AI search optimization has been about getting cited inside AI answers. The next layer arriving now is AI agents acting on your website on a user's behalf, booking reservations, comparing products, filling forms, completing purchases.

The two require different work. Citation work earns you a mention in an AI answer. Agent-readiness lets an autonomous agent actually use your website to complete a task, without a human ever loading the page.

Google's developer guidance has made the shift official. The web.dev article "Build agent-friendly websites" tells developers they now have "a new type of visitor" and recommends building for AI agents alongside humans. The Model Context Protocol (MCP) hit 110 million monthly downloads in 2026 per Anthropic, faster than React reached the same milestone. Chrome's WebMCP entered its origin trial in May 2026, backed by Google, Microsoft, Booking.com, and Shopify.

The piece below covers what agent-readiness actually means, the four protocols emerging around it, the practical checklist Google published, and how to think about this alongside the citation work you already do.

What "agent-ready" actually means

An agent-ready website is one an autonomous AI agent can read, understand, and act on programmatically, without scraping pixels or guessing at layout.

The framing matters because it sits one step beyond GEO. Generative engine optimization aims to get your content cited inside an answer. Agent-readiness aims to let an agent execute on your site after the user reads that answer. The two work together but solve different problems.

Three ways an agent reads your website, per Google's web.dev guidance:

• Screenshots through a vision model: The agent takes a snapshot of the rendered page and a vision model identifies elements like buttons, search bars, and form fields.

• Raw HTML and the DOM: The agent parses the page's HTML, reading semantic structure, attribute names, and the relationships between elements (a "Buy Now" button inside a product container belongs to that product).

• The accessibility tree: The browser-native API that distills the DOM into roles, names, and states of interactive elements. Google calls it the page's "semantic summary," and notes it functions as a high-fidelity map that ignores visual styling and focuses on functional intent.

The accessibility tree is the most important of the three. Built for assistive technology like screen readers, it now doubles as the cleanest machine-readable representation of your site for AI agents. Google's own line: "Everything we suggest to make a site 'agent-ready' also makes sites better for humans."

That overlap is the strategic detail most marketers miss. Agent-readiness is not a new tax on top of existing work. The discipline is largely the WCAG accessibility playbook restated for a second audience.

The four protocols emerging in the agent layer

A handful of standards are forming around how agents discover, read, and act on websites. None of them is settled yet, but the pattern is consistent enough to plan around.

The four worth knowing:

• MCP (Model Context Protocol): Anthropic's open standard, launched in 2024, for connecting AI models to external tools and data sources. MCP hit 110 million monthly downloads in 2026. Microsoft's CTO Kevin Scott has compared its role to HTTP for the agentic web.

• WebMCP: A web-native implementation of MCP, championed by Google and others, that lets websites expose structured commands directly to browser-based agents over HTTPS. Chrome 149 began an origin trial in May 2026.

• UCP (Universal Commerce Protocol): Google's protocol for cross-platform shopping and checkout, anchoring the agentic commerce work Google expanded heavily at I/O 2026. Our breakdown of what UCP means for SEO strategy and the March 2026 update covering Cart, Catalog, and Identity Linking walk through the mechanics for ecommerce brands.

• AP2 (Agent Payments Protocol): Google's payment-agnostic protocol announced in September 2025, designed for multi-agent configurations. Compatible with both MCP and UCP, intended to make agent-led payments work safely across providers.

The four overlap by design. MCP and WebMCP define how agents read and act on tools. UCP defines how agents interact with commerce. AP2 handles the payment layer. Together they sketch the protocol stack for what Google and others are calling the agentic web.

The honest caveat: every one of these is early. WebMCP is in origin-trial. AP2 is months old. UCP is rolling out by geography and vertical. No marketer should bet their roadmap on any single protocol winning. The right move is to understand the direction, prepare the foundations that work across all of them, and watch which standards consolidate.

Google's seven-rule agent-friendly checklist

Google's web.dev guide, authored by Kasper Kulikowski and Omkar More and last updated April 1, 2026, ships seven concrete rules for making a site agent-friendly. Each rule maps to an existing WCAG accessibility practice, which is the point.

