Shopping in ChatGPT: what OpenAI's commerce push means for brands

In early 2026, ChatGPT stopped being a place you research a purchase and started being a place you make one. OpenAI added checkout inside the chat, then reshaped the approach around brand-owned apps. For anyone selling a product, this collapses the distance between "the AI mentioned us" and "the customer bought" - and makes being recommended more valuable than ever.

The mechanics have shifted more than once this year, which is itself the point: AI commerce is being built in real time. Here is where it stands and what it changes for your brand.

From "Buy it in ChatGPT" to brand apps

In February 2026, OpenAI launched Instant Checkout - "Buy it in ChatGPT" - letting users complete a purchase without leaving the conversation. It started with Etsy sellers and extended toward Shopify merchants, powered by an open Agentic Commerce Protocol built with Stripe. The pitch was frictionless: ask, get a recommendation, buy, all in one thread.

By mid-2026, OpenAI adjusted course, steering commerce toward brand-owned apps inside ChatGPT that route the buyer to the retailer's own experience rather than completing checkout entirely in-chat. The reason is control - merchants want ownership of the transaction, the data and the customer relationship. The direction of travel is clear either way: ChatGPT is becoming a storefront as well as an advisor.

Apps are a new visibility surface

Alongside checkout, OpenAI opened ChatGPT to third-party apps and launched an in-chat app directory, with early partners spanning travel, design and media. That directory is a new place to be found - and, candidly, a crowded and still-immature one. Early reporting suggests many integrations are hard to discover and have driven limited traffic so far. It is a surface to watch and experiment with, not yet a channel to bet the quarter on.

The honest read: apps in ChatGPT are an emerging opportunity with real uncertainty. Worth a considered presence if it fits your product; not worth abandoning the fundamentals of being recommended in the first place.

Why recommendation now equals revenue

Here is the shift that matters most. In classic search, being recommended earned a click, and the sale happened later, elsewhere, after more steps. When the assistant both recommends the product and enables the purchase, being named in the answer sits one tap from the transaction. The funnel compresses to almost nothing.

That cuts both ways. If the model surfaces you for a buying question, you are extraordinarily well placed. If it doesn't, you are not merely losing a visit - you are losing the sale, because the entire journey now happens inside the assistant and never reaches your site at all. Absence from the recommendation is absence from the purchase.

"When the assistant recommends the product and enables the purchase, being named sits one tap from the transaction."

What brands should do

1. Get your product data right everywhere. Accurate, structured, consistent product information - on your site, on marketplaces, in reviews - is what lets an assistant describe and trust your offer well enough to recommend it.
2. Earn corroboration for your products. The reviews and community discussion AI leans on decide which product it reaches for. Being well-reviewed in the right places is now merchandising.
3. Watch the commerce surfaces, invest carefully. Explore Instant Checkout eligibility and a ChatGPT app if they fit, but treat them as experiments with clear eyes about their current maturity.
4. Measure whether AI recommends you for buying questions. Not just "does AI know us," but "when a buyer asks the assistant what to buy, are we named - and if not, who is?" That is the metric that now maps to revenue.

How agentic commerce actually works under the hood

The plumbing behind "Buy it in ChatGPT" is the Agentic Commerce Protocol (ACP), an open standard OpenAI and Stripe published together and released under an Apache 2.0 license so any agent and any compatible payment provider can implement it. It is worth understanding the flow, because it tells you exactly where your brand has to show up to be part of the transaction.

In plain terms, ACP breaks a purchase into three moves:

1. Discovery. The assistant reads a structured product feed you supply and decides your item is a good answer to a buying question.
2. Checkout. The buyer confirms in the conversation, and the agent sends a checkout request to your systems on the buyer's behalf.
3. Payment. The agent passes a secure delegated payment credential (a scoped token, not raw card details), and you - the merchant of record - accept or decline using your own fraud and payment signals.

The important detail for brands: you remain the merchant of record. You own the order, the fraud decision and the customer data, even when the buyer never touches your site. That is the same instinct driving the mid-2026 pivot toward brand-owned apps that deep-link buyers into the retailer's own checkout. Whether the sale closes in-chat or on your storefront, discovery still happens inside the assistant, and discovery still depends on the feed.

The product feed is the new shelf

An assistant cannot recommend what it cannot describe with confidence. For ChatGPT that description comes from a structured product feed you submit directly, not from the model guessing off your marketing pages. If you sell on Shopify, this is close to a switch you flip: enabling the ChatGPT sales channel from the Shopify admin wires up the feed with no code. Everyone else builds an ACP-compliant feed and delivers it to OpenAI, typically as a full file once a day with lighter updates layered on throughout the day (the spec reportedly accepts refreshes as often as every 15 minutes).

