Strip away the hype and an AI shopping agent's purchase attempt is a pipeline of unforgiving, mechanical steps. Understanding exactly where that pipeline breaks on your site is worth more than any think-piece about the agentic future — because each break is a fixable line of markup.
Step 1: Discovery — can it find the product?
The agent arrives with intent ("40-inch USB-C monitor under $300") and needs a path from your homepage or search to a product page. It follows real links and reads accessible names. Sites whose navigation lives in JavaScript handlers, or whose search is unlabeled, fail here — before price ever enters the picture.
Step 2: Extraction — can it read the offer?
On the product page, the agent needs name, price, currency, and availability in machine-readable form. Product/Offer JSON-LD answers this in one read. A price that exists only as styled text in a rendering-dependent component forces inference — and agents sent to spend money are (rightly) conservative: when they can't verify the offer, they move to a competitor they can verify.
Step 3: Cart — can it act?
"Add to cart" must be a real, discernible control. The agent clicks it and then needs confirmation the action worked — a cart state it can read. Silent cart updates that only manifest as a badge animation are invisible to it.
Step 4: Checkout — can it finish?
This is where most attempts die. Checkout forms with unlabeled fields, no autocomplete tokens, human-only affordances (drag interactions, image puzzles), or multi-step wizards whose state lives entirely in JavaScript. Each is a wall. Agents operating for real customers handle payment through their own vaulted flows — what they need from you is a form they can parse and a path they can complete.
Most sites don't lose the agent sale at discovery. They lose it at the last hundred pixels: the checkout form.
What the winners do
Nothing exotic: labeled fields, autocomplete tokens, real buttons, readable cart state, JSON-LD offers, and a deterministic test that runs the whole pipeline on every deploy. The sites capturing agent-driven revenue aren't the ones with an "AI strategy" deck. They're the ones whose checkout a machine can finish.