On September 29, 2025, OpenAI and Stripe jointly added a 'Buy' button to ChatGPT: U.S. users asking about products in chat could place orders instantly—starting with Etsy sellers and quickly expanding to over one million Shopify merchants. According to Fidji Simo, CEO of OpenAI’s Applications business, this was intended to let businesses of all sizes 'capture commerce where people already are,' enabling buyers to complete purchases directly within conversations.
Five months later, on March 4, 2026, Daniel Danker, Senior Vice President of AI Acceleration at Walmart, described the feature during Morgan Stanley’s industry conference as 'a very temporary phase' and added, 'You won’t see that experience next month.' On March 24, OpenAI officially confirmed the demotion of native in-chat checkout to low priority, shifting focus instead to product discovery—and returning checkout entirely to merchants’ own interfaces.
A feature jointly backed by two star companies and touted as 'the next step in AI commerce' lasted only five months. Yet concluding from this that 'AI-powered purchasing is doomed' would misread the entire conflict—because in the month following its retreat, two other fronts accelerated markedly: on April 2, Coinbase donated x402, a payment protocol it had incubated, to the Linux Foundation, establishing the x402 Foundation; on April 28, Google donated its Agent Payments Protocol (AP2)—released just seven months earlier—to the FIDO Alliance. Viewed together, these developments reveal that the so-called 'agent payment protocol war' is not a single battlefield—but three.
## Three Protocols, Answering Three Distinct Questions
The first is OpenAI and Stripe’s Agentic Commerce Protocol (ACP), which addresses the **entry question**: Where should checkout occur? ACP’s approach is to bring the store into the chat—when a user clicks 'Buy' inside ChatGPT, Stripe issues a 'shared payment token' credential, binding the merchant and cart amount while never exposing card numbers. Merchants retain their role as 'merchant-of-record', handling order fulfillment, tax calculation, shipping, and returns as before. The protocol is open-sourced under Apache 2.0, and Stripe claims existing merchants can integrate it with just one line of code.
The second is Google’s Agent Payments Protocol (AP2), launched on September 16, 2025—thirteen days before ACP—and backed by more than sixty institutions, including Mastercard, American Express, PayPal, Coinbase, and UnionPay International. It tackles the **authorization question**: How do we prove this payment was truly authorized by the user? AP2’s core component is the 'mandate'—an immutable, cryptographically signed digital contract. When you tell your AI, 'Find me a pair of white running shoes,' that instruction becomes an 'intent mandate'; when the AI presents a shopping cart and you confirm, you sign a 'cart mandate'—specifying exactly what was purchased and for how much, making it irrevocable. If you say, 'Snag concert tickets the moment they go on sale,' you pre-sign a mandate specifying price caps and time windows; the AI executes automatically once conditions are met. Google maps this design onto three questions: authorization (did the user genuinely grant this permission?), authenticity (does the AI’s request truly reflect user intent?), and accountability (who is liable if something goes wrong?).
The third is x402, which answers the **settlement question**: How do machines pay each other directly? It revives HTTP’s long-reserved 402 status code—'Payment Required.' An AI agent requests a paid API; the server replies with a 402, including amount, payee, and accepted payment methods; the agent signs a payment credential and retries—the entire flow proceeds through verification, settlement, and delivery, requiring no account registration or API keys, and charging per request. It most commonly settles using stablecoins on Ethereum-compatible or Solana blockchains, and the protocol itself incurs zero fees.
These three layers have distinct responsibilities—and thus distinct success criteria: the entry layer competes on traffic and habit; the authorization layer on standards and trust; the settlement layer on the scale of machine-to-machine economies. Framing them collectively as a 'protocol war' was flawed from day one.
## The Five-Month Retreat Proves: Entry-Layer Protocols Cannot Create Demand
In hindsight, Instant Checkout failed—not because of flaws in the protocol itself—but because neither supply nor demand was ready.
On the supply side, merchants did not want to be 'moved into' someone else’s chat window. Harley Finkelstein, President of Shopify, told investors that among Shopify’s millions of merchants, only a dozen had actually launched the feature—the bottleneck remained squarely on the AI company side. According to The Information, even after nearly six months live, the feature supported only single-item purchases, lacked multi-item carts and coupon codes, and hadn’t even built out state sales tax reporting systems. Walmart offered harder numbers: for the same item, conversion rates for direct sales inside ChatGPT were only one-third those achieved when users were redirected to the merchant’s own website for checkout.
Demand-side reality was simpler: users simply don’t treat chatbots as stores. As cited by The Information, sources reported users 'weren’t actually completing purchases via ChatGPT at all.' For OpenAI, e-commerce ambitions were further pressured by market-share anxiety—per Apptopia data, ChatGPT’s share of daily active users among U.S. AI apps fell from 57% in August 2025 to 42% in February 2026, while Gemini rose from ~13% to 25% over the same period. As the moat around its traffic gateway narrowed, the revenue story based on transaction fees grew more urgent—and urgency led to premature button deployment before ecosystem maturity.
