AI Browsers and Agent Experience (AX) — Can an Agent Actually Use Your Website?

For two decades we optimized websites for two audiences: a human with a mouse and a crawler reading HTML. A third one has just arrived — an agent that operates your site like a human: it clicks, scrolls, fills forms and completes tasks, only on the user's behalf and without the user's patience. In AI crawlers I covered who reads your site; in JavaScript SEO — what they see without rendering. This post answers the third question nobody was asking a year ago: can an agent actually DO something on your site — find a product, add it to the cart, submit a contact form. Because if it can't, you are losing conversions that no "abandoned cart" report will ever show you.

Within twelve months we got Comet from Perplexity (free worldwide since October 2025), Claude for Chrome and ChatGPT Atlas — and in July 2026 OpenAI announced it is retiring Atlas and folding agentic browsing directly into ChatGPT. Browser brands come and go, but the capability stays: an agent that clicks, fills forms and completes tasks on your site on the user's behalf. Crawlers only needed readable HTML — an agent has to be able to ACT. What Agent Experience (AX) is, what most often blocks agents (captchas, modal walls, div-buttons, unlabeled forms) and how to test your own site with an agent in 30 minutes.

Let's name the discipline: Agent Experience (AX) — designing a site so that tasks can be completed not only by a person, but also by a machine acting on their behalf. The good news: 80% of the work overlaps with solid accessibility (a11y) and semantic HTML you should have anyway. The bad news: the remaining 20% is exactly what UX checklists miss, because a human works around those flaws intuitively — and an agent doesn't.

A year of AI browsers — and what the Atlas U-turn teaches

The timeline is short and instructive. July 2025: Perplexity ships Comet — a browser whose agent takes over the tab and executes tasks; since October 2025 Comet is free worldwide. August 2025: Anthropic launches Claude for Chrome — a research preview of an in-extension agent, initially for 1,000 Max plan subscribers, with safety front and center. October 2025: OpenAI unveils ChatGPT Atlas — a full browser built around ChatGPT, with agent mode for paid plans. And in July 2026 — the turn: OpenAI retires Atlas as a standalone browser and folds agentic browsing into ChatGPT itself and Codex; Atlas stops working in early August.

Sound familiar? It should — it's the same pattern I described with the Instant Checkout pivot in agentic commerce: interfaces and brands are variable, the capability is durable. Nobody in their right mind "optimizes for Atlas" — you optimize for the capability of agentic action on your site, whether the agent lives in a browser, an extension, a chat or an operating system. Within a year that capability went from demo to product in all three major AI ecosystems at once, and nothing suggests it will roll back.

How an agent sees and operates a site — and how it differs from a crawler

The differences are fundamental and overturn some classic SEO intuitions:

• A crawler reads, an agent acts. Googlebot fetches a document and indexes content. An agent executes a sequence of actions: find, click, type, submit, verify the result. Its "success" is not an indexed page — it's a completed task.
• Robots.txt doesn't apply to it in the classic way. A browser agent operates inside a logged-in user's session, with their cookies and permissions — from the server's perspective it's almost a regular visit. Managing crawlers via robots.txt and bot verification still matters, but it's a different layer: there you control reading, here you serve acting.
• An agent renders full JavaScript — unlike most AI crawlers, which don't execute JS. The paradox: a site invisible to the ChatGPT crawler can be operable for the ChatGPT agent — and conversely, a heavy SPA the agent renders for 20 seconds can blow the task's time budget.
• An agent reads the page through the DOM and the accessibility tree — button names, field labels, element roles, messages. Screenshots help, but it's DOM semantics that decide whether the agent knows what to click. Exactly the same structures a screen reader consumes.
• An agent is expensive and impatient. Every step means model calls — seconds and real money. Platforms impose limits and safeguards (the Atlas agent couldn't run code or download files). When an agent gets lost, it doesn't "click around" like a frustrated human — it aborts the task and reports failure. The user sees: "I couldn't do it on this site." There goes the conversion.

The mechanics are the same as what I described from the automation side in Computer Use — only the roles reverse: there you send an agent to other people's interfaces, here someone else's agent comes to yours.

