An AI shopping agent gets a request: "Find a black running shoe in my size that can arrive by Friday." A store page may look excellent to a person and still leave the agent unable to answer. A photograph does not confirm stock. A price on a listing does not include shipping. A checkout button does not explain what happens when the last pair sells.
For agents to work, the store needs a dependable path from discovery to a verified offer. The aim is not to replace the customer-facing site. It is to make the store's facts and actions clear enough that an assistant can help a buyer without inventing details.
Give every variant a clear identity
A product record should name the exact variant: product ID, size, color, price, currency, availability, seller and a stable link. If a shoe comes in several sizes, "in stock" at product level is not enough. The agent must check whether the requested size is available now.
Show when the data was last updated. A feed helps an agent find options, but the store's live inventory should decide whether an item can be bought. Return a plain out-of-stock response instead of silently substituting a nearby variant.
Return the full offer before checkout
For the chosen item and destination, return the current price, tax, delivery fee, estimated arrival and applicable return terms. If the delivery address changes the quote, say so. The buyer should see the final total before making a commitment.
The assistant can compare shoes, but "find me shoes" is not permission to buy them. The purchase must remain a separate buyer decision. A secure checkout handoff is better than asking an agent to copy a card number into a chat.
Make each action report a real outcome
A useful flow has distinct operations to search, quote, create a cart, start checkout and confirm an order. Each operation should return an ID and a clear success or failure state. A cart is not an order. An initiated checkout is not a paid order. The agent should tell the buyer what actually happened.
The store should also let the buyer or support team trace and correct a failed attempt. A timeout after checkout needs a status check, not a second charge by guesswork.
Test an agent with a concrete task
Try this: "Find two black running shoes in size 9 below my budget, deliverable to my postcode by Thursday. Do not buy anything." Compare the agent's answers to the store's live data. Did it get the size right? Did it include shipping? Did it promise an impossible arrival date? Did it stop before checkout?
Fix the first failure, then run the same task again. It is a better test than counting page impressions alone.
The wider shift is still a forecast. Search ads and store visits will not vanish overnight. But if buyers increasingly ask assistants to do the first pass, stores with accurate records, safe actions and clear checkout handoffs will be easier to recommend. Google's Merchant API checkout settings and OpenAI's commerce documentation show parts of this path emerging, not a universal standard every store can use today.
The checkout path has also changed. In its March 24, 2026 update, OpenAI said the initial version of Instant Checkout did not offer enough flexibility. It is letting merchants use their own checkout experiences while focusing on product discovery. The earlier Instant Checkout launch is historical context, not evidence that buyers can generally complete a purchase inside ChatGPT. A store still needs an accurate offer and a safe handoff to its current checkout.
Sources
Google Merchant API checkout settings, OpenAI Agentic Checkout Spec, OpenAI's September 2025 Instant Checkout launch. Update: OpenAI's March 24, 2026 product-discovery and merchant-checkout update.