In September 2025 OpenAI shipped Instant Checkout in ChatGPT — the ability to buy a product directly inside the chat interface without ever leaving the conversation. Stripe powered the payments. The Agentic Commerce Protocol (ACP) was the underlying standard. Etsy went live first, with 1M+ Shopify merchants — Glossier, SKIMS, Spanx, Vuori — queued up to follow. It was framed as the beginning of agentic commerce: a future where you would never have to visit a brand's website at all.
By April 2026, OpenAI started walking it back.
The official explanation, paraphrased from OpenAI's own communications: the initial version of Instant Checkout did not offer the level of flexibility merchants needed. Different retailers wanted different upsell flows, different shipping options, different loyalty integrations, different post-purchase experiences. A standardized in-chat checkout couldn't accommodate them. So OpenAI pivoted: instead of completing the transaction inside ChatGPT, the new architecture is branded apps that live inside ChatGPT, handle discovery and conversation natively, and then hand the user off to the retailer's own site or app to finish the job.
Target shipped a ChatGPT app for conversational shopping. DoorDash shipped one for grocery ordering. Walmart is integrating its Sparky assistant directly into ChatGPT. Etsy is on its way. The Apps SDK that powers all of this has been live since late 2025 and is built on the same Model Context Protocol (MCP) you may have heard about in the context of Claude Desktop and Cursor.
Most of the coverage of this pivot has framed it as an ecommerce story. It is — but the more interesting story for the rest of us is that "ChatGPT apps" are now a generally available distribution surface, and most non-retail SaaS businesses haven't noticed yet. If your customers already ask ChatGPT about your category — and they probably do — you can ship an app in roughly 24 engineering hours that puts your product inside the conversation as a discovery option, then routes the buyer back to your existing landing page to convert.
This piece is about whether you should, when, and what the honest tradeoff looks like.
What Apps in ChatGPT Actually Are
The Apps SDK isn't a wrapper around your existing API. It's a thin specification on top of MCP: you stand up an MCP server, define a small set of tools, and optionally provide a web UI that renders inside a sandboxed iframe when the conversation calls for richer interaction.
An app in ChatGPT exposes three things:
Tools. Discrete, typed functions the model can call when the conversation calls for them. For a contract management SaaS, the tools might be search_templates, get_template_details, and start_signup. For a meal-kit company, they'd be search_recipes, view_menu, and place_order.
An optional UI surface. The Apps SDK lets you render a small web view inside ChatGPT for cases where text isn't enough — a product gallery, a date picker, a configurator. The view is sandboxed, communicates with the MCP server, and is presented inline in the conversation.
An auth handoff. The user can authenticate into your app from inside ChatGPT, which means subsequent tool calls can run against their account. This is what makes the experience feel like an app, not a chatbot — the model is acting on their behalf inside your system.
Critically, when the user is ready to do something substantial (complete a purchase, sign a contract, finish onboarding), the app can hand off to your own site — opening a new tab with the user pre-authenticated, the cart pre-loaded, the form pre-filled. The conversation continues, but the conversion happens on your terrain. This is the architectural shift OpenAI made when they walked away from Instant Checkout: the conversation lives in ChatGPT, the conversion lives on your page.
The Distribution Math
ChatGPT had 900 million weekly active users as of Q1 2026. OpenAI has publicly disclosed that roughly 50 million product or service discovery queries flow through the platform every single day. Among knowledge workers and high-intent commercial buyers, the share is even higher — internal data we've seen from B2B partners suggests 22% of SaaS evaluation queries now start in ChatGPT rather than Google.
Without an app, your only path to being part of that conversation is being cited organically — through Reddit, YouTube transcripts, your own page being indexed, third-party reviews. We've written extensively about how to win those citations. They work, they compound, and you should still be doing them.
But an app puts you in the conversation by name. When a user asks "what's a tool for managing legal contracts that integrates with Google Drive," and your app is in their App Directory (or surfaces in the discovery layer), you become a first-class option, not a cited footnote. The user can immediately interrogate your product through the model, get answers in their voice, and start the signup with one tap that opens on your site.
To put numbers on the gap: a typical organic AI citation drives ~0.3% click-through to the cited page (our measurement across roast.page customers). A ChatGPT app surfaces to the user as a structured option — and early data from the launch partners suggests handoff rates from in-app interactions to the retailer site in the 8-15% range. That's 30-50x the engagement of an organic citation, on the same query.
The catch: an app only surfaces for queries where the model thinks an app is appropriate. You're not getting in front of every ChatGPT user. You're getting in front of the ones whose query the model interprets as "wants to actually do something now." That subset is small as a percentage, but it's the highest-intent subset of an already-high-intent traffic stream.
Who Should Actually Build One
Not every business should. Here's the four-question filter we've been using with founders.
