users in November 2025. Amazon's Buy for Me, from April 2025, lets Amazon's agent order from other brands' own sites, which raised real concerns among merchants about control over pricing and customer data. Shopify says AI-attributed orders on its platform grew eleven times between January 2025 and January 2026, from a small base.
An agent cannot be charmed by a lifestyle photo. It needs facts it can check. Four gates decide whether you reach the shortlist, and the fifth item below keeps you there once the order starts.
Title, brand, GTIN or other identifier, price, currency, size, colour, material, dimensions, weight and variants, all as separate fields, not buried in a paragraph. If "waterproof" appears only in a marketing sentence, an agent filtering on waterproof may never find the product.
An agent that reports "in stock" and then fails at checkout damages trust in both of you. Stock must come from the same ledger your sales channels use, and price changes must reach every surface quickly. A feed that refreshes once a day is already too slow for an assistant that checks at the moment of purchase.
Delivery cost by region, delivery windows, cut-off times, return period, who pays for return shipping, refund method. Write them as specific rules a machine can read, not "fast delivery, easy returns". Agents rank on what they can verify.
Tracking is part of the sale. After the purchase, the agent must be able to ask "where is order 1042?" and get a status such as paid, packed, shipped or delivered, with a tracking link. If your answer lives only in an email, the agent cannot relay it.
Stable endpoints for catalog, cart, order creation and status, with authentication, rate limits and clear error messages. You do not have to implement every new protocol on day one. If your platform exposes clean data and a documented API, adopting a standard later is an adapter, not a rebuild.
Brand. When an agent summarises your product in one line, your tone of voice disappears unless it is built into the facts. Trust signals that survive summarisation are verifiable ones: consistent reviews, clear policies, a track record of delivering on time.
Merchandising. Product pages written for people still matter, but the underlying data now carries equal weight. Attributes, naming conventions and variants become a competitive asset. Bundles and "complete the look" logic must be expressed as data, not just as a banner.
Metrics. Classic funnel numbers will blur. Sessions may fall while orders hold, because the research happened elsewhere. Add these to your dashboard:
Score each area from 0 (absent) to 3 (complete and automated). A total under 8 means start with data. Between 8 and 12 means fix gaps before adding channels. Above 12 you are ready to connect to agent programmes as they mature.
When you evaluate platforms for this work, look for one inventory ledger behind every channel and an open API, which is the approach of StoreConsole's inventory module.
You do not need to wait for a winner among the protocols. The programmes above have changed shape within months, and one has already been withdrawn. The data work, however, pays off whichever protocol wins: it improves your search ranking, your feeds, your customer support and your own storefront.
A workable order: clean the catalog, tie stock and price to one source of truth, rewrite the policies as rules, expose order status, then connect to the programmes that fit your market. Part 2 shows how to get recommended once the data is ready, and part 3 covers checkout and payments for the moment an agent actually pays.
Back in the outdoor shop, the owner spends three weeks on exactly this. Waterproof rating, fabric and weight become separate fields. Stock reads from the same ledger as the till, and the returns page now says: 30 days, free label for unworn items, refund to the original payment within five days. The next shopper types the same sentence. This time the jacket passes all four gates and is one of the three on the screen, and a week later an order arrives tagged to an AI assistant, the first line in her reports she had never seen before.
Anichur Rahaman is a software architect and the creator of StoreConsole. He designs commerce and ERP systems for growing businesses, with a focus on event-driven architecture, data integrity and self-hosted operations.
About the Author
Anichur Rahaman
Continue Reading
