AI agent transactions arrive on 200M+ Android phones via Chrome auto-browse in late June 2026. What makes a site agent-transactable — not just agent-readable — and how to audit yours.
For two years, "AI visibility" meant one thing: does an LLM cite your site when someone asks? That question is about to get a more expensive sibling. Starting late June 2026, Google's Chrome auto-browse lands on Android at the operating-system level, default-on for everyone with a Pixel 10 or Galaxy S26, with Google's stated path reaching more than 200 million devices by the end of the year. When an agent shows up on a user's phone to book the appointment, the question is no longer just whether it found you. It is whether it can finish.
The old visibility question (does an AI assistant mention you in its answer?) measured whether you exist in the model's worldview. The new one measures whether you can be operated. An agent that cites you but can't complete your checkout sends the user a recommendation. An agent that can complete your checkout sends you a sale.
That gap is the whole story. If your site can be read but not driven, you don't lose a citation, you lose the conversion that the citation used to lead to. The agent reads three roofing companies, picks the one whose booking form it can actually fill, and books it. The other two never find out they were in the running.
App-level agents reach the people who chose to install them. A ChatGPT app or a Perplexity Comet browser is opt-in, which kept agentic transactions in early-adopter territory through 2025 and early 2026. An agent baked into the operating system reaches everyone who bought the phone. There is nothing to download and nothing to enable.
That is what shifts late June 2026 from a product launch into a distribution event. The Pixel 10 and Galaxy S26 are the first wave, and Google has said the same capability extends to watches, cars, glasses, and laptops across the rest of the year. The audience for "can an agent transact on you" jumps from a sliver of power users to a meaningful share of your actual mobile traffic, on a default setting, almost overnight.
Two details say this is real capability rather than a demo. Google's underlying agent work, Project Mariner, scored 83.5% on the WebVoyager benchmark for completing real web tasks. And the feature is metered like something people use for high-value work: the AI Pro tier runs $19.99/month for 20 agent tasks a day, and AI Ultra runs $249.99/month for 200. People paying by the task tend to delegate the tasks that matter: the bookings, the orders, the reservations.
Most "AI-ready" work to date optimized for one bar: being readable. Clean meta descriptions, server-rendered content, a tidy llms.txt, structured headings an assistant can quote. Readable means an agent can fetch your page and understand what it says. It is the bar that earns citations.
Transactable is a higher bar. It means a non-human operator can complete a task on your page: submit the form, pick the slot, reach the confirmation screen. A site can be perfectly readable and completely un-transactable, and until this year that was fine, because nothing was trying to operate it. Chrome auto-browse is the thing that starts trying, at scale, by default.
Auto-browse uses Gemini 3's multimodal model to read a page, identify what is on it, fill forms, navigate the flow, and complete the transaction. Google has not published the exact pathway, but it combines vision with DOM access and accessibility-tree reads. The practical translation: the agent operates your real website the way a user does, faster and without anyone tapping. So readiness is not about adding a protocol. It is about whether a careful, non-human operator can drive the DOM you already have.
There is a 30-second test for this, and it is worth running before you read another word of strategy. Open your booking or checkout flow in Chrome on a phone. Disable JavaScript in dev tools. Reload. Can you see the form, see the buttons, and finish the task with the keyboard alone? If yes, the agent can too. If the page goes blank or the flow breaks, you have work to do.
The failure modes that stop an agent mid-task are mostly old accessibility sins wearing a new consequence:
<label> or an aria-label, or the agent can't tell the phone field from the ZIP field.<div onclick> styled to look like a button often isn't operable; use a real <button> or <a>.The encouraging part: every one of these maps to an existing WCAG accessibility recommendation. The work you do to let a screen-reader user complete your booking is very nearly the same work that lets an agent complete it. Accessibility stopped being only a compliance line item and became a conversion path.
If you have read about WebMCP, you might expect it to be the answer here. It isn't the answer for auto-browse, and being clear about that saves you from chasing the wrong fix. WebMCP lets a site register tools an agent can call directly, a cleaner and more reliable path than driving the DOM. But Chrome auto-browse does not invoke WebMCP today; it reads and operates the page visually. The agents that do invoke WebMCP (Comet, some browser extensions, custom buying agents) are a smaller, opt-in set.
