Selling Into ChatGPT After Instant Checkout

Chapter 1Can You Still Sell Inside ChatGPT?

Not through Instant Checkout. OpenAI discontinued it in early March 2026, reported as 5 March, roughly five months after launching it on 29 September 2025. If your agent-commerce plan says integrate Instant Checkout, that plan is out of date, and a lot of published advice still says exactly that.

What has not changed is the part that was always doing the work. ChatGPT still recommends products, and being named in that recommendation is still the thing that decides whether you are in the consideration set. The purchase moving back to your own site does not reduce that; it makes the recommendation the entire mechanism rather than half of it.

Chapter 2What Instant Checkout Was, and What It Cost

Instant Checkout let a shopper complete a purchase inside ChatGPT rather than clicking through to the merchant. It was built on the Agentic Commerce Protocol, published by Stripe and OpenAI under Apache 2.0, and that protocol still exists independently of the surface being withdrawn.

That rate is worth remembering rather than forgetting, because it is the number any successor surface will be compared against. It is also not settled: the first version of this channel was withdrawn inside six months of launch, so treat a published take rate as a starting position rather than a market price.

Chapter 3Why Did It Fail?

Because it did not convert, and because shoppers who had used an assistant to research a purchase mostly still wanted to complete it on the merchant's own site. The reporting on this is consistent even where the details differ.

Every figure in that table is trade coverage rather than a first-party disclosure, and each cell says so rather than relying on a column heading. What the pair of Walmart numbers shows is the important part and it does not depend on the exact values: the assistant was good at introducing the brand and bad at closing inside the chat. Those pull in opposite directions, and the correct response is to keep the half that worked.

Chapter 4What Replaces It

OpenAI's stated direction after the withdrawal was to route purchases through third-party apps inside ChatGPT rather than to operate checkout itself, with retailers keeping payment processing. That is a real difference in posture: it moves OpenAI from being the merchant of record's counterparty to being a surface that hands off.

The wider landscape did not stop moving when this one surface did. Several other protocols cover adjacent parts of the same problem, and the protocol map sets out who governs each and which are actually in production.

Chapter 5What to Build Instead

Everything below works whether or not any in-chat checkout returns, which is the test a plan should pass in a market that just withdrew its flagship surface inside six months.

1. 1

   Make sure an assistant can read you at all

   This is first because it gates everything else. Six of the 72 DTC brands in our public registry answer every request with an error rather than a storefront, from a crawler user agent and a desktop browser user agent alike. Nothing in their own analytics shows it, because the request never reaches their origin. Registry figures read from our checked-in brand list on 3 August 2026.

2. 2

   Separate the crawlers rather than blocking them as a category

   Blocking the training crawler is a defensible policy. A blanket AI-bot rule also catches the search crawler, which removes you from ChatGPT recommendations entirely and cannot be bought back with ad spend.

3. 3

   Make the landing experience survive an assistant referral

   A shopper arriving from a recommendation has already done the comparison. They land with intent and no patience for a homepage. Send them to the product, not the front door, and make sure price and availability on that page match what the assistant was told.

4. 4

   Instrument the referral you cannot see

   Assistants frequently send no referrer, so these sessions land in Direct and branded organic. A one-question post-purchase survey joined to order IDs is the cheapest way to size the gap on your own data rather than borrowing a benchmark.

5. 5

   Measure whether you are named, before optimising anything

   Unbranded, purchase-intent questions from your own catalogue, asked repeatedly, with failures recorded as failures rather than as absences. If you are not in the answer, checkout mechanics were never your constraint.

The detail behind steps one and two is in the OpenAI crawler reference, step four is the survey protocol in tracking the dark funnel, and step five is measuring AI visibility without fooling yourself.