Understanding OpenAI's commerce retreat
The gap between search and buy
Note: I think there’s enough context now in the essays, insights and predictions for me to now start analysing news. Here goes:
OpenAI has scaled back its plan to introduce shopping directly inside ChatGPT, reports The Information. Instead of allowing users to make purchases from product listings in ChatGPT search results, it is now focusing on checkouts inside specific apps that plug into ChatGPT. The reasons cited: only a small number of merchants were selling through the checkout; users were researching products inside ChatGPT but not using it to actually make purchases.
Six months ago, OpenAI had announced this as a major business opportunity. It partnered with Shopify, Etsy, and Stripe, and said millions of merchants would soon be available for purchase inside ChatGPT. The actual number that went live: roughly a dozen of Shopify’s millions of merchants. OpenAI had to work hands-on with each. It had not set up systems to collect and remit state sales taxes. The user behavior it needed hadn’t appeared.
Shopify told merchants that ChatGPT in Agentic Storefronts will launch later in March, with buyers completing purchases on the merchant’s own storefront rather than inside ChatGPT.
I had expected these problems, among others:
1. The commission made merchant sign-up unlikely from the start:
The Information reported that OpenAI planned to charge merchants up to 4% for completed purchases via Instant Checkout (roughly 7% including card fees and taxes). I had asked in Why commerce isn’t ready:
“Why will merchants sign up for an expensive sale when free alternatives exist?”
Shopify President Harvey Finkelstein confirmed in their February earnings call that for Shopify merchants in the new model:
“The economics are the same as the transaction happened on the online store when it comes to agentic. Specifically on something like ChatGPT, which requires Shopify payments, monetization is through payments.”
The leverage needed to impose a 7% cut requires market dominance. OpenAI doesn’t have that in ecommerce. My read at the time was that “Google may not charge for external checkout just yet, because it knows the platform game, and it can play the long game while OpenAI can’t.” Google’s UCP has 60 partners including Shopify, Etsy, Target, Walmart, Mastercard, Visa, and Stripe.
2. ChatGPT’s app interface degrades the shopping experience
In The Opportunity Trap of the ChatGPT App Store, I wrote:
“If you have a ChatGPT App, you’re not a service provider for the customer: you’re a service provider for ChatGPT. ChatGPT apps are tools, not destinations.”
Testing the Booking.com app inside ChatGPT, I found that even after connecting the app, ChatGPT prioritised generic web results over Booking’s contextual results, and recommendations ignored key filters. The experience and interface inside ChatGPT for apps and commerce is terrible. Users will move to a better experience via ChatGPT, rather than remaining on the platform, especially when it comes to something that involves spending money.
3. It’s clear that merchants aren’t ready for integration into chat apps yet
The merchant-side structural problem is deeper than it appears. The first digitisation made products visible to humans: listings, photos, descriptions, reviews. Agents need something qualitatively different. The Information reported in January:
“ChatGPT has to interpret information like pricing and in-stock availability that is often ambiguous and spread out across multiple systems… If the agent gathers information incorrectly, it might charge the wrong price or place orders for something that’s out of stock.”
For an agent to buy correctly, it must identify the product unambiguously, reconcile attributes described differently across systems, verify whether “in stock” means what it appears to mean, confirm whether the listed price includes taxes and shipping, know when a transaction is complete, and determine who is responsible if something goes wrong. As I wrote in Why commerce isn’t ready for AI yet: “Agents don’t simplify commerce: they force it to be explicit.”
Siddharth Puri, founder of Tyroo, went through Google’s Universal Commerce Protocol and told me that there’s a practical constraint:
“My challenge is they are still trying to set input standard for brands to bring in data - largely such initiatives fail - other than electronics/consumer durables type categories. It’s tough to achieve in fashion, beauty, food.”
A dozen merchants live, out of millions, is a data readiness problem, not a product problem. Finkelstein noted at the February earnings call that orders from AI searches are “up about 15x. Now that’s obviously on a very small base, and it’s still early days.”
Reasoned is where I write about how AI is changing the world, whether its Commerce, Social Media, Content, Classifieds, Payments or even war. I publish twice (sometimes thrice) a week.
Do consider subscribing.
