# The assistant recommended your product. So why is there no order? > Why AI referral traffic does not turn into orders. Four places stores lose shoppers after an AI assistant's recommendation, and how to find your own leaks. Page: https://opsalute.ai/en/blog/ai-recommendation-to-order Русская версия: https://opsalute.ai/llms/ru/blog/ai-recommendation-to-order.md Published: 2026-08-31 · Shelfport On Wednesday evening Olga messages her assistant: "I need a car seat for my son. He's two and weighs about 30 pounds, and the car has ISOFIX. We're driving to my mom's on Saturday, budget up to $250." The assistant asks a couple of questions and suggests three models. About one it says: in stock at a particular store for $229, delivered tomorrow. It also suggests adding a seat protector mat so the upholstery does not get dented. Olga taps the link and lands on the store's home page. She finds the seat through search: $249 and "ships in 5 to 7 days". She looks for the mat separately, finds three similar ones and cannot tell which one fits. Ten minutes later she buys a different seat from a different store. The first store never learned how close it came to a sale. In its analytics, Olga was one more AI referral visit with no order. ## The assistant already did the expensive part It found the shopper, understood the job, ruled out what would not work and chose your product. Stores usually pay for that with ad budgets and sales staff time. The order was lost in the last few steps. These shoppers arrive close to buying. Salesforce found that during the 2025 holiday season, shoppers coming from AI search converted nine times more often than those coming from social media. That makes losing them between recommendation and order all the more painful. There are usually four leaks. Here is each one and how to check for it. ## Leak 1. Price and stock do not match An assistant knows what you gave it, as of when you gave it. If the product feed goes out once a week and the price in your ERP changed on Tuesday, the assistant keeps quoting the old price until the next export. Stock works the same way: the last seat sold in the morning, and in the evening the assistant still promises next-day delivery. To the shopper this looks like a bait and switch, even when nobody meant it. A $20 gap destroys more trust than an honest price that is slightly higher. What to do. Prices and stock for assistants should come from the same source as your website, with no manual exports. OpenAI's documentation recommends sending the full feed once a day and pushing price and availability changes during the day as they happen. Flag backordered items clearly, with a lead time. How to check. Take twenty products from one category and, at a random moment, compare price and stock in your channel data with what your back office shows right now. ## Leak 2. The terms are unclear Olga needed the seat before Saturday. The assistant said "delivered tomorrow", but where did that come from? If delivery times depend on the region and the assistant has no data on them, it either says nothing or guesses. Both are bad. Returns work the same way. A car seat may not fit the car, and the shopper wants to know whether it can go back. If the answer is buried in your terms of sale, the assistant will paraphrase it loosely or say "check with the retailer". The shopper will not check. OpenAI's product feed specification has dedicated fields for this: shipping cost, whether returns are accepted and the return window in days. It states plainly that a missing shipping cost means the charge is unknown. The shopper sees uncertainty, and uncertainty works against you. What to do. Describe shipping and returns briefly, in a form you can pass along as data: regions, cost, delivery times, items that cannot be returned. Include restrictions too: oversized items, assembly, payment methods. ## Leak 3. There is no way to build a cart The assistant recommended a seat and a mat. That is already a two-item cart. Yet the link opens a single product page, without the color the shopper picked. The shopper has to redo what the assistant already did for them. To build a cart, an assistant needs specific product variants with identifiers and a way to pass them to your store with quantities. Then the link opens a cart that is already filled, and all the shopper has to do is review it. Shopify's developer documentation shows one way this works: the agent creates a cart and a checkout session, gets a link and hands it to the shopper. The link takes the person into the store's checkout with the chosen items already in place. How to check. Ask an assistant to put together an order of two or three products from your category, then see where the link goes and what is in the cart. ## Leak 4. Checkout breaks The shopper reaches checkout, and the usual problems start. Mandatory account creation. A cart that empties after login. Shipping cost that only appears on step three. The payment method they wanted is not there. These problems cost money in any funnel. A shopper sent by an assistant leaves faster: everything was just picked for them in a minute, so manual work at the last step feels even more irritating. In many channels checkout already happens on the store's own site. According to Shopify's documentation, for example, shoppers coming from ChatGPT complete their purchase in the merchant's online store. We cover how the market got here in [what already works in agentic commerce and what is still an experiment](https://opsalute.ai/en/blog/when-the-buyer-is-an-agent). Platforms also plan for steps an agent cannot complete. Shopify's documentation describes the case where checkout needs something that is not available through the API. The agent should then send the shopper to a link to finish the purchase in the store's checkout. What matters is that nothing the shopper chose or entered gets lost in that handoff. One more thing: a person confirms the purchase. The description of AP2, an open protocol for agent payments, says plainly that the user must always be in control. So at the confirmation step the shopper should see the final picture with no surprises: items, total including shipping, address and delivery date. What to do. Go through checkout on a phone the way a shopper from an assistant would: without logging in, with a new address. Remove anything that makes them start over. ## How to find where your AI referral traffic leaks Start with one category and one evening. For the full list of what to check, see our [checklist for getting a store ready for AI shopping agents](https://opsalute.ai/en/blog/prepare-store-for-ai-agents). 1. Ask assistants 10 to 15 questions your shoppers ask. Note where your store is named and where answers get price, stock or terms wrong. 2. Compare price and stock in your channel data with your back office right now. 3. Follow the path from the link in an assistant's answer to an order in your ERP. 4. Tag AI referral traffic in your analytics and see at which step those visitors drop off. 5. Fix the biggest leak first, then the next one. Even with everything fixed, the assistant decides what to show the shopper. Nobody can guarantee a recommendation. But once an assistant has picked your product, getting that shopper to an order is the store's job. ## How Shelfport closes the gap Shelfport is an Opsalute product that connects AI assistants to a store, currently in beta. The assistant gets prices and stock from the store's own systems. It can build a cart and hand the shopper over to the store's checkout. The order lands in the back office tagged as coming through an agent. During the beta we test this path on a store's own products, starting with one category. [See how Shelfport works](https://opsalute.ai/en/shelfport), or leave a request and we will walk through the purchase path in your store. ## Sources - [Salesforce: 2025 holiday shopping data](https://www.salesforce.com/news/stories/2025-holiday-shopping-data/) - [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) - [Shopify: Checkout MCP](https://shopify.dev/docs/agents/carts-and-checkout/checkout-mcp) - [Shopify: agentic storefronts](https://help.shopify.com/en/manual/online-sales-channels/agentic-storefronts) - [AP2: Agent Payments Protocol](https://ap2-protocol.org/)