1. Do your customers already use ChatGPT for category research? This is the most important filter. If your buyers are technical (developers, marketers, finance leaders, designers), the answer is almost certainly yes — those audiences over-index on ChatGPT usage. If your buyers are physical-product consumers who've never paid for a chatbot, the answer is closer to "not yet." Run the test: open ChatGPT, type the five most common questions your category gets, and see whether ChatGPT gives a useful answer with named alternatives. If yes, you have demand.
2. Is your conversion a few structured steps an agent could complete? Some conversions are inherently un-agentable. Selling a $200,000 ARR enterprise contract requires three calls, a security review, a legal redline, and a procurement process — no app is going to compress that. Selling a $29/month SaaS subscription is three structured fields and a credit card — an app can absolutely shepherd a user to that. Selling a sandwich is one tap. Map your conversion to a step sequence. If the sequence has more than ~6 steps or any step requires a human conversation, an app probably can't carry it across the line — but it might still be valuable for top-of-funnel discovery.
3. Can your team afford a one-person-week investment? A minimum-viable ChatGPT app — MCP server, 3-5 tools, basic auth handoff, no custom UI — is roughly 24 engineering hours for a competent developer who's never touched MCP before. With a UI surface, it's closer to 60 hours. With tight integration into your existing user system, it's 100+. If you can dedicate one engineer for a week, you can ship a v1. If you can't, hold off until you can — a half-built app in the directory is worse than no app at all.
4. Are you willing to support a second front door? The app becomes a real product surface. Your customers will use it. They'll send you support tickets through it. Their experience there reflects on your brand the same way your iOS app does. If your current support is at capacity, adding a ChatGPT app adds load. Plan for it.
If you answered yes to all four, the case to build is strong. If you answered yes to two or three, it's defensible but not urgent. If you answered yes to one or zero, focus on organic AEO until the conditions change.
What a Minimum-Viable App Looks Like
For the case where the answer is "build," here's the rough anatomy of what you're shipping. This isn't a tutorial — the Apps SDK docs are good and you should read them — but it's the shape of what gets built.
An MCP server. A small HTTP/SSE service that implements the MCP spec. There are SDKs in TypeScript, Python, Go, and Rust. The server exposes a list of tools, handles incoming tool calls, and returns structured responses. For a SaaS app, this is a thin layer on top of your existing API — most of your business logic stays where it is.
3-5 tools. Resist the temptation to expose every endpoint. For a typical SaaS landing page experience, the right tool set is:
search(query)— return a list of relevant entities (templates, products, courses, recipes — whatever your unit is)get_details(id)— return the structured details of one entitycompare(ids)— return a side-by-side comparison (high value for buyer-stage queries)start_signup(email, plan)— initiate signup and return a deep-link URL that opens the rest of the flow on your site, pre-filledstart_purchase(items)— initiate a purchase and return a deep-link URL that opens checkout pre-loaded
That's it. Each tool returns structured JSON. The model uses the structure to compose a useful in-conversation response. The user clicks the deep-link to convert on your site.
A deep-link handoff. The most important integration is the one-tap handoff to your existing landing page or signup flow. When the user is ready to convert, your app returns a URL that, when opened, drops them into a fully-prefilled state on your site — auth pending, cart loaded, form populated. This is why the pivot away from Instant Checkout was actually good news for landing pages: the conversion still happens on your URL, in your design system, with your analytics watching. You get the agent-driven discovery without giving up the conversion surface.
App Directory submission. Once your app passes review, it appears in the ChatGPT App Directory and becomes discoverable through the model's tool-selection layer. Submission involves a manifest, a privacy policy, an icon, and a few screenshots — same shape as the iOS App Store, minus the binary review.
The first version of this — search, details, start_signup, deep-link handoff — is genuinely a 24-hour build for a single engineer who's read the Apps SDK docs. We've watched four customer teams ship v1 apps in May 2026 and the median time was 28 hours from kickoff to App Directory submission.
The Landing Page Side
Building an app is half the work. The other half is making your existing landing page deep-linkable in the way the app needs.
If your signup flow is a five-step wizard, the app's start_signup tool needs a URL that drops the user into step three with the first two steps pre-filled from the in-conversation data. If your checkout is a multi-page funnel, the start_purchase tool needs a URL that lands them on the payment step with the cart already loaded.
This is engineering work that most landing pages haven't done. The default assumption was that every visitor starts fresh from the top. Apps break that assumption. Build deep-link support before you build the app — otherwise the handoff is brittle, and the user drops out at the seam between conversation and conversion.
A few concrete patterns we've seen work:
Signed handoff tokens. The app generates a short-lived signed token with the in-conversation context (email, plan choice, source). The landing page validates the token and uses it to prefill. This is more secure than passing data in query strings and works across domains.