Treat the feed like a physical shelf: what is missing does not get picked up. The fields that decide whether you get shortlisted include:

• Titles and long-form descriptions that read like a human wrote them, not keyword soup - the model uses these to match your product to a specific need.
• Accurate price and real-time inventory, because an assistant will not confidently recommend something it thinks might be out of stock or mispriced.
• Stable identifiers (GTINs, SKUs) so your product can be matched and corroborated across marketplaces and review sites.
• High-quality images and clear shipping and return details, which reduce the assistant's uncertainty about surfacing you.
• Structured review data, where available, so the model has a signal for quality and not just existence.

Feeds also carry control flags worth knowing: reporting indicates an item's searchability and its checkout eligibility are set separately, and that turning on in-chat purchase requires the product to be searchable in the first place. In other words, being buyable is downstream of being findable. Get the feed clean before you chase the checkout button.

Corroboration decides which product the model reaches for

A clean feed gets you into consideration. What tips the assistant toward you over a near-identical competitor is corroboration - the independent signals that say this product is real, good and the right fit. Models cross-check what your feed claims against reviews, community threads, expert roundups and retailer ratings. When those agree with your own description, confidence goes up and you get named. When they are thin or contradictory, the model hedges and reaches for a rival it trusts more.

This reframes reviews from a nice-to-have into merchandising. A few concrete moves:

• Make sure your best products are reviewed in the places the model already reads - major marketplaces, established review platforms, category-specific communities - not only on your own site.
• Keep specs consistent everywhere. If your feed says one thing and a marketplace listing says another, you are handing the model a reason to doubt you.
• Earn genuine third-party mentions for the specific product, not just the brand. "This brand is well known" helps less than "this exact model is the one reviewers recommend for X."

The discoverability reality-check

Stay clear-eyed about the current state. Reporting through the first half of 2026 suggests in-chat checkout saw limited real-world adoption - by some accounts only a few dozen Shopify merchants ever went fully live - which is part of why OpenAI reportedly leaned back toward a feed-first, deep-link model. The in-chat app directory carries the same caution: early integrations have been reported as hard to discover and light on traffic so far. None of this is a reason to ignore the surface, but it is a reason not to bet the quarter on it.

The durable takeaway from the churn is that the volatile part is the checkout mechanism, while the stable part is the product data and corroboration underneath it. Whichever way OpenAI routes the final transaction, the assistant still has to find you and trust you first. Invest there, and you are covered no matter how the plumbing shifts again.

A practical checklist for e-commerce brands

If you want a concrete order of operations rather than a philosophy, work down this list:

1. Ship a clean feed. On Shopify, enable the ChatGPT sales channel. Off Shopify, build an ACP-compliant feed and deliver it on the recommended cadence. Fix titles, descriptions, pricing, inventory, identifiers and images before anything else.
2. Audit consistency. Reconcile the specs, prices and product names across your site, your feed and every marketplace you sell on.
3. Seed corroboration. Concentrate review and community presence for your hero products in the sources models actually read.
4. Decide on checkout deliberately. Check Instant Checkout eligibility and consider a ChatGPT app only where it genuinely fits your catalog and margins. Treat it as an experiment, not a core channel yet.
5. Instrument recommendation. Test the real buying questions in your category and record whether the assistant names you - and who it names when it does not.

How this ties back to measurement

All of this is unmanageable if you cannot see it, which is where measurement becomes the connective tissue. The old question - "does AI know we exist?" - is no longer enough when the answer sits one tap from a sale. The question that maps to revenue is sharper: for the specific buying prompts in your category, does the assistant recommend you, at what rank, and against which competitors?

That is a monitorable thing. Track your presence and position across the buying questions that matter, watch how it moves as you improve your feed and corroboration, and treat every prompt where a rival is named instead of you as a concrete gap to close. When the assistant closes the sale, your visibility inside its answers is your funnel - so the recommendation is the metric worth watching, and the feed and corroboration are the levers that move it.

The takeaway

OpenAI's 2026 commerce push - checkout in the chat, then brand-owned apps - turns being recommended by AI into being bought through AI. The plumbing will keep changing; the strategic fact won't. In a world where the assistant closes the sale, the brands that win are the ones it can describe accurately, trusts enough to recommend, and names for the questions that end in a purchase. Everything else is downstream of that.

Related: How to rank in ChatGPT - a practical guide to getting named →