Checkout vs. AI agents
This article was drafted with AI assistance and reviewed by our team before publishing.
AI shopping agents — autonomous tools that browse the web, compare products, and complete purchases on behalf of real consumers — may already be visiting your store. The problem is that your checkout may not have been built with these agents in mind. And when an agent fails to complete a purchase on your store, no error is logged, no support ticket is raised, and no frustrated customer sends you an email. The sale simply disappears.
Two Kinds of Agentic Checkout
Before getting into what breaks, it helps to understand that "agentic commerce" actually describes two quite different things — and only one of them involves your checkout page. **Protocol-based agentic checkout** works through a direct merchant integration. ChatGPT's Instant Checkout, built on the Agentic Commerce Protocol, is one example: the agent communicates with a merchant's backend via a structured protocol and does not touch your storefront's HTML at all. If a consumer uses this kind of agent to buy from you, and you have the integration in place, your checkout page is bypassed entirely. **Browser-based agents** work differently. They open an actual browser, navigate your pages, read what is on screen, and fill in forms — much like a human would. OpenAI's Operator and Anthropic's Claude both have this capability, completing multi-step tasks on a user's behalf by operating a real browser session. Google is building similar agentic infrastructure into Gemini and AI Mode. The checkout-page problems described in this article apply to browser-based agents. Protocol-based integrations are a separate implementation question entirely.
AI Shopping Agents: Available Now, Adoption Still Early
Browser-based agents act as proxies for real consumers who have delegated a purchase decision to an AI — "order me more of that face wash" or "find the cheapest version of this product and buy it." When the agent succeeds, the consumer gets what they wanted without lifting a finger. When it fails, the consumer often doesn't find out why, and neither do you. A survey published in December 2025 by payments vendor Checkout.com found that 47% of respondents said they planned to use an AI agent for Christmas shopping, with the report suggesting 2026 "could be" the year agentic commerce goes mainstream. It is worth noting that the survey measures stated consumer intentions rather than observed transaction volumes, and no sample size was disclosed — so treat it as a directional signal rather than a verified baseline. Real-world shopping usage by AI agents is not yet well measured. What is clear is that these tools are available, the major AI labs are actively building and releasing them, and some consumers are already using them for shopping tasks — though how routinely and at what scale remains an open question. For direct-to-consumer stores in the $500K–$5M range, that is not a distant trend to monitor. It is worth understanding today.
How Browser Agents Read Your Pages
This is where it gets technically important — and where a common oversimplification can mislead you about what actually needs fixing. Some browser-based agents rely on the page's **accessibility tree** — a structured representation of the page that browsers expose to assistive technologies — to understand and interact with your checkout. SEO and technical accessibility writers have noted that the accessibility tree is how AI agents read your site, and that broken or missing labels in that tree directly impair these agents' ability to complete forms. Others use **screenshots and visual reasoning**, interpreting your page more like a human looking at a screen. Label problems affect these two categories of agent differently: an agent using visual reasoning may successfully identify an unlabelled field from its visual context, while an agent relying on the accessibility tree will not. The practical implication is that fixing your accessibility tree — properly labelling fields, managing focus, structuring your page semantically — meaningfully reduces failure risk for accessibility-tree-dependent agents, and does no harm to agents using visual reasoning. It is the higher-leverage fix. Computer-using agents, such as OpenAI's Operator at its launch, can operate a browser visually. Claude's computer-use features work similarly. Neither is purely dependent on the accessibility tree, but neither is immune to the structural failures described below.
