What Actually Happens When ChatGPT Buys From Your Store: A Frame-by-Frame Walkthrough

Agent Commerce · Walkthrough

Open ChatGPT. Ask it to buy something from your store. Watch what happens.

If you have never run this test on your own brand, you are not alone. Most DTC operators have spent years optimizing a checkout flow for a human visitor (the cart-abandonment emails, the express-checkout buttons, the trust badges, the address autocomplete) and have never once watched an AI agent walk through that same flow on behalf of one of those humans.

The agent does not see the trust badges. The agent does not click the express-checkout button. The agent does not type into the address field. The agent runs an entirely different sequence of operations, and at each step something on your store either cooperates or breaks.

Agent commerce is the layer where an LLM-driven buyer agent completes a purchase on behalf of a human user. This post is the frame-by-frame walkthrough of that layer. The 7-Stage Agent Checkout is the standardized pipeline that agent walks: what it does at each stage, what breaks on the average DTC store at each one, and the operator fix per stage.

You optimized the checkout flow for a human who would never touch it.

Stage 1 and 2: Intent capture and discovery

Stage one is intent capture. A user types a purchase request in natural language: "find me a merino base layer for winter running, under $120, ships in three days." This stage happens entirely inside the agent. There is nothing on your store to fix here, and nothing to break. ChatGPT now serves 900 million weekly active users as of February 2026 (OpenAI), with roughly 50 million shopping queries per day flowing through it (OpenAI Economic Research). The volume of intent being captured this way is no longer a rounding error against search.

Stage two is discovery. The agent assembles a candidate set: it reads its training corpus, runs a live web search, and queries structured commerce data where it exists. Most stores pass this stage, because discovery is forgiving. The agent will find you if you are crawlable. What it finds, and how confidently it can act on what it finds, is decided downstream. Discovery is where you get into the room. The next five stages decide whether you leave with the sale.

Stage 3: Ranking and filtering, and why your brand never makes the list

This is the first stage where most stores quietly fail. The agent takes the candidate set from discovery and ranks it against the buyer's stated constraints (price ceiling, shipping window, category) and a set of inferred preferences it reads from the conversation and from the buyer's history. Then it filters. A candidate set of forty becomes a shortlist of three. If your product is not in those three, nothing else in this walkthrough matters, because the buyer never sees you.

Ranking is not a black box. It runs on signals you can audit: review density and recency, structured product data, third-party citations, and price-and-availability accuracy. We catalogued these in the 7 things that make AI agents recommend your brand. The brands that make the shortlist are the ones whose data answers the agent's filter questions without ambiguity. The brands that do not are usually missing two or three fields the agent treats as disqualifying, and they have no idea, because no dashboard reports a ranking they never entered.

The volume behind this stage is the reason it is worth fixing now rather than next year. Bain projects agentic AI will account for 25% of U.S. ecommerce sales by 2030. Gartner projects 90% of B2B purchases will be mediated by AI agents by 2028, representing over $15 trillion in spending. The share is growing faster than the number of brands optimizing for it, which is why brands now bid for the agent's attention through OpenAI Ads. Paid placement amplifies a strong ranking signal. It does not substitute for one.

Stage 4: Verification, the structured-data hop

Once you are on the shortlist, the agent verifies. Before it will recommend you with confidence, it confirms the things a careful human would check: is this a real product, is it in stock, is the price current, when will it arrive, can it be returned. It does this by reading structured data, not by reading your marketing copy. This is the structured-data hop, and it is where a brand that ranked well still loses the sale to a hedge in the agent's answer.

The fields live in your Schema.org Product and Offer markup, the structured-data standard the agent actually reads. Most DTC stores populate some of them and leave gaps in the rest. By various public structured-data audits the share of ecommerce product pages with complete, valid Product schema is well under half (figure illustrative, since audit methodologies differ). Shopify is a useful case: Shopify Payments handles settlement cleanly, but Shopify's default product schema is incomplete for agent verification, so the gap is often invisible to operators who assume the platform handles it. The technical baseline for being readable at all is covered in the visibility audit most brands skip; this stage is what the agent does with that readability once it has it.

Stage 5 and 6: Authorization and payment, the protocol layer

Stage five is authorization. The agent has a verified candidate. Now it either surfaces that candidate to the human for a yes, or it proceeds directly on stored preferences the user has already approved. Which path it takes depends on the trust the user has granted the agent and on whether your store speaks a protocol the agent can transact against. A store that can only be bought from by a human filling a form forces the agent back to a hand-off, and every hand-off is a place the purchase can stall.

Stage six is payment execution: the token exchange, 3-D Secure where the card network requires it, and the agent-payment handshake. This is the layer where the protocol standards live, and it is worth being precise about who does what. OpenAI launched in-chat Instant Checkout in 2025 and retired it in early 2026, moving completion back to the merchant side. Stripe is the standout here: it is the only payment provider with a working agent-commerce primitive in 2026, having shipped agent-commerce support with OpenAI that merchants enable through their dashboard. Apple Pay, by contrast, has no agent flow: the agent falls back to a card on file, which reintroduces the friction the protocol layer exists to remove.

The protocols themselves (ACP, AP2, and the commerce extensions to MCP) are easy to conflate and easy to over-invest in. The operator-relevant distinction is short: one of them asks a DTC team for real work this quarter and the rest are calendar items. We laid out the full taxonomy in the protocol map. For this walkthrough, the point is narrower. Authorization and payment are the stages where the protocol layer decides whether the agent can finish the purchase on its own or has to punt back to a human, and the punt is where conversion leaks.

Stage 7: Post-purchase context, the memory that compounds against you

The purchase completes. Most operators think the flow ends here. The agent does not. Stage seven is post-purchase context retention: the agent stores what happened (which brand, what product, how the fulfillment went, whether the buyer was satisfied) and carries it into future sessions. The next time the same user asks for something in your category, the agent does not start from zero. It starts from memory.

This is the stage that compounds, in your favor or against you. A clean purchase with accurate fulfillment writes a positive entry. A purchase where the price was stale, the item arrived late against a shipping claim the feed got wrong, or the return was harder than the policy implied writes a negative one. That entry shapes the next ranking. When we analyzed what AI agents actually recommend when 10,000 buyers ask, the pattern was consistent: agents reward the brands whose past interactions resolved cleanly and route around the ones that did not. AI-driven traffic to U.S. retail sites grew 393% year-over-year in Q1 2026 (Adobe Analytics). ChatGPT now drives 20% of Walmart's referral traffic and over 20% of Etsy's; Amazon, having blocked ChatGPT crawlers in robots.txt, receives under 3% and declining 18% month-over-month. The memory is the part of the checkout you cannot optimize after the fact. You earn it at stages four through six or you do not.

The operator's 7-stage audit

You can run the whole walkthrough against your own store in about thirty minutes, one check per stage, no engineering required. The point is not to fix everything today. The point is to find out which of the seven stages your store is currently failing, because most operators discover they are losing the purchase at a stage they have never looked at. Public audits suggest the average ecommerce store carries several structured-data errors per product page (illustrative; exact counts depend on the scanner), and you will not know yours until you check.

The agent walks through your checkout every day, and you have never watched it do it. That is the uncomfortable part: not knowing each stage means forfeiting control of each stage. The 7-Stage Agent Checkout is not a future you are preparing for. It is a flow that is running against your store right now, deciding at ranking whether you make the list, at verification whether you earn the recommendation, and at memory whether the next buyer ever hears your name. By 2028 a meaningful share of your revenue will pass through these seven stages whether you instrumented them or not. If you want Cresva to instrument them on your own store and show you exactly where the agent leaves, request early access.