Cart abandonment counts a shopper who looks at a basket and walks away from it. Remove the basket and the measurement does not get harder. It stops having a subject. Baymard Institute puts the average documented rate at 70.22%, drawn from fifty separate studies. The page does not say how any of the fifty defined the event.

OpenAI’s Agentic Checkout Spec collapses two funnel steps into a single call. ChatGPT posts to the merchant’s /checkout_sessions endpoint with the items already chosen, so the merchant’s first sight of the shopper is a basket that arrives full. Add-to-cart and begin-checkout stop being separate events on the retailer’s property, because neither happened there.

The same spec allows a session four states: not_ready_for_payment, ready_for_payment, completed, canceled. There is no fifth. A session the shopper quietly stops attending simply stops, at whatever state it last reached, with no timeout defined and no event to fire. The merchant holds a cart that nobody has left, because leaving was never written down.

Baymard concedes that much of the 70% was never a failure. 42% of US shoppers say they abandoned because they were browsing rather than buying, and the browsing has now been handed to software. Imagine an agent opening a live session at three merchants and completing one. Two merchants are left holding a full basket that no human ever set down. One may send a reminder email to a customer who has already bought.

The object has not vanished. It has changed owner. Google’s Universal Cart, shown at I/O in May, holds items from several retailers at once and carries them from Search into Gemini. Google’s own example is a shopper gathering parts from different shops into one cart. That cart sits in nobody’s funnel, and therefore in nobody’s denominator.

Google’s merchant documentation commends the arrangement partly on the grounds that it may reduce cart abandonment, which is true in the sense that a shop with no door records no theft.

The trade has decided the problem cancelled itself. OpenAI scaled Instant Checkout back in March, and Walmart’s Daniel Danker told Wired that conversion ran three times lower inside the chatbot than on paths that clicked out, as Modern Retail reported. His stated reason repays attention: shoppers had items sitting in a Walmart cart already, and did not want the assistant buying one of them alone. One shopper, two baskets, and only one of them countable.

Forrester finds that buying inside an answer engine is the least-adopted use case among people who use them regularly, and that 35% of US Gen Z online adults searched for products in ChatGPT last month. Neither figure is large. Size is not the failure mode here. A denominator does not break by shrinking; it breaks by going partial.

Baymard prices the recoverable loss at $260 billion and locates the cure in better checkout design. Design presumes a surface, and the surface is what has gone. Guest checkout and shorter forms are answers to a shopper who is looking at something. The rate will go on being quoted to two decimal places. That is now the most dependable thing about it.