How do I prepare an ecommerce site for AI agents?
Serve agents a version of the store without the elements that only help humans: overlays, hover menus, anti-bot challenges and client-side rendering. Then give them what they came for as plain text: product facts, current price and stock, and a checkout they can complete. The products, pricing and business logic stay the same. Only the rendered surface changes, the way a mobile site differs from desktop.
Where agents fail on a store built for humans
Every pattern that lifts human conversion is a failure mode for an agent.
- Overlays and promotional interruptions. Every popup, modal and newsletter gate is a step the agent has to detect, dismiss and verify.
- Navigation instability. Hover menus, dropdowns and injected elements produce click paths that break under automation timing.
- Anti-bot layers. CAPTCHAs and bot mitigation block identified agents along with scrapers, often silently, with no error the agent can read.
- Clutter. Ad blocks, chat widgets and decorative elements widen the surface an agent has to scan and raise the number of screenshots a computer-use agent needs.
- Client-side rendering. Pages that re-render into a different structure between visits force the agent to re-parse from scratch, and simple bots never see the rendered content at all.
Product data as text
A simple bot reads the HTML your server returns. If the product name, price, variants, availability, shipping terms and returns policy are present in that HTML as text, the bot has them. If they load from an API after the page renders, sit inside an image, or appear only after a size is selected, the bot has nothing.
Put the facts an assistant will quote on the page in words: "in stock", "ships in 2 business days", "fits models A and B", "free returns within 30 days". These are the pre-computed answers that end up in a recommendation. A spec table in HTML beats a spec sheet PDF, and both beat a configurator.
Price and stock freshness
AI apps cannot cache the internet. Retrieval loads your page into a context window every time someone asks, so agents re-fetch, and the answer they give is only as current as the page they received. A stale price in a cached edge copy becomes a wrong price in an answer. Serve current values to agents, and make sure the agent-facing page updates when the human one does.
Checkout an agent can complete
For computer-use agents, the measure is task completion: did the agent add the right item and reach the confirmation page, in how many steps, at what token cost. Forms rendered in HTML with stable field names, visible validation messages in text, no CAPTCHA for identified agents, and a confirmation state written in words all raise completion. In the benchmark we ran with KERNEL on five test stores (1,000 runs), an agent site instead of the default store produced 24 to 35% faster task completion, 100% completion with zero failed runs across 500 agent-site runs, and up to 38% fewer steps. Fewer steps means fewer screenshots, and the bill for those screenshots increasingly lands on your customer.
Machine payments are next
Today most commerce reads are research. On tollbit.com Amazonbot, which Amazon documents as powering Alexa and other services, made 2.9K successful reads in 30 days, while Amzn-User, the retrieval agent that acts on a shopper's request, appeared in single digits. Meta launched Muse, a personal assistant, in September 2026. Personal assistants that hold a login, preferences and a payment method will need authentication, APIs and a way to pay. HTTP 402 Payment Required is the standard response for "this has a price", and it is already in use for content. The same rail generalizes to carts.
Doing it without a replatform
An agent site is the agent-facing version of your store on a subdomain of your own domain, with identified AI agents routed to it by a CDN or WAF rule. Humans and Googlebot never go there. The site fetches your origin, strips what agents do not need, and returns a page they can read and act on. Setup is two IT changes, and it works alongside the CDN and bot management you already run. TollBit does this for stores on Shopify and other commerce platforms.
Frequently asked questions
Does any of this change what shoppers see? No. The human store is untouched. Agents are routed to a separate rendering of the same pages, with the same products and prices.
Are agents actually buying, or just browsing? Mostly researching, today. Retrieval bots and shopping assistants read product pages to answer "which one should I buy". Completed purchases by agents are early, and the stores that are usable when they arrive will take them.
Is serving agents a different page a form of cloaking? Same facts, cleaner format, bots only. The prices, stock and claims match the human page. Search engines are never routed to the agent site, so there is no difference between what Google indexes and what humans see.