OpenAI Killed Its Own In-Chat Checkout and eBay Banned Buy-for-Me Agents. Agentic Commerce Is Not the Land Grab You Were Told to Prepare For.

OpenAI launched Instant Checkout at the end of September 2025 and pulled it back in early March 2026. According to reporting at the time, roughly 30 Shopify merchants were participating as of February. The company that defined the category shut its own version of it after about five months.

In the same window, eBay posted a user agreement update naming and prohibiting buy-for-me agents. Amazon had already sued Perplexity over its Comet browser placing orders on Amazon accounts.

I bring this up in August, five months after the fact, because I keep seeing indie builders plan 2026 roadmaps around agent-ready checkout. The evidence has been in for a while and it points somewhere else.

Instant Checkout did not fail on demand. It failed on data.

This is the part worth understanding properly, because "users did not want it" and "it did not work" lead to very different conclusions.

The reporting on why it was pulled converges on a data problem. OpenAI was obtaining product information by scraping retail sites, which meant the catalogue inside ChatGPT could not reliably reflect live inventory and live pricing. If you have ever run a store, you know what that produces: confident recommendations for things that are out of stock, at prices that are no longer real.

Walmart's numbers, as reported, put ChatGPT checkout conversion at about a third of its own site's rate, with product data inaccuracy named as a major contributor. That is not a demand signal. That is a plumbing signal.

I want to flag my confidence level honestly here. Everything in this section comes from trade press and analyst commentary rather than from OpenAI or Walmart publishing the figures directly. The direction is consistent across multiple outlets. The precise conversion ratio, I would not stake anything on.

eBay's ban is the more interesting document

OpenAI retreating is a product decision. eBay changing its user agreement is a structural one, and it tells you what platforms with real transaction volume think about agents touching their checkout.

The update was posted in January and took effect in February. What makes it notable is the specificity. eBay already had broad anti-bot language. The revision reportedly calls out buy-for-me agents, LLM-driven bots, and end-to-end flows that place orders without human review, unless the tool has prior express permission from eBay.

That last clause is the whole thing. The door is not locked, it is gated. Agent access exists as a partnership you negotiate, not as a capability you build against a public interface. Which means the economics of building an agent that shops on eBay are the economics of business development, not engineering.

eBay's stated reasoning is about auction integrity, fraud and mistaken orders, and on an auction marketplace that is not a pretext. Bots sniping bids is an old and real problem there. But the same instinct shows up at Amazon in litigation form, and the common thread is that marketplaces do not want a third party sitting between them and the buyer, mediating the relationship and the data.

Where the industry actually landed

The phrase that captures the current state is "discover in AI, buy on your own site." People genuinely do use chat assistants to research purchases. They do not, at scale, complete the purchase inside the chat.

OpenAI's own stated direction after the retreat fits that shape: route purchases through third-party apps inside ChatGPT, with retailers keeping payment processing, rather than operating checkout itself. The Agentic Commerce Protocol work with payment companies continues. What changed is who holds the transaction.

If that sounds familiar, it is the same pattern as Stripe buying OpenRouter, which I wrote about yesterday. The interesting position is not the consumer-facing checkout. It is the layer underneath that everyone routes through, and the companies that already own payments are the ones assembling it.

For a solo operator, the takeaway is that the transaction layer is being decided by companies with balance sheets, and you are not going to win a seat at it. The discovery layer is different, and it is open.

What I would actually do

Treat AI assistants as a traffic source, not a sales channel. That reframing decides everything downstream.

Concretely, that means your product data needs to be legible to something that is reading rather than browsing. Structured product markup with real prices and real stock status. Specifications in text on the page, not baked into images or loaded by a script that a fetch will not run. A page that answers the question a buyer would ask, in words, rather than deferring everything to a tabbed interface.

None of this is new SEO advice, which is exactly why I trust it. It is the same discipline that made your pages parseable to search crawlers, applied to a reader that is summarising rather than ranking. The work compounds across both.

The thing I would not build right now is an integration that assumes an agent will complete a purchase on a marketplace you do not control. That capability is being actively withdrawn by the platforms that have the volume, and building against a gated interface means your product's existence depends on a partnership approval you have no leverage to obtain.

If you sell direct from your own site, the situation is genuinely better, because you control the checkout and nobody can ban you from it. The question there is simply whether AI-referred traffic converts, and the honest answer is that you should measure it rather than assume it. Tag the referrals you can identify and look at the numbers in ninety days.

Where I could be wrong

The obvious counter is that I am calling a race five months in. Instant Checkout failing on scraped product data is a solvable problem, and the fix is exactly what OpenAI says it is pursuing: merchant-supplied feeds through a protocol rather than scraping. If that lands well, in-chat purchasing could work fine in 2027 and this post will read as a snapshot of a trough rather than a verdict.

I also want to be careful about generalising from eBay. Auction marketplaces have unique reasons to fear automation that a normal retailer does not. A direct-to-consumer brand has no equivalent objection to an agent buying from it, and some will actively court that traffic. eBay's ban tells you about eBay, and about Amazon by analogy, and less than I implied about the general direction.

What I hold to with more confidence is the narrower claim: right now, today, betting a solo project on agent-completed checkout on a marketplace you do not own is a bad bet, and the discovery-side work is a good one because it pays off regardless of how the checkout question resolves.

Author

Lukas

@lukcombinator

Sources

• What It Means That The Leader In Agentic Commerce Just Pulled Back, Forrester
• What went wrong with ChatGPT's Instant Checkout, Modern Retail
• eBay Blocks Use of AI Buy-For-Me Agents, PYMNTS
• eBay explicitly bans AI buy for me agents in user agreement update, Value Added Resource
• Why OpenAI's Checkout Retreat Spells Trouble For Its Commerce Strategy, Forbes