On September 29, 2025, Stripe announced that ChatGPT users in the United States could buy goods from US-based Etsy businesses inside the chat, with “over a million Shopify merchants” promised soon after. OpenAI’s own launch post framed it as the first step toward ChatGPT helping people buy things, powered by the new Agentic Commerce Protocol. Shopify’s president called it the “new frontier” for online retail.

Six months later the frontier was closed. By February 2026, Forrester counted roughly 30 Shopify merchants actually live on the feature, a rounding error against the million once promised. In March 2026, OpenAI confirmed it was deprioritizing Instant Checkout as a standalone product, saying it “did not offer the level of flexibility” merchants and shoppers needed, as DigitalSilk’s post-mortem documents. CNBC’s reporting added the operational texture: item information was often inaccurate, and onboarding merchants proved far harder to scale than OpenAI expected.

The industry read this as a retreat. It was the opposite. OpenAI did not abandon shopping. It abandoned checkout, kept the part of the purchase that actually happens inside a model, and handed the transaction back to retailers. That relocation, more than any protocol announcement this year, defines where agentic commerce is going: agents decide, merchants transact. And it makes the quality of the data behind the decision the single point of failure for the entire industry.

What actually died

The failure was structural, not cosmetic. Checkout.com’s analysis of the shift, published in April 2026, is the clearest technical account of why Instant Checkout could not scale. Behind the native-looking chat interface, merchants were still operating the entire commerce stack: payments, order management, fulfillment, refunds, and customer support. ChatGPT surfaced the experience. The merchant carried the burden.

That asymmetry could survive a handful of curated launch partners. It could not survive a million merchants. Every merchant connection required deep integration work, and every integration exposed a new surface for the data-quality problems CNBC documented: wrong prices, stale inventory, mismatched product information. When the interface promises one-click purchase inside a chat, every one of those errors converts directly into a wrong order, a frustrated customer, and a support ticket the merchant pays to answer.

The numbers tell the story plainly. Thirty live merchants out of a million promised, five months in. OpenAI’s own language, per The Information’s original reporting cited by Forrester, described the move as “temporary” and a decision to “scale back.” Forrester’s verdict was blunter: with checkout moving into merchants’ apps within ChatGPT or back to their websites, “the answer engine [is] just a referral, as search engines always have been.”

What replaced it

The replacement is not a smaller ambition. It is a different one, and it is already deployed. Per DigitalSilk’s documentation of the current experience, every ChatGPT tier, including the free plan, now gets:

  • Visual product browsing, with real product images instead of text-only recommendations
  • Side-by-side comparison tables showing price, ratings, and key features
  • Image-based search, where a shopper uploads a photo and gets similar products
  • Conversational refinement that narrows options by budget, style, or use case

Checkout happens on the retailer’s own site. Major retailers, including Walmart, Target, Best Buy, Sephora, Nordstrom, and The Home Depot, are feeding live product data into ChatGPT through the expanded Agentic Commerce Protocol. Checkout.com’s Rami Josef described the new architecture in one sentence that deserves to be quoted in full: “AI is becoming a new shop window for commerce, while merchant infrastructure remains the system of record.”

Two checkout models survived the culling. In the redirect model, the shopper moves from ChatGPT to the merchant’s website or app to pay. In the app model, merchants build dedicated experiences inside ChatGPT using OpenAI’s SDK, with checkout powered by their own infrastructure. In both, per Checkout.com, “AI handles discovery and intent; merchants retain control of checkout.”

Meanwhile the Agentic Commerce Protocol itself is expanding beyond payments into product discovery and richer commerce interactions. The protocol that once carried the transaction now carries the product data instead. Follow that thread to its end and you arrive at the uncomfortable destination: the most consequential data flowing through agentic commerce’s flagship protocol is ratings, prices, and descriptions, and none of it is verified by anyone.

The referral engine converts like crazy

Here is the paradox that explains why OpenAI “won” by retreating. DigitalSilk reports that ChatGPT sessions convert at roughly 15.9 percent, against 1.8 percent for Google organic traffic. The platform was never bad at generating purchase intent. It was bad at being a storefront. The high-intent audience was always there; what failed was the plumbing that tried to absorb the whole transaction.

The demand-side numbers back this up. Forrester’s December 2025 Consumer Pulse Survey found that 23 percent of Gen X US online adults had used ChatGPT in the past month to search for products, climbing to 32 percent of Millennials and 35 percent of Gen Z, with older generations in the single digits. Product discovery inside answer engines is now a mainstream behavior for everyone under 60, and it compounds yearly.