Why Your Checkout May Fail Agents
Here are the specific failure points that cause silent drop-off in browser-based agentic sessions: **Unlabelled form fields.** A CVC field that has placeholder text reading "123" but no proper HTML `<label>` element looks fine to a sighted human. To an agent reading the page's accessibility tree, it is ambiguous or invisible. Even agents using visual reasoning may misinterpret fields that lack clear textual labels. The agent may skip the field, enter data incorrectly, or abandon the form. **CAPTCHAs and bot-detection challenges.** reCAPTCHA and similar tools are designed to block non-human sessions. An AI agent is, by definition, a non-human session acting on behalf of a human. Invisible CAPTCHAs that trigger on behavioural signals — mouse movement patterns, typing speed, session duration — can flag agent traffic and block it. The agent hits a wall and no purchase is made. **Payment iframes that break keyboard navigation.** Many payment widgets load inside an iframe — a separate embedded document within the page. Keyboard navigation must flow into the iframe, through all its fields, and back out to the main page without getting stuck. This is also a WCAG 2.1 requirement (No Keyboard Trap): users and agents must never be unable to move focus away from an embedded component using standard keyboard controls. If your payment iframe does not correctly release focus after its final field, agents — and keyboard-dependent users — will stall at the payment step. **Non-standard or dynamic step sequences.** Agents handle predictable, linear flows better than unpredictable ones. A checkout that injects unexpected modals, requires interaction with conditionally appearing elements, or restructures the page mid-flow based on input creates unpredictable behaviour that agents handle poorly. These are not exotic edge cases. They are extremely common in storefronts built on templates, assembled from third-party plugins, or updated incrementally over time without systematic accessibility reviews.
The Hidden Revenue Drain
Here is why this matters more than a standard UX issue. When a human encounters a broken checkout, something happens: they get frustrated, they might try again, they might contact support, and in the worst case they abandon and you lose one sale. That failure has a signal attached to it — a cart abandonment, a support ticket, a session recording in your analytics tool. When an agent fails, none of that happens. The agent simply stops, the task is marked as incomplete or fails silently, and the consumer may not even know a purchase was attempted. There is no signal. No session recording of a confused user. No spike in support contacts. The revenue just doesn't arrive. At low volumes of AI-referred traffic, this may be a rounding error. But if browser-based agent usage grows as the major AI labs are betting it will, a checkout with structural agent-compatibility problems becomes a revenue leak that compounds quietly without ever appearing in your dashboard as a labelled problem.
How to Audit Your Checkout for Agent Compatibility
The good news is that fixing agent compatibility is largely the same work as fixing accessibility — and accessibility improvements benefit all users, not just agents. Here is where to start. **Run your checkout through an accessibility audit.** Tools like Axe, WAVE, or a Lighthouse audit in Chrome DevTools will surface missing form labels, focus management issues, and iframe problems. These are free and can be run without deep technical knowledge. Any critical or serious issue flagged by these tools may be an agent failure point. **Test your checkout using only a keyboard.** Open your store, start a purchase, and complete the entire checkout without touching your mouse. If you get stuck anywhere — a field you can't reach, a button you can't activate, a modal you can't dismiss — an agent relying on keyboard navigation will get stuck there too. Tab through every field in order and confirm each one is clearly labelled in the page structure, not just visually. **Review your bot-detection setup with your vendor.** If you use reCAPTCHA or a similar challenge, talk to your payment provider or fraud-tool vendor about how their system handles legitimate agentic sessions. Do not reduce fraud protections based on referral data alone — referrer signals can be spoofed, and loosening controls without proper safeguards can increase card-testing risk. Your vendor is best placed to advise on any agent-compatible authentication options they support. **Confirm your payment iframe handles focus correctly.** Ask your developer or payment provider to verify that your payment widget correctly receives focus when a user or agent tabs into it, and correctly releases focus when they tab past the last field — without trapping them inside the iframe. If your provider doesn't support this, it is worth raising as a support request or considering alternatives that do. **Test with a structured, step-by-step walkthrough.** Go through your checkout and document every single step in sequence, including any conditional steps — address validation prompts, discount code fields that appear after a minimum order, upsell modals before payment. Then consider: would an agent following this sequence encounter anything unexpected? Anything that requires a visual cue without a text equivalent, or an interaction pattern outside a standard form submit? None of these fixes require a full rebuild. They are targeted and specific, and in most cases something a developer familiar with accessibility standards can implement without a major project. The investment is modest. The return — reliably capturing a growing category of transactions that currently disappear without a trace — compounds in the other direction.