ChatGPT's Checkout Button Flopped. The Recommendation Didn't.
OpenAI scaled back in-chat checkout after ~30 merchants signed on. Yet AI retail traffic jumped 393%. The sale moved upstream, to the recommendation.
ChatGPT’s Checkout Button Flopped. The Recommendation Didn’t.
In September 2025, OpenAI launched Instant Checkout inside ChatGPT with Etsy, Shopify, and Stripe. The pitch was that you’d ask ChatGPT for a product, see it, and buy it without leaving the chat. The Agentic Commerce Protocol would handle the rest. A million Shopify merchants were “coming soon.”
Six months later, OpenAI pulled back. By the time CNBC reported the retreat in March 2026, roughly 30 Shopify merchants were live with the checkout feature — a rounding error against Shopify’s millions of stores. OpenAI called it a transition, not a death, and said purchases would move to apps instead. Either way, the flagship “buy in the chat” experience had stalled.
Here’s the part most people read wrong. The checkout button failing does not mean AI shopping failed. It means the sale was never going to happen at the button. It already happened three messages earlier, when ChatGPT named three products and your brand wasn’t one of them.
The data says shopping moved to AI. Buying didn’t.
While in-chat checkout collapsed, AI-driven retail traffic exploded. Adobe found that traffic from AI sources to U.S. retail sites rose 393% year over year in Q1 2026, on top of a 693% surge over the 2025 holidays. That traffic isn’t junk, either. By March 2026, Adobe measured AI-sourced visitors converting 42% better than non-AI traffic and generating 37% higher revenue per visit. A year earlier, the same AI traffic converted worse than human traffic. The line crossed.
So people are clearly using AI to shop. They’re just not checking out inside the chat. Adobe’s survey of more than 5,000 U.S. consumers found that 73% of those who shop with AI use it as their primary source of product research. Research, compare, decide, then click out to a retailer to finish the purchase.
Forrester saw the same split from the other side. In its March 2026 consumer data, regularly used answer engines were dominated by asking questions and researching products. Completing a purchase inside the engine was the single least-adopted use case. OpenAI built a checkout for a behavior that barely exists, and skipped past the behavior that’s growing fastest.
| Stage of the AI purchase | What’s happening in 2026 | Who’s winning |
|---|---|---|
| Research & discovery | AI retail traffic up 393% YoY; 73% of AI shoppers call it their primary research tool | The brands the model names |
| Comparison | Buyers ask “best X for Y,” weigh 2-4 options the assistant surfaces | The brands with strong third-party corroboration |
| Checkout | In-chat purchase is the least-adopted behavior; OpenAI scaled it back | The retailer’s own site, still |
The strategic read: AI owns the top of the funnel and the click is moving back to your own store to close. The decision gets made in the recommendation. The transaction happens wherever it always did.
Why the checkout was the wrong thing to build first
A checkout is a hard, unglamorous machine. It needs live inventory, accurate pricing, tax collection across every U.S. state, fraud liability, returns, and refund disputes. CNBC’s reporting noted OpenAI hadn’t solved state sales tax remittance as of early 2026. None of that is an AI problem. It’s a commerce-operations problem that Shopify, Amazon, and Stripe spent a decade grinding through.
And it taxed merchants for the privilege. ChatGPT’s Agentic Commerce Protocol charges a 4% fee on top of existing payment costs. Perplexity, by contrast, charges merchants nothing and funds shopping through subscriptions. When the behavior is rare and the toll is high, you get 30 merchants.
The recommendation has none of those frictions. It doesn’t need a payment rail or a tax engine. It just needs the model to have a reason to say your name. That’s a content and authority problem — exactly the kind your team can actually move.
What this means for brand visibility teams
Stop watching for the checkout integration. Start treating the recommendation as the conversion event, because for most buyers, it is. By the time someone clicks through to a retailer, the assistant has already narrowed a thousand options to three. If you’re in that three, the checkout button’s owner barely matters. If you’re not, no amount of agentic-commerce plumbing saves you.
Here’s where to put the effort instead:
- Audit the comparison queries, not the brand queries. “Tell me about Acme” is a vanity check. The money is in “best running shoe for flat feet” and “Acme vs. competitor for small teams.” Those are where the model builds its shortlist. Track them, and track who else shows up. (We break this down in the metrics that actually matter for AI search.)
- Feed the model reasons to recommend you. Models lean on third-party corroboration (reviews, comparisons, roundups, forum threads) far more than on your own marketing copy. Earned mentions across the sites the model trusts move recommendations in a way an on-site product page can’t. This is the core of getting your brand recommended by AI.
- Make your product data machine-legible. Adobe’s own finding was that many retailers aren’t structured for AI to read them cleanly. Clear specs, structured product data, and consistent naming let the model match you to the query with confidence instead of guessing.
- Win the click-out, not the in-chat sale. Since buyers leave AI to purchase, the recommendation has to carry intent that survives the jump. A vague mention loses to a specific one. “Acme, known for X” travels; “Acme is an option” evaporates.
The protocol war is a sideshow
There’s a real fight underway between OpenAI’s Agentic Commerce Protocol, Google’s Universal Commerce Protocol, and Perplexity’s approach. Shopify now syndicates catalogs to ChatGPT, Perplexity, and Copilot at once. It’s worth tracking. It is not worth waiting on.
The protocols decide who processes the payment. They don’t decide who gets recommended, and the recommendation is the part that’s already worth 393% more traffic to your category this year. Plumbing for a checkout nobody uses yet, while ignoring the recommendation everybody uses now, is optimizing the wrong end of the funnel.
OpenAI just learned that lesson in public. The brands paying attention won’t repeat it. The buy button can wait. Being the answer can’t — and once the model settles on its default three for your category, dislodging it is the hardest visibility problem there is.
RivalHound tracks your brand’s visibility across ChatGPT, Google AI, Perplexity, and more. Start monitoring to see where you stand.