So the realistic near-term picture of “agentic commerce” is not an agent with your card completing orders autonomously at 3 a.m. It is a comparison engine with elite conversion rates, feeding shoppers into retailer checkouts. The agent’s deliverable is not a paid order. It is a ranked, side-by-side recommendation.

That is a much bigger responsibility than it sounds like.

When the decision is the whole job, the decision is the attack surface

Two weeks ago, Visa published the first rigorous benchmark of AI shopping behavior. The results, in Visa’s own release: 22 models from 12 labs failed 42.8 percent of realistic shopping runs, only 34.2 percent of scenarios passed three consecutive attempts, and, most damning, 54 percent of failed runs still ended in a purchase. Agents do not fail toward walking away. They fail toward buying.

Now combine that finding with the post-checkout architecture. The transaction, the last three seconds, sits back on retailer infrastructure where fraud checks, chargebacks, and customer service live. The decision, the three minutes that determine which product wins, sits entirely inside the model. An agent biased toward action, pointed at a comparison table, with no verification of the table’s contents: that is the exact configuration the Visa benchmark measured, and it fails silently.

The comparison table deserves special attention, because it is a fraud amplifier disguised as a neutral UI. A human browsing a marketplace scrolls past dozens of listings and applies accumulated skepticism. A comparison table in an answer engine renders three to five options with price, rating, and features. The rating column decides the winner. A 4.6 against a 4.8 is a 0.2-star margin, and a few hundred purchased reviews will produce it.

Review manipulation has always been profitable. In a comparison-first agentic world it becomes deterministic: the fake reviews do not need to persuade a human, they only need to shift two decimal places of a star average that an agent reads programmatically, ranks mechanically, and presents as the answer. The FTC’s consumer review rule, in force since October 2024, treats fake reviews as deceptive practices with per-violation penalties, and enforcement actions through 2026 have made the liability real. But enforcement is a deterrent that operates on the timescale of lawsuits. Ranking operates on the timescale of milliseconds.

The layer the retreat exposed

This publication has argued since the fall that the agentic commerce stack verifies the transaction and not the product. The Instant Checkout retreat sharpens the argument into its final form. The payments layer is no longer even contested inside the flagship assistant: OpenAI gave it back to the merchants. What OpenAI kept, and what every answer engine from Google’s AI Mode to the next entrant is now racing to own, is the recommendation layer.

That layer has three inputs: prices, which merchants now feed directly and accurately because errors cost them sales; product attributes, which come from the same feeds; and reputation signals, ratings, review counts, and badges, which come from marketplaces whose review base everyone in this industry knows is polluted. Amazon itself has blocked hundreds of millions of suspected fake reviews, and the FTC’s updated complaint over its ad auction alleges the ranking signals around those reviews were manipulated for years.

Two of the three inputs are getting cleaner every quarter, because merchants own them and errors cost money. The third is getting more valuable to manipulate at exactly the rate agents gain influence over purchases. That is the arbitrage of the next two years, and it will be exploited.

What fills the gap

A recommendation layer running on unverified reputation data needs an evidence layer, and the requirements are now specific:

  • Score review quality, not review quantity. A 4.8 average across 30,000 reviews carries no information about authenticity. GoBuy’s Smart Score runs 0 to 100 and is computed from review authenticity, weighting verified signals up and fabricated patterns out.
  • Filter continuously. Review fraud is a stream, not an event. A trust score computed once decays on arrival.
  • Return a shortlist, not a catalog. An agent making a recommendation does not need ten thousand options. GoBuy returns the top seven products per category, which forces the ranking to be defensible rather than comprehensive.
  • Reward track records, not spikes. A product holding a Smart Score above 80 for 90 days earns GoBuy Verified. The 90-day window is what separates a durable product from a manipulated one.
  • Arrive as data, through the same protocol everything else uses. GoBuy’s MCP server at gobuy.ai/api/mcp exposes this layer to any agent, and the agent docs cover integration. For the human still shopping in a browser, the Chrome extension injects the trust panel directly onto Amazon product pages.

OpenAI’s retreat settled an argument the industry has been having since September 2025. Agents will not own the transaction for a long time, if ever. They own the decision, starting now. The decision runs on trust data nobody verifies. Whoever fixes that, fixes agentic commerce.

Agentic commerce coverage, product trust analysis, and the daily state of AI shopping run on GoBuy.ai. Building a shopping agent? Start with the GoBuy agent docs and the MCP server at gobuy.ai/api/mcp.