1. Use semantic HTML

Use the right HTML element for the right purpose. Buttons should be <button>, links should be <a href>, headings should be <h1> through <h6>. Agents (and screen readers, and search engines) infer purpose from semantic tags. A <div> styled to look like a button is invisible to all three.

2. Keep layouts stable

If an agent identifies an "Add to cart" button at time T and the page shifts because a banner ad loaded at T+1, the agent clicks the wrong element. Cumulative Layout Shift was a Core Web Vital ranking factor. For agents it becomes a blocking issue. Reserve space for ads and embeds, set explicit width and height on images, and avoid injecting content above existing content after first paint.

3. Use the cursor:pointer signal

Interactive elements should change the cursor to a pointer on hover. Agents use this signal to recognize what is clickable. Tailwind v4 introduced a default that strips cursor:pointer from native button elements, which silently breaks this rule on sites that upgraded from v3. The fix is a three-line base style.

4. Link form labels to inputs properly

Every input field needs an explicitly linked label. Either wrap the input in a <label> element or use the for attribute pointing to the input's id. Unlabeled inputs are functionally invisible to agents trying to fill a form.

5. Hit minimum interactive size

Touch targets should be at least 24x24 pixels (WCAG AAA recommends 44x44). Smaller targets fail for both accessibility and agent vision. The fix is padding, not just hitbox tricks.

6. Avoid ghost overlays

Invisible elements that block clicks (carryover modals, hidden overlays, decorative div layers) break agent interactions. If something is interactive, it should be visible. If it is not interactive, it should not intercept clicks.

7. Show visible state changes on action

When an agent clicks a button, the page should provide visible feedback (a loading state, a confirmation, a state change). Without it, the agent has no way to know its action registered, which leads to repeated clicks or abandoned tasks.

The whole list is the WCAG accessibility playbook restated for a second visitor type. Brands that already maintain accessibility have done most of the work. Brands that have not just got a second business reason to do it.

How to prepare your site, in priority order

The work breaks into three layers. Do them in this order.

Foundation: fix the accessibility tree

If your site fails an accessibility audit, it fails agent-readiness. Run Lighthouse 13.3's new Agentic Browsing audit category (available in Chrome DevTools), note the pass ratio, and start with the highest-impact failures: unlabeled buttons, missing form labels, alt text gaps, layout shift. Each fix improves accessibility, SEO, and agent-readability at once.

The step is unglamorous and uncontroversial. The accessibility-tree work is also where 80% of the gains live for most brands.

Identity: ship clean structured data

Standard schema.org markup remains worth the effort, even though it does not directly cause AI citations. For agent flows, schema is how an agent identifies your products, your prices, your inventory, your reviews. For commerce especially, Product, Offer, and Organization schema are the difference between an agent confidently recommending you and an agent guessing. Our research on why brands show up differently on every AI platform shows how unstable identity already is across platforms; clean structured data is what reduces that variance.

The Ahrefs May 2026 study on schema is worth holding alongside this. Adding schema to existing pages did not produce a measurable lift in AI citation share on its own. That finding is correct for the citation scope. The finding does not change the agent scope, where schema is part of how an agent reliably reads what your product is and what it costs.

Protocols: prepare for, do not bet on, the emerging stack

For most brands today, the right protocol decisions are watchful and minimal. Sign up for the WebMCP early preview if your developers have capacity to experiment. If you are ecommerce on Shopify or one of the launch UCP partners (Nike, Sephora, Target, Ulta, Walmart, Wayfair), explore the UCP integration scope. If your business depends on agent-led payments, AP2 is worth tracking, especially in concert with MCP-based agent flows.

For everyone else, the foundation and identity layers carry the load. Agent traffic is not yet large enough that protocol-specific work earns immediate return, but the foundation work compounds across whichever protocols win.

How this connects to citation work

Agent-readiness does not replace GEO. The two layers stack.

A user asks ChatGPT or Gemini for a recommendation. GEO determines whether your brand gets cited in the answer. Once cited, the user (or an agent acting for them) goes to your site to complete a task. Agent-readiness determines whether that task succeeds.