AX is UX for machines — seven principles

1. 1.Semantic HTML is your site's API. A button is a `<button>`, not a `<div onClick>`. A link is an `<a href>`. A field has a `<label>`. Agents locate elements by role and accessible name — a div styled like a button has role "generic" and an empty name.
2. 2.State always as text in the DOM. Validation errors, submission confirmations, out-of-stock notices, loading states — if the message exists only as a border color or an icon, the agent can't read it. Text plus proper attributes (aria-invalid, aria-live for dynamic messages).
3. 3.ARIA patterns wherever a component is custom. Tabs, accordions, comboboxes, dialogs — the W3C ARIA Authoring Practices Guide provides ready-made semantics. This is not an "accessibility extra" — it's the communication protocol for every machine client of your site.
4. 4.Content and actions reachable without hover or gestures. Hover-only menus, tooltip-only content, long-press actions — for an agent that content doesn't exist. Everything important must be reachable by click and keyboard.
5. 5.Predictable navigation and stable names. An agent plans: "contact is probably in the footer or menu". Creative labels ("Let's talk about your journey" instead of "Contact") raise the cost of every task. In information architecture, boring is a virtue.
6. 6.Performance is the task budget. An agent waits for network and rendering at every step — a slow site multiplies that by the number of steps. Solid Core Web Vitals — especially INP, the response to interaction — is not a "ranking factor" in AX; it's the condition for finishing the task at all.
7. 7.Don't deceive the agent. Hiding instructions for models in invisible text is prompt injection — platforms actively defend against it (researchers demonstrated vulnerabilities in Comet and the industry responded by hardening safeguards), and detected manipulation means losing the system's trust with no support ticket to win it back.

What blocks agents — the removal list

Agent testing keeps surfacing the same short list of culprits:

1. 1.Captcha on the task path — a wall for the agent; platforms deliberately hand control back to the human. Keep captchas where they prevent real abuse (registration, payment), remove them from search, filters and the contact form — less invasive protections suffice there.
2. 2.Modal walls: a cookie wall covering the content, a newsletter popup after three seconds, "rotate your phone", a chat widget gnawing at the corner. A human closes them reflexively; an agent must identify and handle each one — and a modal without an accessible close button can end the task.
3. 3.Div-buttons and unlabeled forms — AX principles 1 and 2 in their "how to lose a customer" form. Add placeholder-as-label fields (the placeholder disappears once typing starts — the agent loses context) and red-border-only validation with no words.
4. 4.Infinite scroll with no alternative — an agent won't scroll forever; it needs pagination, search or filters to reach anything beyond the first screen.
5. 5.Content locked in canvas or images — a price list as a JPG, a configurator living entirely in canvas: a black box to the DOM.
6. 6.Fragile sessions — carts lost on navigation, logouts on return, forms wiped after a single field error. An agent steps through flows more slowly than a human; a site that loses state along the way loses the agent's task.
7. 7.Payment and strong authentication — 3-D Secure, SMS codes, biometrics. This block is correct: it's where the human handoff should happen. Just make sure the interrupted flow can be resumed after authorization without losing the cart.

An agent-proof lead form and checkout

Two flows where AX pays back fastest. The lead form: complete `<label>` coverage, `autocomplete` attributes (name, email, tel, organization — the agent fills the user's data without guessing), validation as text next to the field, and after submission — a confirmation as text in the DOM with a concrete promise ("Thank you, we reply within 24h"), not just a toast that vanishes in two seconds. Checkout: guest checkout without forced accounts, an address form aligned with autocomplete, prices and shipping costs visible before the final step (an agent comparing offers treats surprise costs the way a human does — it abandons), stock status as text. The transactional layer — product feeds and the ACP/AP2/UCP protocols — is covered in agentic commerce; AX is its browser-side complement: where no protocol integration exists, the agent simply walks through your checkout by hand.

Test your site with an agent — the 30-minute protocol

Don't guess — measure. You need access to any browser agent (agent mode in ChatGPT, Comet) and a list of 5 tasks that are money for your business:

1. 1.Information: "Find the opening hours / scope of service X / return policy on [domain]". The simplest task — and it fails surprisingly often because of modals.
2. 2.Product: "Find [product] in size M under $50 in the store and add it to the cart".
3. 3.Lead: "Send a quote request for X via the contact form" (on a staging environment or clearly marked as a test).
4. 4.Comparison: "Compare [your product] with [competitor's] and recommend one" — you'll see which data the agent finds on your site and which it pulls from elsewhere.
5. 5.Navigation: "Book a meeting / find the nearest branch / download the spec PDF".

For each task note three things: did it succeed, where the agent hesitated (repeating or backtracking steps signals unclear semantics), and what it told the user at the end — because that report ("I couldn't submit the form on site X") is literally a review of your website delivered to your customer. Cross-check with data: in-app browser sessions and AI referrals are already covered in AI traffic analytics. Fix the three most frequent blockers first — usually the modal, the captcha and the form.

---

I audit websites for agents end to end: agent testing of key flows, semantics and accessibility review (a11y = AX), blocker fixes and impact measurement in the logs. I do this as part of technical SEO and web engineering. Get in touch — I'll start with a 30-minute agent test of your site: five tasks, a list of blockers and fix priorities.

Worth reading next:

• AI Crawlers — Who Scans Your Site and How to Manage It
• JavaScript SEO — What Bots See Without Rendering
• Agentic Commerce — How to Sell When the Buyer Is an Agent
• Computer Use — AI That Operates Any App Like a Human