# How to get your store ready for shoppers who send an AI agent

> An agentic commerce readiness checklist for online stores. Six things to check on one category, from product data and live prices to agent rules and checkout.

Page: https://opsalute.ai/en/blog/prepare-store-for-ai-agents
Русская версия: https://opsalute.ai/llms/ru/blog/prepare-store-for-ai-agents.md
Published: 2026-10-15 · Shelfport

The head of e-commerce at a home appliance retailer wanted to know how ready his store was for agentic commerce, where an AI assistant picks products for the shopper. So he typed what his customers type: "I need a quiet espresso grinder under $200, delivered by the weekend."

The answer came back fast and confident. Three models, one of which his store has sold for two years and keeps in stock. But the assistant sent him to buy it somewhere else.

About his own store it said the grinder cost around $230, although the price had been cut a month earlier. Delivery times? "Check the retailer's website."

Nobody did anything wrong on purpose. The new price lived in a banner, and the shipping terms in a pop-up.

A person would have found them. Software reading the site on that person's behalf did not.

## What changes when an assistant does the choosing

More and more often the shopper does not browse the catalog. They describe the job, and the assistant searches, compares and asks follow-up questions. Your store now has a second reader, and it reads differently.

Most of the data so far comes from the US, but the direction is clear. Adobe reports that traffic from AI services to US retail sites grew 393% year over year in the first quarter of 2026. In the same report, Adobe says major portions of US retail websites are not entirely readable by machines.

And in its 2025 holiday season review, Salesforce found that shoppers arriving from AI search converted nine times more often than those coming from social media.

A person on your site forgives a lot. They zoom into photos, find the size chart, read the fine print in the footer.

An assistant needs facts in a clear structure, and it needs to trust that they are true today. When a fact is missing, it guesses, or it takes the fact from a store that has it.

The path to purchase is changing too. Assistants are learning to build carts and pass the shopper on to checkout. For that, platforms agree on [agentic commerce protocols](https://opsalute.ai/en/blog/agentic-commerce-protocols): shared rules for how an assistant and a store exchange data and requests.

OpenAI and Stripe maintain the open Agentic Commerce Protocol. Google co-developed the Universal Commerce Protocol with Shopify, Etsy, Wayfair, Target and Walmart. Shopify turns on agentic storefronts for eligible stores, so their products can be discovered in ChatGPT, Google AI Mode and Gemini, and Microsoft Copilot.

Where the order is placed depends on the platform. In March 2026 OpenAI said ChatGPT would focus on product discovery and comparison, with purchases completed in the merchant's own checkout.

Google's checkout inside AI Mode and Gemini is open to select merchants and, for now, only to products eligible in the United States, Canada and Australia. The rules move fast, so it makes sense to be ready for both setups.

## The agentic commerce readiness checklist

Here are six things an assistant needs to consider your product and carry the shopper through to an order. You can check each of them on a single category.

### 1. Product data

Assistants compare products attribute by attribute. For a grinder that means burr type, number of grind settings, noise level, hopper size and warranty. If the product page says "made for true coffee lovers" instead, there is nothing to compare.

Make sure every product has a title without marketing fluff, a brand, a model, a manufacturer part number or barcode, a plain-text description and its key attributes as separate fields. Treat colors and sizes as variants of one product, each with its own identifier. Otherwise the assistant cannot tell that the black and the white grinder are the same model.

### 2. Prices and stock from a live source

A stale price is the most common mistake and the most damaging one. The assistant quotes one figure, the shopper sees another on your site and stops trusting both of you.

Prices and stock levels in the data you give assistants should come from the same source as your website: your commerce platform or ERP. A manual export once a week will not do.

OpenAI's product feed documentation recommends sending the full feed once a day and pushing price and availability updates during the day as they change. If an item ships on backorder, say so and say when.

### 3. Shipping and returns

Shoppers do not only ask "what should I buy". They ask "when will it arrive" and "can I send it back".

OpenAI's feed specification has dedicated fields for shipping cost, whether returns are accepted and the return window in days. It also states that a missing shipping cost means the charge is unknown. Nobody will assume it is free.

Write your shipping and return terms so they fit in a couple of sentences: where you deliver, what it costs, how many days it takes, which items cannot be returned. Terms buried in a twenty-page legal document will be summarized badly or not at all.

