ChatGPT's app store is six months old. Should you build there yet?

OpenAI launched apps inside ChatGPT last October with seven pilot partners: Booking.com, Canva, Coursera, Expedia, Figma, Spotify, and Zillow. In December it opened submissions and shipped an app directory. Six months and roughly 300 apps later, ChatGPT has most of the parts of an app store, attached to 900 million weekly users.

The pitch writes itself. Someone is planning a trip in a chat, the model pulls your booking flow into the conversation, and the user never sees an install screen or a signup form. We wrote about this exact mechanic when it was Telegram doing it. The disappearing funnel is the story, again.

But six months of real usage gives us something the launch coverage couldn't: evidence. And the evidence is mixed in ways that should change when you build, more than whether.

What a ChatGPT app actually is

Not an app in the iOS sense. A ChatGPT app is an MCP server plus a set of UI components the Apps SDK renders inside the conversation. MCP (Model Context Protocol) is the open standard Anthropic published in 2024, and it's the part worth paying attention to: you expose tools, the model decides when to call them, and your components draw the result as a card, a map, a player. There is a small irony in OpenAI's store running on its rival's protocol, but for you it's good news, and we'll get to why.

The practical consequence: most of the build is your API. If your product already has a clean, documented one, wrapping it in an MCP server and a few components is weeks of work, not months. If your "API" is a pile of internal endpoints nobody wrote down, that's the real project, and the ChatGPT part is the cheap bit at the end. Submission itself wants the boring things platforms always want: a privacy policy, terms, a verified domain, screenshots, and a review against guidelines that require your app to do something ChatGPT can't already do by itself.

The pilot apps show what works in a chat: Zillow draws listings on an interactive map, Canva turns a described idea into an editable design, Spotify assembles the playlist you asked for. Interactive results the user can act on, not a mini website squeezed into a bubble.

Distribution is the model, not a shelf

The directory exists, but nobody browses their way to your app the way they scroll a store category on a Sunday. Real distribution happens when the model decides, mid-conversation, that your app is relevant to what the user is trying to do. OpenAI says it weighs conversational context, past app usage, and user preferences, and admits it is still experimenting with how surfacing works.

That cuts both ways. You can't buy placement, and there's no keyword grid to grind, so most of the app store optimization playbook simply doesn't transfer. On the other hand, the model's habits are forming right now, and being the app it reaches for in your category is a position that could compound quietly for years. Adobe found that around 47% of people who use AI assistants for search are at least partly asking for product and app recommendations. Those recommendations have to come from somewhere.

Now the cold water. Reporting this spring found partner companies getting little traffic from their ChatGPT apps, users struggling to find them at all, approval queues moving slowly, and developers receiving almost no data about how their apps perform. Some partners deliberately stripped features out of their apps to keep the customer relationship, and the payment, on their own turf. Others found users hesitant to hand card details to an AI assistant. That doesn't kill the thesis, but it does mean the channel is early, and anyone selling you guaranteed ChatGPT traffic in 2026 is guessing.

The money part is half-built

Monetization is where the store metaphor breaks down, at least as of July 2026. The path that's generally available is external checkout: the user finishes the purchase on your domain, and OpenAI currently approves that for physical goods. An in-conversation payment sheet exists in beta for a handful of marketplace partners, built on the Agentic Commerce Protocol that OpenAI developed with Stripe. Broader monetization tools, including revenue for developers, are promised for later this year.

If you sell software subscriptions, notice what's missing: there is no sanctioned way to charge for anything digital inside ChatGPT today. For SaaS products, an app today is a way to win users and keep them close; the actual revenue still happens on your own site. That may be perfectly fine. It's just worth knowing before you scope the project, because it changes what "success" should mean in the first year.

Build now, wait, or hedge

Build now if your product is API-shaped, your buyers already live in ChatGPT (at 900 million weekly users, that covers most consumer categories), and being early where the model forms its habits matters more to you than clean attribution this quarter. Physical goods commerce has the clearest case, because checkout actually works there.

Wait if your moat is the experience you own: a checkout tuned over years, a UX that is the product, margins that can't absorb an intermediary later. Wait, too, if you need analytics to justify the spend internally (you will barely get any), or if your compliance team will ask exactly where user data flows in an OpenAI-mediated conversation and you'd like to give them a confident answer.

The hedge is the option we find ourselves recommending most, and it works because of that irony from earlier: the SDK sits on an open protocol. The MCP server you'd build for ChatGPT is the same server Claude and a growing list of agent platforms can call. Build the server, treat OpenAI's components as one client of it, and you've made your product agent-ready instead of OpenAI-dependent. The question stops being "should we bet on OpenAI's store" and becomes "should agents be able to use our product at all". For most products, that second question answers itself.

We went through a version of this with Telegram: a platform opens a surface, early builders get outsized attention, then the rules tighten and the tax arrives. Expect the same arc here. The difference is the ticket price. If your API is in order, finding out costs a few weeks, and the plumbing is reusable across every agent platform that matters. Just don't confuse a directory listing with distribution. The model decides that part, and it hasn't made up its mind yet.