AI shopping agents: what Walmart's gateway and Amazon's block mean for your store

Two of the largest US retailers now disagree in public about whether AI shopping agents may buy anything at all. On October 21, 2026, Walmart announced an Agent Gateway, per-agent profiles and formal terms for agents that shop on a customer's behalf. Amazon has been going the other way since it cut off Meta's Muse agent on the night of September 20. If you run an online store, the standoff is not the useful part. In our view the work that pays off either way is your product data, not an in-chat buy button.

What AI shopping agents can actually buy right now

AI shopping agents are assistants that search catalogs, compare items and, on some surfaces, place the order. What they can complete depends entirely on the retailer or platform at the other end, and that picture changed twice this year.

The clearest reversal came from OpenAI. In March, a spokesperson said that "Instant Checkout is moving to Apps, where purchases can happen more seamlessly", as reported by Retail Insight Network from CNBC's story. Products stay discoverable inside ChatGPT, but shoppers finish on the merchant's own store in an in-app browser or a separate tab. Etsy, listed in our E-commerce category, told the same reporting that its own app would give it more control over how its shopping experience appears; Etsy had been one of the first Instant Checkout sellers.

So the honest answer to "can an agent buy from me today?" is: on Walmart's rails, soon and under rules; through Shopify's agentic channels, on some surfaces; and in ChatGPT, usually by sending the shopper to your checkout.

Walmart's Agent Gateway gives agents house rules, not free rein

Walmart's announcement has three parts, and the detail matters more than the headline.

• The gateway. An MCP server - Model Context Protocol, the open standard Anthropic created - lets approved agents search Walmart's online inventory and place orders at Walmart and Sam's Club without receiving the shopper's personal details or purchase history.
• Agent Profiles. Rolling out next month, each agent gets its own profile inside the customer's account, so its purchases sit separately from the customer's own. Customers can require approval for every transaction or pre-authorize purchases under a spending limit they set, which Walmart compares to adding an authorized user on a credit card. Profiles are created the first time an agent shops, can be deleted at any time, and need no Walmart+ membership.
• Terms for agents. Walmart's returns process applies even when an agent does not act as intended; agents may not exceed what the shopper intended, may not touch payment information, and may only access the data the task needs. Agents that ignore the terms get blocked.

All of that comes from Walmart's launch, which CEO John Furner discussed at the Masters of Scale Summit in San Francisco. Walmart has not said how third-party Marketplace sellers' items are handled inside the gateway, and we would not assume parity with first-party stock until it does.

Amazon's block shows the real lever is terms of service

Amazon's block page told Muse users that "continued access by an unauthorized AI agent violates Amazon's Conditions of Use, to which our customers have agreed." Amazon said Meta never told it Muse would access the store, that the agent does not identify itself while browsing, and that it "appears to capture and store customer credentials" - a claim Amazon published no evidence for, and which Meta disputes.

The mechanics are the part worth copying down. When Search Engine Journal fetched Amazon's robots.txt on September 22, it found 101 user-agent blocks, 100 of them refusing the whole site, covering 99 distinct crawlers including GPTBot, PerplexityBot, ClaudeBot and GoogleAgent-Shopping. None of them stopped Muse, because Muse sends no user-agent string of its own. Amazon enforced its terms against the logged-in shopper instead.

For your store, that is two separate dials: robots.txt only refuses agents that name themselves, and your terms of service reach the customer, not the agent vendor. If you want a policy on agent traffic, it belongs in both places. Our view is that blocking is the wrong default for a store that needs demand - Amazon can afford to be a destination, and most stores cannot.

The rails: UCP, Shopify's channels and Stripe's feed

The plumbing consolidated faster than the retail politics. Shopify and Google introduced the Universal Commerce Protocol on January 11, an open standard for agents to transact with any merchant, which Shopify says was endorsed by more than 20 retailers and platforms and works with any payment processor, including Shopify Payments, over REST, MCP, AP2 and A2A.

For merchants on Shopify, the practical surface is Agentic Storefronts. Shopify says products "become discoverable in ChatGPT by default", that the AI channels are "managed centrally from the Shopify Admin", and that there are "no transaction fees beyond standard processing rates" on those sales, with Shop Pay "coming soon" to Microsoft Copilot, per its agentic commerce update of March 24. Brands that do not run their store on Shopify can list products in Shopify Catalog through its Agentic plan, which the same page says is now publicly available.

If you are not on Shopify, Stripe documents the other path. Its catalog feed takes a CSV through the Product Catalog Import API and syndicates it to agents, and its agentic commerce docs cover selling to agents as well as building them. Stripe states plainly that availability, launch stage and supported countries vary by integration, so check your own before planning around it. Upfynd lists 124 products in Payments & Banking and 147 in E-commerce as of October 2026; the agent-ready subset is much smaller than either number.

What to do with your store this quarter

The requirements in Stripe's feed spec are, in our view, the clearest public statement of what an agent needs to consider your product at all, and they are platform-agnostic enough to work as a checklist.

1. Fill the required fields first. Stripe's spec marks `id`, `title`, `description`, `link`, `image_link`, `availability` and `price` as required, `brand` as required for most categories, `gtin` as recommended and `mpn` as required when a GTIN is missing. Start with your best sellers.
2. Keep stock and price honest. Stripe provides separate inventory and price feeds precisely so you can push changes often without resubmitting the catalog. In our view an agent ordering something you do not have is the failure hardest to recover from, because the shopper never saw your listing.
3. Decide discovery versus checkout per product. Setting `disable_checkout=true` keeps a product syndicated for discovery while sending the shopper to its `link` to buy. In our view that is the right setting for anything with sizing, personalization or fragile margins.
4. Add the signals that rank you. Stripe lists `popularity_score`, `return_rate`, `product_review_count` and `product_review_rating` as recommended, aggregated fields. Reviews you already collect become ranking input - the same pattern we found in AI search visibility for SaaS.
5. Carry your compliance text. The `product_warning` field exists for legal disclaimers, safety warnings and California Proposition 65 notices, and Stripe says it shows them on the agent checkout screen.

Skip the deeper work if you sell one-off or made-to-order goods where every order needs a conversation. Structured feeds reward catalogs with stable SKUs, and in our view a 40-product handmade shop gets more from accurate stock and clear returns copy than from protocol work.

Which assistant your customers hand their card to is still unsettled - we compared two of them in OpenAI Dots vs Meta Muse. Start from the part that outlives the winner: a feed that describes your products accurately. Then browse the E-commerce category for the storefront and feed tools, or submit your product if you build one.