So WebMCP is the forward "sell" layer, not the thing that makes you transactable to 200 million auto-browse phones this month. It has real value: where an agent supports it, tool calls are faster and don't break when you change your CSS, and the snippet no-ops harmlessly on browsers that don't. Ship it as the next layer once the DOM is clean. Just don't mistake it for the layer with the June deadline.
Crawlytics frames AI-readiness as three jobs, and they happen to be the right priority order for this moment too.
Detect — are agents already here? Before you change anything, find out whether AI agents and crawlers are already fetching your transactable pages, and which ones. If auto-browse and its cousins are hitting your checkout, you have an immediate stake. The AI citation and bot-tracking guide walks through reading this from your traffic.
Serve — clean DOM plus a map. This is the layer with the deadline. Fix the eight blockers above so the agent can operate your flows, then ship llms.txt so it knows what your site offers and which pages are worth acting on. Readable and operable, together.
Sell — the protocol path. Once the DOM is clean, add the WebMCP commerce snippet so the agents that prefer a direct tool call can take one. It is the efficiency upgrade, valuable as adoption grows and free of downside where it isn't supported yet.
Run them in that order. A WebMCP snippet on a site whose checkout breaks without JavaScript is a lock on a door that won't open.
Late June 2026 is a matter of weeks, not a roadmap entry. The desktop preview shipped in January, the Android version was announced May 12, and the OS-level rollout begins this month. The sites that capture the first wave are the ones that fixed their DOM blockers before the agents arrived, because an agent doesn't give you a second attempt. It either completes the task or moves to the competitor whose page let it.
You don't need to predict the agentic web to act on it. Run the disable-JavaScript test on your single most valuable flow today. Whatever breaks is your list.
Written by Crawlytics Team. Crawlytics tracks AI bots, generates llms.txt, and powers WebMCP commerce, all from one snippet on any stack. See how it works →
AI agent transactions are tasks an AI assistant completes on your website on a user's behalf (booking an appointment, submitting a lead form, or placing an order) rather than just reading or citing your page. The shift matters because being mentioned by an AI assistant only creates value if the agent can then finish the action the user wanted. As of late June 2026, with Chrome auto-browse rolling out on Android, completing the transaction becomes part of what "AI visibility" is worth.
Chrome auto-browse is Google's agentic browsing feature that uses the Gemini 3 model to read web pages, fill forms, navigate flows, and complete transactions automatically. It begins rolling out at the operating-system level on the Pixel 10 and Samsung Galaxy S26 in late June 2026, default-on rather than opt-in, with Google's stated path reaching more than 200 million Android devices by the end of 2026 and expanding to watches, cars, glasses, and laptops over time.
Run a quick manual test: open your most important flow (checkout or booking) in Chrome, disable JavaScript in dev tools, reload, and try to complete it using only the keyboard. If you can finish, an agent likely can too. If the page breaks, look for the common blockers: client-side-only rendering, unlabeled form fields, div-based buttons, cookie-wall traps, CAPTCHA, and slow dynamic loads. Each maps to a WCAG accessibility fix.
No. Chrome auto-browse operates your site by reading and driving the page visually through the DOM and accessibility tree, the way a person would. It does not require WebMCP or any custom API. WebMCP is a separate, optional path that lets some agents (Perplexity Comet, certain extensions, custom buying agents) call your site's tools directly, but it is not what makes you transactable to auto-browse. Get a clean, accessible DOM first; treat WebMCP as a later efficiency layer.
Largely, yes. Google has noted that its agent-readability rules map to existing WCAG recommendations, and the failure modes that stop an agent (unlabeled inputs, non-semantic buttons, keyboard-inoperable flows) are the same ones that stop assistive technology. Investing in accessibility now does double duty: it serves users with disabilities and makes your site operable by the AI agents arriving on hundreds of millions of phones.
This page is part of Crawlytics.app. View all pages: llms.txt · llms-full.txt