Why AI agents can't shop online (yet)
The web has hundreds of millions of product pages, checkout flows, and pricing tables. Almost none of them work for an agent. Here's why.
There's a reasonable assumption that because AI agents are good at using computers, they're good at shopping. They can read a webpage. They can fill in a form. They can, technically, complete a checkout flow.
But technically isn't the same as actually.
The problem isn't capability. It's structure.
When a human lands on a product page, they're doing a lot of low-level work they don't notice: skimming for the price, finding the right variant selector, parsing whether "in stock" means today or in three weeks, deciding if the seller is trustworthy based on some combination of reviews, site design, and gut feel.
An agent can simulate that process. But it's slow, fragile, and expensive. It's using a large language model to parse visual noise that only exists because the page was designed for a human eye.
The deeper issue: none of that information was meant for the agent. The price is buried in a <span> with seventeen CSS classes. The shipping estimate is rendered by JavaScript three seconds after the initial load. The return policy requires clicking "expand" on a modal. These are design choices made for a specific kind of reader — one with eyes, a cursor, and patience.
Agents have none of those things.
Captchas are the obvious part
Everyone mentions captchas when this comes up. They're a good symbol but not the real bottleneck. Captchas are solvable. The structural problem isn't.
The structural problem is that the modern web is optimised for conversion, not comprehension. Every dark pattern, every upsell interstitial, every "only 2 left!" urgency nudge — these are features for humans. For an agent, they're noise at best, blockers at worst.
An agent trying to buy something on a standard e-commerce site has to:
- Parse unstructured product data across inconsistent schemas
- Navigate checkout flows designed to slow you down (so you add more to your cart)
- Handle authentication, sessions, and cookies built for browsers
- Deal with payment flows that require a human to verify with their bank
- Figure out, without explicit guidance, what the right price actually is
None of this is impossible. It's just wasteful. You're spending compute budget fighting the friction that was specifically engineered into the buying experience.
The agent doesn't need a better browser. It needs a different store.
The fix isn't a smarter agent. It's a marketplace that was built assuming the buyer is a program.
That means structured listings. Machine-readable pricing. APIs instead of forms. Offer-and-response instead of "add to cart." A trust layer that doesn't depend on recognising a logo.
That's what we're building. Not a better scraper. A different kind of store.