Can an AI Assistant Buy From Your Site Without a Human? Test It

You can run the ai checkout test in the next twenty minutes, and the first time you do it, it stings. Open a capable AI agent, the kind a customer would actually delegate an errand to, point it at your own site, and give it the real job: "order two of these and check out," or "request a quote for this," or "book the first available slot." Then sit back and watch. Not a survey, not a report. Watch an agent try to complete the exact thing you sell, on the exact pages you sell it on, and see how far it gets before it stops. Most owners assume it sails through, because people buy from that site every day. Most owners are wrong in one specific spot, and the spot is the point.

Here is what the test tends to look like. The agent does well, better than you expect. It reads the product, understands what you offer, finds the thing the customer asked for, gets it into the cart or starts the quote form. You watch it work and you relax a little; this is going to be fine. Then it reaches one step, just one, that it cannot operate. It pauses. It tries again. It tries a slightly different way. And then it gives up and tells the customer it could not finish. The cart was full. The quote was half-typed. The sale was sitting right there, already won, and it evaporated at a single step that every human clears without thinking, because a human has eyes and hands and the agent has neither.

That is the whole frontier in one sentence: a sale you have already earned can be lost at a step an agent cannot operate, and you will never see it leave.

Can an AI agent complete a purchase on your site?

The only honest way to know is to test it. Point a capable AI agent at your site, ask it to complete your core action, and watch where it stops. Common dead-ends are unreadable pickers, human-only steps, and unlabeled fields, and the lost sale leaves no trace.

This is not a hypothetical you can answer from your chair. You cannot reason your way to it by looking at your own checkout, because you already know how to use it; your eyes fill in everything the agent cannot. The test exists precisely because the failure is invisible to the person who built the page. You have to hand the wheel to something that cannot see, and watch.

A quick, honest caveat before the method, because the hype here is loud and mostly wrong. Agents are not doing the bulk of buying and booking yet. Anyone who tells you the robot-shopping era has fully arrived is selling you something. What is true is narrower and more useful: the behavior is real, it is spreading, and the part of it that should worry you is not the volume. It is the silence. When an agent fails on your site, there is no bounce in your analytics, no abandoned-cart email, no angry message. The customer simply hears "I could not finish that" and moves to a business the agent could finish with. You are not losing a visible fight. You are losing one you never knew started.

How to run the test this afternoon

You do not need special software, a developer, or a budget. You need a capable AI agent and your own site. The agent might be powered by Claude, it might be an assistant built into a major platform, it might be one of several others now sitting in a customer's pocket. From your side of the glass the brand barely matters; what matters is that it is an AI agent acting the way a delegated customer would. Pick one and give it a real task.

Keep the steps plain:

• Choose your core action. The one thing a customer most often comes to your site to do. Book the appointment, buy the item, request the quote, reorder the part. Not "browse." The thing that ends in money or a committed booking.
• Tell the agent to complete it, like a customer would. Phrase it as an errand, not a question. "Add two of the medium to the cart and check out." "Request a quote for a kitchen install." "Book me the earliest slot this week." You want it to act, not to describe.
• Watch every step, and do not help. This is the hard part. The instinct is to nudge it past a stuck spot the way you would nudge a confused relative. Do not. The spot where you want to reach in and help is the exact spot where the sale dies when no human is there. Let it fail.
• Write down where it stopped. The first place the agent stalls is the first place an agent picks you and then loses you. That single line is the most valuable thing the test produces.

Run it twice if you like, once for the purchase and once for the booking or quote, since those flows fail in different places. The goal is not a clean pass. The goal is to find the wall, because the wall is real and it is costing you quietly.

The dead-ends agents tend to hit

When the agent stops, it almost always stops at one of a small, recognizable set of steps. None of them look broken. All of them work fine for a person. That is what makes them dangerous: they pass every human test you have ever run, and fail the one you never ran. Here are the ones that show up most on a typical small-business site.

