How selling on ChatGPT works: Instant Checkout, the Agentic Commerce Protocol, the feed and endpoint you expose, and what you give up in return.
The short answer
Selling on ChatGPT means the assistant completes a purchase inside the chat through Instant Checkout, which runs on the Agentic Commerce Protocol from OpenAI and Stripe. You expose a machine-readable product feed and a live checkout endpoint, you stay merchant of record, and you give up the site visit, the upsell and most session data.
Key takeaways
ChatGPT can take a shopper from question to paid order without sending them to your site. The mechanism is Instant Checkout, and it runs on an open spec OpenAI published with Stripe. For a merchant the question is practical: can you be in it, what do you have to build, and what do you lose.
This page is the merchant version of that answer. It covers what Instant Checkout is, the product feed and checkout endpoint you have to expose, who can switch it on today, and the parts of your funnel that disappear when the buying moment happens somewhere you do not own.
Selling on ChatGPT means a shopper completes the purchase inside the assistant instead of on your website. OpenAI calls this Instant Checkout. Your product data and your checkout system still do the work, but the buying moment happens in a chat window.
Three different things get described as selling on ChatGPT, and they have different requirements. Keeping them apart saves weeks of misdirected effort.
Most brands should start with the first one. An agent that cannot tell a 2-pack from a 6-pack, or that reads a stale price, will not recommend you and cannot transact you either. Shopify's guidance on agentic-ready product data puts it plainly: machine-parsable, structured and real-time are the bar, and Shopify reports AI-referred orders grew nearly 13x year on year.
The integration work described below is genuinely small compared with the data work behind it. Teams that sequence it the other way round end up shipping a clean API into a catalog that misrepresents what they sell. Our page on product feed optimization covers that groundwork, and the agentic commerce pillar covers the wider shift.
ChatGPT instant checkout works by letting the assistant build and complete a cart against your systems over an open API, then handing your processor a scoped payment credential to charge. OpenAI and Stripe published that API as the Agentic Commerce Protocol, and it is public, so any merchant or platform can implement it.
The flow has four moving parts. None of them replace your commerce backend.
You remain merchant of record throughout. The charge appears on your statement descriptor, the fraud exposure is yours, and returns and chargebacks come to your support team the way they always have.
The important design detail is that pricing stays server-side with you. The agent never invents a total, never applies a discount you did not authorize, and never holds card credentials. That is why the protocol asks for a live endpoint rather than a static file: a cached price is a broken order.
The endpoint surface itself is small. The Agentic Checkout Spec defines five calls on /checkout_sessions: create, update, complete, cancel and retrieve. Requests carry Idempotency-Key, Request-Id, Signature, Timestamp and API-Version headers, so retries and request signing are part of the contract rather than an afterthought. The spec is versioned, so read the current version at agenticcommerce.dev before you build against it.
You expose two surfaces: a readable product feed and a checkout endpoint. Everything else stays where it is. Here is what each one has to carry.
The endpoint accepts a session, prices it, updates it as the buyer changes address or quantity, and completes it. It has to be idempotent, because agents retry. It has to return machine-readable errors, because an out of stock response that arrives as HTML is a silent lost sale.
It also has to be honest about what it cannot do. If you cannot ship that SKU to that state, say so in the session response rather than accepting the order and cancelling it later. Cancellation after confirmation is the fastest way to be dropped from an agent's recommendation set.
On customer data the spec is specific. The buyer object carries a name, an email and an optional phone number, and the session carries a fulfillment address and an optional billing address. You never receive a stored payment credential, only a token good for that one transaction.
Treat those contact details as order data. Whether you can market to that buyer afterwards is a separate question of consent and of OpenAI's merchant terms, so check both before you add agent-sourced customers to any list.
Eligibility is gated by OpenAI and by your commerce platform, not by your willingness to build. Two routes exist: your platform ships the integration for you, or you implement the protocol directly.