### 4. Rules for the agent

When an assistant builds an order, it acts for the shopper but under your rules. Write those rules down in advance.

Which actions are allowed: search only, or also carts and reservations. Whether discounts can be offered and within what limits. What to say about backordered items. When to hand the conversation to a person on your team.

One rule matters most: prices, discounts and stock are set by your store's systems. Text generated by a model does not change them. If a shopper talks a model into "giving them 30% off", that must not turn into a discounted order.

### 5. The path to checkout

A recommendation with no way to buy is worth little. The assistant needs a clear route: put a specific product variant in a cart and hand the shopper over to checkout. Shopify's developer documentation shows how this works: the agent builds a cart, gets a checkout link, and the shopper continues the purchase in the store.

Check that a link from an assistant lands in the cart or on the right variant's page, not on your home page. And check that the cart survives when the shopper is not logged in. We look closer at [where stores lose shoppers between a recommendation and an order](https://opsalute.ai/en/blog/ai-recommendation-to-order) in a separate article.

### 6. Where the shopper confirms the purchase

The final yes comes from a person. They review the items, the total, the address and the payment method. The protocols build this in as well: the description of AP2, an open protocol for agent payments, says the user must always be in control.

Decide where that confirmation happens for you: in your own checkout, or in the platform's interface if you are connected to one. Then check that an order from an assistant lands in your systems like any other order, tagged with where it came from. Without that tag you will never know whether the new channel works.

## Where to start: one category

You do not need to prepare the whole catalog at once. Pick one category where shoppers take their time and ask questions: coffee grinders, child car seats, tents. Stable stock and simple shipping make it easier.

1. Write down 15 to 20 questions shoppers in this category ask your sales and support teams.
2. Ask them to several assistants. Note where your store is named, where competitors are, and where the answers get price, stock or terms wrong.
3. Fix the data: attributes, variants, shipping, returns.
4. Trace where prices and stock come from and how long a change in your ERP takes to reach the data assistants see.
5. Walk the whole purchase path, from the assistant's answer to a confirmed order in your back office.
6. A month later, ask the same questions again and compare.

To be straight about the outcome: getting ready does not guarantee that an assistant will pick your product. The assistant decides what to show. What preparation gives you is a fair chance to be considered, compared on accurate data and carried through to an order when the choice falls on you.

## Where Shelfport fits

Shelfport is an Opsalute product for selling through AI assistants, currently in beta. A store connects its catalog through a feed, an API or its ERP and decides what an agent is allowed to do.

Prices and stock come from the store's own systems, and the shopper places the order in the store's checkout. The website, payment provider and back office stay as they are.

During the beta we help stores get ready and test the purchase path on their own products, starting with one category. [See how Shelfport works](https://opsalute.ai/en/shelfport).

## Sources

- [Adobe: AI traffic to US retail sites, April 2026](https://business.adobe.com/blog/ai-traffic-surge-retail-sites-not-machine-readable)
- [Salesforce: 2025 holiday shopping data](https://www.salesforce.com/news/stories/2025-holiday-shopping-data/)
- [OpenAI: Powering Product Discovery in ChatGPT, March 24, 2026](https://openai.com/index/powering-product-discovery-in-chatgpt/)
- [OpenAI: product feed specification](https://developers.openai.com/commerce/specs/feed)
- [OpenAI: getting started with product feeds](https://developers.openai.com/commerce/guides/get-started)
- [Agentic Commerce Protocol on GitHub](https://github.com/agentic-commerce-protocol/agentic-commerce-protocol)
- [Google: Universal Commerce Protocol, January 11, 2026](https://blog.google/products/ads-commerce/agentic-commerce-ai-tools-protocol-retailers-platforms/)
- [Google Merchant Center: UCP-powered checkout](https://support.google.com/merchants/answer/16837055?hl=en)
- [Shopify: agentic storefronts](https://help.shopify.com/en/manual/online-sales-channels/agentic-storefronts)
- [Shopify: Checkout MCP](https://shopify.dev/docs/agents/carts-and-checkout/checkout-mcp)
- [AP2: Agent Payments Protocol](https://ap2-protocol.org/)