ChatGPT’s retail apps solve the wrong problem

When OpenAI abandoned Instant Checkout in favour of retailer-branded apps, the official rationale was flexibility: merchants want control of the customer experience and the transaction process. Fair enough. But look at what the actual user journey now looks like.

You open ChatGPT. You describe what you’re looking for. You get a recommendation. If the relevant retailer has a ChatGPT app – and you somehow know that – you activate it, switch into that retailer’s environment, and are then directed to their website to complete the purchase, either in an in-app browser or a separate tab. If they don’t have an app, or you don’t know one exists, you follow a link out to their site anyway.

At every step, the user carries the cognitive load. Which retailer has an app? Why should I use this app and not directly shop on the retailer’s website? How do I find it? How do I trigger it? As PYMNTS flagged shortly after the announcement, shoppers simply “may not know that the retailers’ ChatGPT apps exist or how to access them.” That’s not a minor UX wrinkle. It’s the central problem of the model.

The friction argument for Apps goes like this: users have trusted checkout flows on their favourite retailer’s website. Their card is saved. Their loyalty points are there. Why would they want to pay inside a chatbot? It’s a fair point but it only works if the App model actually delivers less friction. Right now, it delivers more. You still end up on the retailer’s website. You’ve just taken a longer, more confusing route to get there.

The original promise of agentic commerce was elegant: an AI that finds the best product for your specific needs, across the entire merchant web, and handles the purchase. The AI is neutral. The outcome is optimised for the buyer.

The App model abandons this entirely. Instead of a neutral agent, you get a storefront selector. Walmart has Sparky. Target has its own chatbot. Sephora has a beauty advisor. Each one is a walled environment built to keep you inside that retailer’s ecosystem. This isn’t agentic commerce. It’s the 2011 app store model wearing a conversational UI as a costume, and the shopper still has to know where they want to shop before they begin.

The honest description of what ChatGPT has become is a referral engine. A smarter, more conversational Google Shopping. TD Cowen analysts called OpenAI’s reversal “a stunning admission,” noting that “the news signals that AI platforms replacing apps to become the ‘new OS’ is either not playing out, or at a minimum is pushed back significantly.” That’s not a pivot but a concession.

On March 24, the same day OpenAI killed Instant Checkout, Shopify flipped a switch and made all 5.6 million of its merchants discoverable inside ChatGPT by default – no setup, no integration, no app. That’s a genuine step toward democratising AI commerce, and it’s worth acknowledging.

But here’s the detail that gets glossed over in the optimistic coverage: nobody actually completes a purchase inside ChatGPT. Not Walmart, not Sephora, not Target. Sephora’s in-app checkout is “coming later.” Target redirects to its website. Walmart’s Sparky is a branded environment inside ChatGPT that processes payments through Walmart’s own infrastructure, closer to a mini-app than a checkout. The only retailer with true in-chat checkout is Instacart. One. Out of a hundred apps.

So what exactly is the “App model” delivering right now? A two-tier discovery system. Tier one: large retailers with the resources to build branded ChatGPT integrations – Walmart’s Sparky with cart sync and loyalty, Sephora with Beauty Insider data, Target with its own chatbot. Tier two: 5.6 million Shopify merchants surfaced through the catalog, with a product card and a redirect.

Both tiers bounce users out of the chat to complete a purchase. But the experience is not equivalent. A Walmart shopper inside Sparky has account continuity, synced carts across platforms, and loyalty points applied. A small Shopify brand gets a link. That’s the same retail hierarchy, with a conversational interface on top – and the promise of in-chat checkout perpetually deferred to a future release.

While OpenAI was pivoting, Google released real-time product data, multi-item carts, and loyalty membership integration into its AI shopping layer – the exact three features OpenAI cited as reasons to abandon native checkout. Google’s Shopping Graph holds over 50 billion product listings, refreshed two billion times per hour. Stale data was OpenAI’s problem because it was scraping. Google solved it at the infrastructure level and kept building.

The contrast is instructive. Both companies hit the same hard engineering wall. One built through it. The other reorganised around it and called it a strategy.

That said, Google has been here before. It launched “Buy on Google” native checkout in 2018 and quietly killed it in 2023 after failing to pull shoppers away from Amazon and retailer websites. Real-time data and multi-item carts are necessary conditions for agentic commerce. They are not sufficient ones. History suggests some scepticism is warranted.

There’s one more problem with “focusing on discovery” that hasn’t received enough attention: it creates an unanswered monetisation question. The transaction fee model was clean: OpenAI earned when users bought, so its incentives aligned directly with finding the best product. A discovery and referral model brings a different gravitational pull. Which merchants get prioritised in recommendations? Who pays for better placement in the feed? Only 16.7% of US AI users currently start product searches with an AI assistant – that number only grows if users trust the results are neutral. The moment sponsored placement enters the picture, that trust erodes. And sponsored placement always enters the picture eventually.

The App approach isn’t going to collapse visibly. Walmart will keep iterating on Sparky. Shopify will proudly point to 5.6 million merchants “discoverable in ChatGPT.” OpenAI will find a monetisation path of some kind.

But look at what’s actually been built. Nobody completes a purchase inside ChatGPT except through Instacart. Everyone else gets sent to a website. The promise of in-chat checkout has been deferred so many times it’s starting to look less like a roadmap item and more like a marketing position.

The genuinely ambitious idea – an AI indifferent to which retailer cut the better platform deal – has been quietly shelved. What replaced it is big brands with better placement, small merchants with a product card and a redirect, and a checkout that still happens somewhere else.