The platform route is the realistic one for most merchants. Shopify merchants get the protocol work done at the platform layer, which means the integration reduces to enabling it and making sure your catalog data is good enough to be picked. That is the whole reason catalog quality dominates this project: the API part is somebody else's code.
The direct route is for merchants on custom stacks or on platforms that have not implemented ACP. You build the checkout session endpoints, handle the delegated payment credential with your processor, and maintain the feed yourself. It is a normal API integration, measured in engineering weeks rather than months, and the maintenance burden sits with whoever owns your pricing logic.
Two constraints are worth planning around before you pick a route. Agents are still weak on complex variants, meaning size plus pack count plus style in combination, so catalogs with deep variant trees see more failed matches. And some large retailers actively block agent traffic, which tells you the strategic question is not only can we, but do we want this channel on these terms.
Eligibility is OpenAI's to set, and it has kept moving since launch across countries, categories and platforms. Read OpenAI's current Instant Checkout documentation for where it is live today, including whether multi-item carts are supported on your surface, rather than planning against a snapshot.
On cost, OpenAI states that merchants pay a fee on completed purchases, that the service is free to the shopper, and that it does not change your prices or influence which products ChatGPT surfaces. The rate is not published. Treat it as a number you get from OpenAI or your platform rather than one you can model from public sources.
If you are on Shopify, the practical prep work is catalog-side and platform-specific. Our guide to Shopify agentic commerce covers variant grouping, taxonomy and live inventory sync in that context.
You give up the session. That is the honest summary, and it is worth pricing rather than arguing with.
The site visit disappears. No hero image, no reviews block, no trust badges, no merchandising sequence you spent two quarters testing. The agent shows a name, a price, an image and a delivery estimate, and the shopper decides on that.
The upsell disappears with it. Cart-level bundles, free shipping thresholds, post-purchase offers and subscription toggles all live in a checkout you no longer control. Your average order value on this channel will look different from your site, and comparing the two directly will mislead you.
Email capture gets thinner. Depending on what OpenAI passes through, you may receive a transactional contact and nothing that gives you permission to market to it. Treat every agent order as a customer you have not yet earned the right to email.
Session data goes too. No referrer chain, no on-site behavior, no scroll depth, no A/B assignment. Attribution for this channel is order-level, which means your analytics team needs a new bucket rather than a new dashboard.
What you gain is reach into a purchase that was not going to pass through your site at all. Productrise, studying 2M+ listings across 100k+ SERPs in the US and UK in August 2026, found only 1.28% product overlap between classic Google search and AI Mode, and that the main seller differed on 49.6% of matched products. That is a different shelf, not the same shelf rendered differently.
You also gain a forcing function. Every fix that makes you eligible for agent checkout, clean variants, literal attributes, live inventory, improves your feed for paid shopping and for every other shopping agent too. The work is not single-channel.
Ads in ChatGPT and Instant Checkout are two different programs with different owners, different buying motions and different requirements. Do not let one budget conversation cover both.
OpenAI sells ads inside ChatGPT through its own Ads Manager at ads.openai.com. Since 10 September 2026 there is a second door into the same surface, an Amazon Ads pilot letting select US advertisers extend campaigns into ChatGPT as conversational ads bought through Amazon DSP, with Delta Vacations among the first testers. Either way that is paid placement, bought the way you buy other inventory, with the assistant as the surface.
Instant Checkout is unpaid plumbing. It does not buy you visibility, it lets a shopper who already found you complete the order without leaving the chat. A merchant can be in one, both or neither.
The distinction matters for planning because the failure modes differ. Ads fail on creative and targeting. Organic recommendation fails on data quality. Checkout fails on endpoint reliability and stock accuracy. Three teams, three fixes.
There is a shared dependency worth naming. Both routes put your product in front of a shopper with almost no room for brand storytelling, which pushes the persuasion work earlier, into the creative that made them ask for your category in the first place. That is a video and creative problem, not a feed problem.
The other thing to watch is that assistant surfaces keep adding commerce layers. Meta launched Muse, a personal AI agent, on 8 September 2026, running in a private cloud VM with a visible browser and a separate Sentinel permissions system, on free, $20 and $100 monthly tiers. Instinct, an invite-only personal agent, pays via Stripe's Link wallet. None of these require you to build anything today, but they all reward the same underlying readiness.
Ads in ChatGPT have moved past the initial closed test. OpenAI began testing them in the US in February 2026 and has since opened self-service buying through its Ads Manager across 31 European markets, with eligibility still set per country and per vertical. The unit shown publicly is a sponsored product card placed below a response.
Before you ask whether you can sell on ChatGPT, find out whether an agent can currently describe your catalog correctly. This takes an afternoon and it is the only honest starting point.
The harder work is structural, and it is the part that decides whether you are competitive in agent-mediated buying rather than merely present.
Rebuilding your variant model is the big one. Most catalogs encode meaning in titles and images because humans read titles and look at images. Agents read fields, so the meaning has to move into fields, and that usually touches your PIM, your feed logic and your theme at once.
Taxonomy is the second. A product classified into a vague category competes against everything in it. Precise taxonomy plus literal attributes is what lets an agent shortlist you for a narrow request, which is where agent traffic concentrates.
Real-time inventory is the third, and it is an engineering commitment rather than a project. Every stale price or phantom stock record is a cancelled agent order, and cancelled orders are the strongest negative signal you can send a system that ranks on reliability.
The fourth is creative. When the product page stops doing the selling, demand has to be created before the shopper ever opens an assistant. That means more concepts, more formats and faster iteration, which is a capacity problem long before it is a taste problem.
Two of those four are things we do directly. On the catalog side we fix Shopify product data for agent readability: variant grouping, taxonomy, literal product fields and live price and stock sync. On the creative side we run performance creative as a service, 100+ creatives a month at a 5-7 day turnaround, with creative intelligence trained on 32M+ ads.
We cover the whole chain: intelligence, strategy, production and campaign management. Analytics-only tools tell you which ad worked and stop. Premium production shops make the asset and have no view of what to make next. Neither touches your product data.
If you want the diagnostic run for you rather than by you, the free ad account audit is the entry point.
Yes, through Instant Checkout, if your commerce platform or your engineering team has implemented the Agentic Commerce Protocol and OpenAI has enabled your account. The shopper pays inside the chat, you capture the payment on your own processor, and you stay merchant of record. Eligibility is still limited by country, platform and category, so confirm your status before planning around it.
Two things: a machine-readable product feed carrying stable IDs, variant grouping, literal attributes, real-time price and stock, and a live checkout endpoint that can open a session, price a cart with tax and shipping, and complete the order. If you are on Shopify, the endpoint work happens at the platform layer, which leaves catalog data quality as the real project.
You keep the order and the fulfillment relationship, since you remain merchant of record. You lose the site session: no landing page, no upsell, no on-site behavioral data, and thinner marketing permission than a checkout you control. Treat agent orders as a separate cohort in your analytics rather than blending them with direct traffic, because average order value and repeat behavior will differ.
No. Instant Checkout is an unpaid transaction flow that lets a shopper complete a purchase in chat. Advertising in ChatGPT is a paid placement, bought either from OpenAI itself through its own Ads Manager at ads.openai.com, with US testing from 9 February 2026, or through Amazon DSP, where an Amazon Ads pilot opened on 10 September 2026 lets select US advertisers extend campaigns into ChatGPT. You can use either, both or neither, and they are owned by different teams with different failure modes.
Test rather than assume. Productrise, studying 2M+ listings across 100k+ SERPs, found AI Mode showed prices 21.6% higher than classic search on matched products, with a median of $149 against $100. That is Google rather than ChatGPT, but it suggests assistant surfaces do not automatically favor the cheapest option, so low price alone is neither a qualification nor a disqualification.