What happened after the retreat is even more revealing: ACP didn’t die—it pivoted. In the same announcement, OpenAI shifted focus to product discovery—image search, price comparison, conversational filtering. Per media reports, ACP transformed into a pipeline for merchants to feed product catalogs and promotions directly into ChatGPT; Target, Sephora, Nordstrom, Best Buy, and Home Depot have all joined. The protocol remains—but concedes one truth: the checkout action must, for now, stay firmly on the merchant’s own turf.
Google and Shopify’s Universal Commerce Protocol (UCP), launched January 11, 2026, offers another answer to the same problem: rather than moving checkout into chat, UCP enables merchants and AI agents to mutually declare supported capabilities and negotiate per transaction—payment tools are negotiated, not prescribed; when human intervention is needed, a dedicated 'escalation' state hands the transaction back to a person; merchants remain the merchant-of-record throughout, retaining full control over customer relationships and data. Walmart, Target, Etsy, and Wayfair co-developed UCP, which will also power checkout in Google Search’s AI mode and within the Gemini app. The entry-layer contest hasn’t ended—it has retreated from 'bringing the store into chat' to 'connecting the store with chat.'
## Signing Authority Is the Slowest—and Deepest—Variable
The noise around the entry layer obscures a critical fact: the authorization layer is the hardest to ship.
When Google donated AP2 to the FIDO Alliance, it also contributed the 'Verifiable Intent' standard co-developed with Mastercard—designed to create tamper-proof evidence for 'AI signing on your behalf.' But look at the standards bodies’ own timelines: in April 2026, two specifications were just merged into a single repository; at the October 2026 W3C meeting, integration approaches were still under discussion; draft v0.1 targets December 2026, and the integrated final version isn’t expected until mid-2027. In other words, technical demos of 'AI spending autonomously' are already running—but the industry-wide rules defining 'whose signature counts' remain in draft form.
More substantive gaps persist in accountability: under pre-authorization, if an AI buys the wrong item, who handles refunds—the merchant, the platform, or the model provider? Credit-card-era chargeback rules—refined over decades—still lack any publicly available, unified industry standard covering 'human-absent' agent transactions. Even the FIDO Alliance, in its announcement accepting the donation, explicitly lists 'verifiable accountability' as a goal—not a current reality.
The participant list itself reveals industry sentiment: Mastercard and American Express appear both among AP2’s sixty-plus partners and as founding top-tier members of the x402 Foundation; Stripe co-developed ACP, endorsed UCP, and holds a top-tier seat on x402; Google launched AP2, co-developed UCP with Shopify, and is also a top-tier member of the x402 Foundation. No one is fighting a single, focused protocol war—instead, giants are placing simultaneous bets across all three layers, hedging precisely against the uncertainty of 'which layer will stabilize first.'
## My Three Judgments
First, 'Which protocol will win?' is the wrong question. Entry, authorization, and settlement are three distinct contests—each with non-interchangeable success criteria—and highly overlapping participation. The likeliest outcome isn’t one protocol dominating all layers, but layered coexistence: entry-layer protocols follow traffic flows; authorization standards settle into formal standards bodies; settlement-layer adoption depends on the real-world scale of machine economies.
Second, the true bottleneck in the entry layer is merchants’ fear of commoditization—not technology. Walmart’s conversion gap shows that breaking products down into isolated chat cards discards the brand context essential for conversion. UCP’s insistence on preserving merchant-of-record status and enabling capability negotiation on a per-transaction basis is a direct response to that fear—a better direction than 'moving the store into chat.'
Third, for ordinary users, the interface most worth paying attention to right now is the mandate. 'AI buying for you' is shifting from click-by-click confirmation toward pre-signed, autonomous execution. Before entrusting your wallet to any agent, ask three questions: Does the mandate include constraints—budget cap, merchant scope, and time window? (All three are indispensable.) Is every transaction backed by tamper-proof evidence? (In disputes, that evidence chain is your contract with the AI.) And who handles refunds? (Knowing who the merchant-of-record is matters far more than knowing which protocol was used.)
The five-month retreat wasn’t of AI commerce—but of the fantasy that 'slapping a checkout button into a chat window equals revolution.' Until the rules governing signing authority are written, the most compelling thing to watch isn’t who launches the next protocol—but who first learns to guard that pen for the user.
Would you sign a pre-authorization mandate with a budget cap, letting an AI snag tickets or track price drops for you? Share your bottom-line conditions in the comments.
Source: Google Cloud Blog, Powering AI commerce with the new Agent Payments Protocol (AP2), https://cloud.google.com/blog/products/ai-machine-learning/announcing-agents-to-payments-ap2-protocol