4. Users research in ChatGPT, but buying is a different decision
OpenAI found that users researched products inside ChatGPT but didn’t buy. In What happens when AI buys or sells for you, I wrote that “Content is easy. Money is where people start thinking about what could go wrong, above all else.” Commerce businesses spend a lot of time and effort on creating an environment where buyers feel comfortable, and a space that they’re comfortable with. Buying in chat is a whole new experience, and chat hallucinations don’t necessarily help with confidence. Research is low-stakes. Purchase requires trust in delivery, returns, transaction state, and the entity holding your payment. ChatGPT had not earned that trust before moving to checkout.
5. Shopify’s silence on ACP is itself a signal
OpenAI and Stripe are continuing to develop the Agentic Commerce Protocol (ACP). Shopify also co-developed the Universal Commerce Protocol (UCP) with Google, a separate standard. At Shopify’s February earnings call, Morgan Stanley analyst Keith Weiss asked Finkelstein directly whether UCP and ACP were competing or complementary, raising the VHS vs Betamax concern. Finkelstein answered entirely on UCP:
“The goal is simple with UCP. It’s one common language for agents and retailers. The idea is that merchants can keep the brand, the attributions buyers get these incredibly trustworthy experiences and agentic commerce can scale. UCP is specifically geared towards being a protocol that covers the full commerce journey end-to-end, from search to cart then checkout, it includes post order.”
He didn’t address ACP, and that’s a signal. I wrote in Why commerce isn’t ready that “standards succeed only when economic incentives precede compliance.” Google has more merchants, more existing commerce infrastructure, and AI Mode in Search as a forcing mechanism for adoption. Yes, forcing, but it’s there.
6. Shopify won’t build a dedicated ChatGPT app, and that’s a signal
Shopify already has a consumer discovery app — the Shop app — for users to discover products from Shopify merchants. It clearly invests in discovery surfaces when it sees a viable commerce audience. The Information reported that Shopify has no plans to build a dedicated ChatGPT app. Either it doesn’t want to cede its own discovery to ChatGPT, and fall into the “Opportunity Trap”, or its waiting to see how this plays out, while still monetizing transactions flowing to Shopify merchants via ChatGPT. Finkelstein confirmed at the February earnings call that the transaction still flows through Shopify regardless:
“LLMs do not bypass Shopify’s checkout... OpenAI will run the front end... Shopify still runs the back end.”
Shopify captures the transaction wherever discovery happens. It just doesn’t think discovery is happening inside ChatGPT yet.
7. Amazon’s investment announcement is also silent on OpenAI’s commerce plans
Amazon also recently announced a $15 billion (of a planned $50 billion) investment in OpenAI. The announcement made no mention of Amazon selling within ChatGPT. Amazon had already locked its site down against AI apps including ChatGPT, and has got an injunction against Perplexity’s bots (court order). Amazon CEO Andy Jassy has said he would be open to working with outside AI shopping tools if terms were attractive. OpenAI’s largest new investor still doesn’t find those terms attractive enough to be present inside the product it just backed.
7. Will it be agents vs platforms?
While it’s too early to call a winner here, we also need to look at the agents versus platforms situation. Platforms like ChatGPT intermediate the supply side of commerce: merchants. Agents are on the demand-side of commerce, representing users who wish to make a purchase. In What happens when AI buys or sells for you, I wrote that buying agents will matter most “when outputs are standardised and predictable”, and they serve users better,
“When buyers have infinite time, infinite comparison, and zero fatigue, price stops being negotiated and starts being engineered.”
Agentic commerce has its own constraints, though: First, there’s invisibility. You can’t tell whether the agent is going to the right website and picking up the right product. Second, the absence of friction in the buying process:
if it’s too easy, it means there is too much risk. Friction here wasn’t inefficiency. It was a safety mechanism. Boundary conditions are essential.”
*
Shopify’s own agentic commerce integration is not currently open to all merchants. Asked at an investor conference why access remains limited, Harley Finkelstein said “the only reason it’s gated is we’re just waiting for the agent applications to continue to open the doors.” That is a generous framing. The gating is infrastructure that wasn’t ready, user behavior that wasn’t there, and data readiness that was never going to resolve across millions of merchants in six months.
*
This piece draws on four earlier Reasoned essays: