When AI comes to shop, can it get in? Meta and Sierra’s PAP and the US retail gap

An AI assistant recommends your product. The customer then says, “Buy it for me.” Can that assistant complete an order at your store?

Two announcements on October 6, 2026 bring that question into focus. Meta and Sierra introduced Personal Agent Protocol (PAP), a proposed way for personal AI agents to connect with businesses. Rezolve Ai also published research in which it could not confirm a public purchasing interface at 97 of America's 100 largest retailers.

For a merchant, the useful connection is practical: getting recommended and accepting an agent's order require attention at different stages of the shopping journey. Here is what the announcements mean, and what you can check in your own business now.

What is PAP, the proposed connection protocol for agentic commerce?

PAP connects a customer's personal AI with a business

PAP is a proposed set of rules for interactions between personal agents and businesses. Its aim is to let customers authorize access while companies define the actions they make available. Sierra's announcement

Consider a shopper looking for a bag for business trips. They ask an assistant to compare options, select a product, and eventually delegate the purchase. The shopper wants the chosen item; the merchant wants to accept a valid order from the right customer.

Even a capable product recommendation cannot resolve uncertainty about whether the assistant may enter the site or change an order. A shared way to establish access and permitted actions would help the two sides work together. That is the problem PAP is intended to address.

In this article, AI commerce includes AI-assisted product discovery and purchasing. Agentic commerce is the part where an agent takes actions on the customer's behalf—the area most directly relevant to PAP.

Walmart, Shopify and Stripe's participation does not establish availability for every store

Participants include Meta and Sierra alongside Walmart, Shopify and Stripe. Their involvement makes the initiative relevant to retailers assessing future support from commerce and payment providers.

At the time of the announcement, however, PAP was under development. Sierra said it planned to release the first specification, v0.1, later in October 2026. The announcement reviewed on October 7 does not establish that the participating companies have made it available to every customer. Official rollout plans

When assessing your own store, ask which features will be available under your contract and checkout configuration. Keep the list of participating companies separate from the list of functions your business can actually use. That distinction makes an internal rollout decision easier to defend.

How PAP approaches the actions an AI agent may take

Separate product research from access to a customer account

Sierra describes a design in which an agent can begin without account access and obtain access when the task requires it. OAuth, an established authorization mechanism, provides the foundation. PAP's proposed design

For example, stock availability and return conditions may be public information. Changing the shipping address on an existing order requires the business to identify the order and establish the customer's permission to change it.

Map that distinction onto your own operations. Reading product details, accessing a customer account and changing an order have different requirements. Recording the identity checks and evidence needed for each action gives your provider something concrete to evaluate.

Connections can use websites, APIs such as MCP, or a business's own agent

The proposed routes include regular websites, APIs using interfaces such as MCP and OpenAPI, and a company's own agent. An API is an interface through which software can use a service's functions. Proposed connection routes

These routes also help identify who should join the discussion. The ecommerce team may own the shopping interface, developers may own connections to order systems, and support staff may need to review conversational return requests.

Having an MCP connection alone does not settle the business requirements. Your team still needs to establish who permits an action, how its outcome is recorded and who handles an exception.

Our Japanese guide to Shopify's WebMCP checkout looks at specific operations exposed in a purchasing flow. PAP raises the broader question of how the customer's agent connects with a business when using such capabilities.

Detailed action limits and payment extensions remain future possibilities

Sierra discusses more detailed permissions and payment extensions as potential future developments. They should not be presented as settings a merchant can already enable. Future development areas

You can still document your own requirements. Decide what a customer should confirm before an order is placed, what should happen if the total changes, and who will handle cancellation or returns. These are criteria for evaluating a future implementation, rather than instructions for installing PAP today.

Why the study could not confirm purchasing interfaces at 97 of 100 US retailers

Distinguish 84 interfaces not found from 13 scans denied

Rezolve Ai reports that its October 3–5, 2026 public-site scan found no published purchasing interface at 84 of the top 100 US retailers. Another 13 denied the scan. A public interface was detected at three. Purchasing-interface results

Observed result	Retailers
No public purchasing interface found	84
Scan denied; interface could not be checked	13
Public purchasing interface found	3

The investigation looked for publicly discoverable purchasing connections or browser tools. Private partnerships and purchases made by an agent operating the ordinary shopping interface cannot be inferred from this count. Scope and method

Describing all 97 as rejecting AI shopping would erase a useful distinction. Keep “not found” and “could not access” separate in your own assessment: the former calls for a feature inquiry, while the latter may require a conversation with the team responsible for access controls.