Anonymous-to-authenticated bridges. If the user converts directly from the conversation (signs up, buys), the app returns a session URL that drops them straight into their authenticated dashboard without a second login. Friction here is where conversion leaks.
Visible attribution. Tag the entry point so your analytics can distinguish ChatGPT app traffic from other sources. This is the only way you'll know whether the app is producing real conversion value vs. vanity engagement.
The Apps-vs-AEO Question
This is the question I get most often. If I have 40 hours to invest, do I build a ChatGPT app or do I invest in organic AI citations and AEO work?
The honest answer is: it depends on which lever is more underbuilt for you right now.
If you're already cited well organically in ChatGPT — your brand shows up, your URL gets linked, your category queries surface you in the top three — then the app is the next leg up. You're capturing intent that organic citations alone can't fully convert.
If you're not cited at all organically — your brand is invisible in the conversation and the AI says "I don't know much about that company" — building an app is putting the cart before the horse. The model won't surface your app to users searching your category if it has no organic context for what your company does. Build the citation base first, then the app.
The cleanest sequence we've watched succeed:
Months 1-3: Establish organic citation footprint. Win mentions on Reddit, YouTube, comparison sites. Make sure your landing page is AI-readable (clear positioning, structured copy, llms.txt, schema). Track citation rate weekly.
Months 4-6: Once your brand has a recognizable presence in conversational AI, ship the app. The app will be surfaced for users who already encountered your brand organically — and the citation context teaches the model when to recommend your app vs. competitors.
Months 7+: Iterate on the app based on which tools get called, which handoffs convert, and which queries the model uses your app for. This is product work, not marketing work.
Skipping the citation work and starting with the app is the most common mistake. Apps without citation context get listed in the directory and ignored. Apps with strong citation context get surfaced into conversations and convert at the 8-15% handoff rates the launch partners are seeing.
What's Going to Happen Next
A few predictions for the rest of 2026, with reasonable confidence:
The App Directory will become a real product surface. Right now it's a discoverable list of ~200 apps from launch partners and early adopters. By Q4 2026 it will be in the tens of thousands, with search, ranking, and quality signals — the way the iOS App Store evolved between 2008 and 2010. Early entrants will accrue rank advantages the way early iOS apps did.
The MCP standard will absorb Apps SDK conventions. Claude Desktop, Cursor, Windsurf, and now ChatGPT all consume MCP servers. The Apps SDK adds a few ChatGPT-specific patterns on top (UI surfaces, App Directory metadata), but the core protocol is the same. A well-built MCP server can be one app surfacing in multiple agentic clients with minimal additional work — that's the leverage move.
Apps will become a competitive signal in vendor evaluations. Procurement-stage buyers in B2B already use ChatGPT to summarize vendor differences. By late 2026, "does this vendor have a ChatGPT app" will be one of the implicit shortlist filters — the way "does this vendor have a mobile app" became a filter in 2014. Being early is cheap. Being late means rebuilding.
The pricing-page-as-tool pattern will emerge. Several teams are experimenting with exposing pricing-quote generation as a tool — the user describes their team size and use case in conversation, the app returns a personalized quote and a deep-link to start. For categories where pricing is "contact us" today, this collapses a 3-day sales motion into a 30-second conversation. Worth thinking about for your business.
What to Do This Week
If you've read this far and the four-question filter pointed toward "build," here's the next-week plan.
Day 1: Read the Apps SDK docs and the MCP spec. Both are short. Together they're a one-afternoon read.
Day 2: Audit your landing page for deep-link support. Identify which conversion steps need to be reachable from a URL with pre-filled context. Spec the changes.
Day 3-4: Stand up an MCP server with the five tools above. Wire each tool to your existing API. Implement signed handoff tokens.
Day 5: Ship the deep-link support on the landing page side. Test end-to-end: ChatGPT conversation → tool call → handoff URL → completed conversion on your site.
Day 6-7: Polish the app metadata (name, description, icon, screenshots, privacy policy). Submit to the App Directory.
By the end of next week, you'd be one of the first 1,000 apps in the directory. The compounding advantages of being early in a new distribution surface are real and they're earned by the teams that don't wait for permission.
If you're not sure whether your landing page is even ready to be a handoff destination from a ChatGPT app — whether the conversion path is clean enough for an agent-routed user to complete — our free landing page analysis walks through that exact question. The pages that are ready for the app era look different from the pages that aren't, and the gap is visible in the analysis output.
The shift OpenAI made in mid-2026 wasn't a retreat. It was a recognition that the conversion belongs on your site, the discovery belongs in the conversation, and the seam between them is the most valuable real estate in the agent era. The teams that own that seam are going to compound.