What If an AI Agent Tries to Check Out on Your Site?
Non-human visitors are booking and buying now. Here's what small sites should fix first.
For the past two years the question was: how do we get cited by AI search? In 2026 the question moved one step further. Once you're cited, can an AI actually complete something on your site?
Tools that click the web on a user's behalf are already in ordinary hands: ChatGPT's agent mode, Claude's Computer Use, Perplexity's Comet browser. There are also efforts to standardize the payment step itself, such as the Agentic Commerce Protocol published by OpenAI and Stripe. In short, it's no longer unusual for something non-human to press your booking button.
How far can an AI agent get on your site today?
Most agents can read a page, fill a form, and click a button. Where they get stuck is usually not model capability but your site's design. Here are the failure points we see most often.
- Prices and stock levels that exist only inside a banner image (for example, a graphic reading “September only — $99”)
- A booking flow split across four or five steps, with a popup at each step
- Phone or messenger inquiries only, with no path that can be completed on the web
- Essential details — hours, shipping fees, cancellation policy — living only in social posts
- Key information hidden behind a login wall or captcha
If you run a hair salon, “Cut $30, no deposit, free cancellation up to 24 hours ahead” has to exist as text for an agent to answer on your behalf. If it's only in an image, the agent quietly drops you from the shortlist.
What should you check right now?
You don't need a redesign. Pick one core action and work through the list below; most sites can clear it in half a day.
- Choose a single core action — booking, quote request, or add-to-cart — and make it completable on one page.
- Expose price, stock, hours, and shipping/cancellation policy as text, and add structured data (Product, Offer, LocalBusiness, FAQPage).
- Give every form field a real label and name attribute. Placeholder-only fields leave agents guessing.
- State your agent traffic policy in robots.txt and your terms. Treat training crawls and user-delegated actions as separate decisions.
- Split your server logs by user-agent and review weekly. Where agents stop is your improvement backlog.
- Keep a human confirmation step right before an order or booking is final — email verification or additional card authentication.
Does this create new risks?
Yes. Agents make mistakes faster and in larger volumes than people do. As the door widens, the locks need attention too.
- Bulk orders for out-of-stock items: validate inventory server-side, not just in the UI.
- Coupon code guessing: enforce one-per-account limits and cap failed attempts.
- Contact form spam: rate limiting plus email confirmation fits the agent era better than captchas.
- Search API abuse: cache heavy queries and cap response sizes.
Finally, run a 15-minute test. Give an agent tool your domain and ask it to “book tomorrow at 2 p.m. here.” Wherever the agent stalls is almost always where human customers drop off too. Getting agent-ready is really just the work of building a clearer site for people.