Where Agent Journeys Break: What Happens When Your Visitor Is a Robot

There's a version of the AI commerce conversation that stops too early. It goes: get the AI to mention you, get recommended, done.

Being mentioned is not the transaction. A shopper who receives a glowing recommendation for your product and then lands on a page where the agent helping them cannot work out which size to pick is a sale you won and then dropped. Those losses are worse than a ranking miss, because you paid all the acquisition cost and got none of the revenue.

This post is about the second half: what actually happens when a non-human tries to get through your store, and how to find out where yours breaks.

First, an honest word on timing

Fully autonomous agent checkout is not settled. Agent payment rails launched across 2025 (OpenAI's Agentic Commerce Protocol with Stripe, Google's UCP, agent payment products from Visa and Mastercard), and then in March 2026 OpenAI pulled back its Instant Checkout in ChatGPT after a five-month beta, shifting toward merchant-run checkout instead.

So: anyone telling you that autonomous agents are buying at scale today is overselling. Mostly, agents research and refer, then hand the transaction to a human.

Here's why that doesn't make this a 2028 problem. The handoff is exactly where you lose people. An agent that cannot resolve your variant, cannot confirm your price, or cannot get past a step doesn't hand the shopper to you. It hands them to a competitor whose flow it could navigate. That is happening today, at volume, and it is measurable today. The rails settling later just raises the stakes on something already costing you money.

The seven places it actually breaks

In rough order of how often they come up.

1. The variant wall

The most common failure, by a distance. Size, colour, capacity, condition, bundle. A human sees a swatch grid and picks. An agent sees a set of controls it must operate correctly to reach a specific SKU, and if the wrong choice means the wrong product, an agent that isn't sure will typically stop rather than guess.

Worst offenders: swatches whose only label is a colour image, size selectors that disable unavailable combinations without saying why, and "select options" flows where price only appears after a valid combination is chosen.

2. Price that requires execution

If the price appears only after client-side rendering, then from the outside your product doesn't have one. Any question with a budget in it, which is most of them, quietly excludes you. Same story for stock status fetched asynchronously: unknown availability reads as unavailable.

3. Buttons that aren't buttons

An <div onclick> styled to look like a button is fine for a mouse and invisible as an action to something reading structure. Add-to-cart is the expensive place for this to happen.

4. Interstitials

Cookie banners, age gates, region selectors, newsletter modals, "are you shopping from the right country" prompts. Each one is a small tax on a human and a hard stop for something that doesn't know that the grey box over the page is dismissible and unimportant.

5. Flows that assume a human "just knows"

A form field labelled only by a placeholder that vanishes on focus. A two-step address entry where step two only appears after a valid postcode. A required field that isn't marked required and only reveals itself as an error after submission. Every one of these is a small piece of tribal knowledge your regular customers absorbed and an agent has no access to.

6. Bot defences that can't tell the difference

CAPTCHAs, rate limits, and WAF rules that block anything non-human. This one is genuinely hard, because those defences exist for good reasons. But it's worth knowing whether your protection is currently turning away shoppers with intent, because right now most merchants have no idea which side of that line their configuration sits on.

7. Cart and checkout state that doesn't survive

Sessions that expire mid-flow, carts that empty on navigation, checkouts that require a login the agent can't complete. The furthest-along failures and the most expensive, because everything before them worked.

Run the test yourself

The best thing about this problem is that you can reproduce it without buying anything. You can be the agent.

Pick five products. Your best seller, one with lots of variants, one that's nearly out of stock, one at a price point where budget filtering matters, and one from a category you'd like to grow.

Write one instruction per product, phrased like a shopper would: "Buy the cheapest waterproof hiking boot in size 10 that ships this week." Not a click path, an outcome.

Point an off-the-shelf browser-automation or computer-use agent at your live store with that instruction. Don't help it. Don't pre-clear the cookie banner. The friction is the finding.

Record four things at the point it stops: which step, what it was trying to do, what it saw instead, and a screenshot. That last one matters more than you'd think: a recording of an AI shopper failing on your own store is remarkably persuasive to people who control roadmap.

Then do the same on two competitors. If they pass where you fail, you've found not just a bug but a live transfer of revenue.

An afternoon. Five products. You'll have a list.

What to do with the results

Sort by two things: how far along the failure was, and how much traffic that product gets. A break at checkout on your best seller is worth ten breaks on a long-tail product's spec tab.

Then split the list. Some failures are structural: a div that should be a button, a modal that traps focus, a checkout step that needs a session. Those go to engineering, and they're usually small.

The rest are comprehension failures: the agent got where it needed to be and still couldn't work out which variant was which, or whether the thing was waterproof, or what it cost. Those aren't front-end bugs. Those are the page not stating what it needs to state, and they're the majority.

That second pile is what Product Gale repairs automatically, and then re-tests to prove the run gets further than it did before. The first pile still needs your developer, but at least now they'll have a replay instead of a hunch.

The point

Your store was built on an assumption that held for twenty-five years: the visitor is a person, with eyes, patience, and context.

A growing and disproportionately valuable share of your visitors now have none of those. They will not squint at a photo to infer waterproofing, will not intuit that the grey overlay is dismissible, and will not guess at a variant. They will go somewhere that doesn't require it.

The only way to know whether that's happening to you is to try it. Go run the five-product test.

---

Agentic checkout timeline: OpenAI's Agentic Commerce Protocol and Instant Checkout beta (2025), Instant Checkout shift reported by Digital Commerce 360, March 2026.