The brands that win the agentic web have to do both. Get cited in the answer, then be usable by the agent that arrives to act. Brands that do only the citation work risk being recommended but unable to convert when the agent shows up. Brands that do only the agent-readiness work risk being usable but never recommended. The case studies of brands building durable AI visibility on original content show how the citation foundation pays off; the agent-readiness layer is what cashes it in when an agent arrives.

The good news for marketers: most of the agent-readiness work is accessibility work, and most of the accessibility work is already being done somewhere in the organization. The job is to make sure it ships and stays shipped, not to invent a new discipline.

Don't wait for the agent traffic to show up

Agent visits will not appear in your analytics as a clean separate channel for a while. Most agents either spoof as regular browsers or get filtered out as bots. Once agent traffic gets large enough to show up in a dashboard, the brands that prepared early will already have the foundations in place. The brands that did not will be retrofitting a year late.

The foundation is accessibility, semantic HTML, stable layouts, and clean structured data. None of that is new. All of it pays off across human users, search engines, and the agents that are now starting to read your site too.

The cleanest first step is a clear view of where your site stands across citation visibility and agent-readability, so the foundation work targets the right gaps. See how Passionfruit's GEO service builds AI search visibility on top of a solid SEO foundation, look at the cross-platform citation tracking inside Passionfruit Labs, and talk to the team about preparing your site for both layers.

Frequently asked questions

What does it mean for a website to be agent-ready?

An agent-ready website is one that AI agents can read, understand, and act on programmatically to complete tasks like booking, comparing, or buying. Google's web.dev guidance identifies three ways agents read sites: screenshots through vision models, raw HTML and the DOM, and the accessibility tree. Agent-readiness is the next layer beyond GEO citation work, focused on letting agents execute on your site rather than just earning a mention in an AI answer.

How is agent-readiness different from GEO and AI citation work?

GEO and AI citation work focus on getting your content mentioned inside AI-generated answers in tools like ChatGPT, Perplexity, Gemini, and AI Overviews. Agent-readiness focuses on what happens after the citation: whether an autonomous agent can use your website to complete a task on a user's behalf. The two layers stack. Citation work earns the recommendation; agent-readiness lets the agent act on it. Brands that win the agentic web do both.

What protocols matter for AI agents in 2026?

Four protocols form the emerging stack. MCP (Model Context Protocol), Anthropic's open standard launched in 2024, hit 110 million monthly downloads in 2026. WebMCP is a Google-championed web-native implementation, in Chrome origin trial since May 2026. UCP (Universal Commerce Protocol) is Google's cross-platform commerce protocol, expanded at I/O 2026. AP2 (Agent Payments Protocol) is Google's September 2025 payments standard, compatible with both MCP and UCP. None is fully settled. The right move is preparing the foundations that work across all of them.

How do AI agents actually read my website?

In three ways, per Google's web.dev guidance. First, screenshots, where the agent takes a snapshot of the rendered page and a vision model identifies elements. Second, raw HTML and the DOM, where the agent reads the page's structure and the relationships between elements. Third, the accessibility tree, the browser-native API that exposes roles, names, and states of interactive elements. The accessibility tree is the most important of the three, which is why agent-readiness overlaps heavily with WCAG accessibility practices.

What is the easiest first step toward agent-readiness?

Run an accessibility audit. Use Lighthouse 13.3's Agentic Browsing category in Chrome DevTools to scan your homepage and a key product or service page. Note the pass ratio. The top failures are usually the same: unlabeled buttons, missing form labels, alt text gaps, and layout shift. Fixing them improves accessibility, SEO, and agent-readability simultaneously, which makes the work pay off across three audiences at once. Roughly 80% of the gains for most brands live in this foundation layer before you ever touch protocols like WebMCP or MCP.

Do I need to implement MCP or WebMCP on my website now?

For most brands, not yet. WebMCP is in origin trial as of May 2026, and the early adopters are large platforms (Microsoft, Booking.com, Shopify) and developer-led products. Sign up for the early preview if your team has capacity to experiment. For ecommerce brands on Shopify or in the UCP partner program, the commerce protocols are worth exploring sooner. For everyone else, the foundation work (accessibility, semantic HTML, structured data) compounds across whichever protocols ultimately consolidate, and is the higher-return investment until the standards settle.