Finding an interface and completing an order need different checks

An interface scan does not demonstrate that shipping calculations, customer permission, payment and order processing all worked. These results are not a measure of completed purchases. What the investigation checks

Ask a provider to walk you through one purchase flow. Follow the product selection, shipping details, final total, customer confirmation and the order's arrival in your management system. This turns a broad claim of support into a set of observable steps.

Also ask what happens if an item goes out of stock or the shipping address cannot be accepted. Those answers help your support team prepare. Any test that could create a real order should use the provider's test environment or an agreed procedure with a clear owner for checking the result.

AI visibility and successful agentic purchasing are different stages

About one in four companies was not named in the sampled answers

Rezolve Ai also examined AI answers concerning 2,047 companies. In that answer set, 498 companies—about one in four—were never named. AI visibility report

The answer study ran from September 23 to October 1. It used 15 unbranded questions per company across four AI systems. It did not collect every question real consumers asked. Answer-study methodology

AI search optimization starts with understanding how products and businesses appear in answers. Choose questions that reflect your customers' comparisons, then examine which products are recommended and what information supports those recommendations.

For a bag retailer, a customer asking for a lightweight cabin bag needs different information from someone shopping for a rainy commute. Read the explanations of weight, dimensions and use cases as well as checking whether the product name appears. That gives the content team a more useful starting point for improving product information.

The 1.5% citation share is not a share of traffic or revenue

In the report, the target company's own website accounted for 1.5% of citation sources. This describes the sources cited by the AI answers, rather than website visits or sales. Citation-source results

Someone who sees a recommendation may visit your store or continue comparing products elsewhere. Counting more mentions alone will not show which part of the path to an order improved.

Likewise, an agent-compatible checkout has fewer opportunities to be used if shoppers never encounter the product. Taken together, the announcements suggest a practical way to organize work: review both the information that gets a product considered and the purchasing process available after it is selected.

Rezolve Ai sells services in this market. Treat the results as observations under the company's published research conditions, rather than a direct forecast of demand or revenue for your store or for Japanese ecommerce as a whole.

What merchants in Japan can check now

Organize product information and purchase conditions

Start with one representative product and list the information needed before someone buys it. The following bag-store example can be adapted to your own products and policies.

Question	Information to check
Can the shopper choose the right product?	Model, dimensions, weight, materials and differences in intended use
Are purchase conditions clear?	Price, availability, shipping charges and destination restrictions
Is delivery timing understandable?	Dispatch estimates and conditions for preorders
Can the shopper get help afterward?	Return and exchange conditions, plus a contact route

Record gaps as specific changes to specific pages. “Link the product page to the return conditions” is easier to assign than “improve returns information.” If the product page and FAQ disagree, resolve the discrepancy first.

Providing consistent answers about the same product also improves ordinary customer guidance. It gives the team a useful task now, while connection specifications and provider support continue to develop.

Ask your platform about connections and customer permission

Contact the commerce platform or payment provider with details of your contract and checkout. Specific questions tend to produce answers that are more useful than a general inquiry about AI support.

• Which AI-facing connections are available for our contract and checkout configuration today?
• Can an agent search products, edit a cart, place an order, or perform only some of those actions?
• Where is the customer's permission obtained, and which action does it authorize?
• Can a person take over after a failure, and how are duplicate orders prevented?
• Can we review documentation and test procedures that distinguish available, trial and planned features?

Keep the contact, confirmation date and relevant features with the answer. As specifications and product availability change, you will know which assumptions need to be checked again.

Track recommendations, visits and completed orders separately

Maintain separate records for AI recommendations, website visits and orders. These help determine whether the next task concerns product information, the path through the site, or purchasing operations.

Stage	Useful records
AI recommendation	Question, date, AI system, recommended product and cited sources
Website visit	Observable source, landing page and subsequent actions
Order completion	Order result, identifiable path, and any failed step or explanation

You may not be able to connect all three records to the same customer. Leave an order's source unknown when the evidence is unavailable, and explain that limitation in reporting. Assigning sales to AI based on an impression can distort the evaluation of your work.

Choose one product and assemble its customer-facing information, your provider's confirmed capabilities and a checklist for completing an order. That gives your team a concrete starting point while it waits for the PAP specification and applicable services.