• A control it cannot read. A date or time picker rendered as a grid of bare cells, a slider, a styled set of boxes with no readable names and no machine-readable mark on which options are open. A person recognizes the shape and clicks. The agent has nothing under the appearance to act on, so it cannot choose.
• A step that needs a human action. A "call to confirm," a "tap to add," a quantity stepper or a cart control that only responds to a precise tap or drag, a step that quietly assumes a mouse and a hand. The agent reaches it and finds no path it can take.
• A form field with no label it can match. A box a person knows is "delivery address" only because of where it sits on the page, with nothing in the markup naming it. The agent cannot tell what to type where, so the form stalls before it submits.
• A gate that asks the agent to prove it is a human. An "are you human" challenge, an image puzzle, a verification step built to stop bots that also stops the legitimate agent a real customer sent. It does exactly its job, and its job is now sometimes wrong.

Each of those is a single attribute of a single step. A step stops an agent when it depends on reading something only the eye can see, or requires an action only a hand can make, or carries no label the agent can match, or gates on proving a human is present. One attribute, one wall, one lost sale. You are not looking for a broken site. You are looking for the one step on an otherwise-fine site that an actor without eyes cannot get through.

Picture it concretely, the same checkout run by the two visitors you actually have.

Nothing in that checkout is broken. It converts fine for the people who reach it. It simply cannot be operated to the finish by the visitor a growing share of customers now send ahead, and the business attached to it never learns it was first choice.

Why the lost sale is silent, and what that means

This is the part worth sitting with, because it inverts the reassurance most owners lean on. "My checkout works fine" is true and beside the point. It works fine for people, who improvise past anything ambiguous and push through a clumsy step because they want the result. An agent does not improvise and does not push through. It reaches the step it cannot operate and it leaves, and because no human was ever sitting at that checkout, there is no human-shaped trace of the leaving.

A person who fails on your site might call, might come back tomorrow, might leave a review you can learn from. An agent that fails on your site tells the customer "I could not," names a competitor it could finish with, and moves on within seconds. Nothing lands in your analytics, because the customer never arrived to be counted. There is no abandoned cart, because the agent abandoned it on its own side of the glass. The sale did not bounce. It was never recorded as having shown up.

So the danger is not that agents are buying everything today. They are not. The danger is that the share which is real fails quietly, blamed on the season or the market, when some of it is a step on your own site that an agent could not operate. The test turns that invisible leak into a written line you can act on.

Where the fixes point, and where to take them next

Once you have your line, the spot where the agent stopped, the direction of the fix is not mysterious, even if the depth of it is. The aim is to turn your core action from a path only a sighted person can walk into a path an agent can walk too: every step readable, every field labeled, every control operable without a human-only tap or call, no gate that mistakes a customer's agent for an attacker. Not flashier. Not rebuilt. The same flow, made completable by an actor that works from structure and text and the states the page actually exposes, rather than from appearance.

I am keeping this at "what to aim for" on purpose, because the doing of it is its own discipline and it has a proper home. How a step becomes readable, what real structure versus visual-only looks like, how a quantity control or a date picker gets a label and an observable state, how the whole flow becomes your core action designed as a path an agent can actually traverse, is covered in full there. The test in this post is the consumer-facing, run-it-yourself version that shows you the wall. The guide is how you take the wall down.

It also helps to place this test in the larger picture, because completing the action is the last link in a chain, not the only one. Before an agent ever tries to check out, it has to choose your business at all, and what makes an AI agent pick you over the business down the road is its own short chain of find, weigh, and complete. This test is that final link, completion, made runnable; the pick-and-weigh depth lives there, and I will not re-teach it here. Earlier still is the framing that the first visitor to your site is increasingly an AI agent arriving in place of the customer, working your pages before any person looks. That post sets up the "agent as visitor" picture; this one runs the specific test of whether that visitor can finish the buy or the booking.

There is a fair question of who does this work once the test shows you the wall. Some owners will read their own flow against the dead-ends above and fix it themselves; plenty of it is visible once you know to look. Others will find the steps a customer completes by reflex are exactly the ones an agent stalls on, and that making them operable is real, unglamorous plumbing. Making a small business's core action operable by the agents now arriving is a service like any other, and the point of running the test first is that you can now name exactly which step is failing instead of paying to fix a flow that was never the problem.

So do the one thing that turns worry into a list. Open an agent this afternoon, send it through your most important action, and write down the first step it cannot finish. That line is your starting point. Take it to the agent-legibility guide above, fix the step, and run the test again. The sale you save is one you already won and were about to lose without a sound.