The Checklist for Making Your Site Readable to AI Agents

On September 8, 2026, Meta launched Muse, an AI agent that opens a browser and does the task for you. ¹ It can search, navigate sites, fill out forms, and complete bookings and purchases. ² The catch is that its browser does not see your site the way you do.

The checklist below has 11 items that help this new visitor understand your site. Each item tells you what to check, how to check it in a minute, and where the rule comes from. If you came for the list alone, it is also ready to copy.

What Meta launched on September 8

Muse runs on its own computer, in the cloud, with its own browser. Meta calls that environment Muse Secure VM. ¹ You talk to it in the app, on the muse.ai website, or in WhatsApp. ³

Two design choices change how it relates to your site. It keeps working after the person closes the app. And it stops short of anything sensitive: it checks with the person before sending an email or making a purchase. ¹

For now, the rollout is limited to the United States, on iOS, Android, and the web, with AI glasses announced for later. ¹ Basic use is free, and there are paid plans at $20 and $100 a month. ³

Meta is not the first tech giant with a shopping agent. Google introduced the Universal Commerce Protocol on January 11, 2026, and Microsoft launched Copilot Checkout three days earlier. ⁴ What is different is reach: Meta’s family of apps averaged 3.6 billion daily active people in June 2026. ⁵ That figure covers the whole family of apps. Muse itself is available only in the United States today.

How the agent reads your page

Think of the floor plan of an apartment. It does not show the paint color or the couch. It shows where each door and each window is, and the name of every room.

The browser builds a plan like that for every page it opens. The plan includes headings, links, buttons, form fields, and the major regions of the page. The visual design stays out. Each item in the plan has a role, like button or heading, plus a name, a state, and a description. ⁶ This is called the accessibility tree.

This tree is how a page reaches assistive technologies, like the screen reader, which reads the page aloud for people who cannot see it. ⁷ And it is what Muse reads.

The wording comes straight from Meta, in the engineering post it published alongside the launch. The browser sub-agent “sees an accessibility tree snapshot of the page”, and Meta is explicit that this is “not the raw DOM”. The sub-agent also has “no ability to run javascript in the page context”. ⁸ In plain terms, what reaches it is a snapshot of the plan. The raw code of the page stays out.

The reason Meta gives is security. A sub-agent that only sees the plan does not touch the page structure and does not see passwords. ⁸ For a site owner, the consequence weighs more than the reason: whatever is not in the plan does not exist for this reader.

One caveat for technical readers. The source says the agent does not run JavaScript in the page context. That means it does not inject scripts to dig out content. What your own JavaScript draws on the page is another story, because the page is built by the browser inside the virtual machine. No source says that content disappears for the agent, and I will not say it either.

This is not a Meta matter alone. Chrome’s documentation for agents states: “Agents rely on the accessibility tree as their primary data model.” ⁹ And it puts both audiences in a single sentence: “missing labels can block users with visual disabilities and agents from completing a task”. ¹⁰ The label missing for someone who cannot see is the same label missing for the agent.

Here is the same product page, seen both ways. We built this example ourselves, in simplified form. It is not a capture from Muse.

Now the agent picks the site

When it books travel, Muse takes two routes, and Skift tested both. Flights go through a direct connection to Duffel, a travel infrastructure company, with live inventory from more than 500 airlines. Hotels go through the browser. The agent visits sites like Expedia and Hotels.com and compares options the way a person would. ¹¹

On the hotel route, the agent is the one deciding where to search, what to compare, and which site to book on. ¹¹

Products tell the same story. According to a reading of the code by Juozas Kaziukėnas, reported by PPC Land, Muse has two ways to find products. One browses the web, “slow and fragilely but almost anywhere”. The other pulls from Shopify’s structured catalog, “exact and universal”. ⁴ That reading is secondhand, because the original post was not accessed.

One detail carries weight here. Meta says there is no advertising inside the product. ¹² The journalist who reported it added a fair caveat: for now, this is company policy, and nothing in the system guarantees it will last. ¹² So far, nobody has documented a way to pay your way into the agent’s decision. What is left is what the agent can read and understand on your page.

That is why SEO has a new reader. SEO is not only about being found by people who search. It is about being understood by whoever buys on their behalf. Meta made a point of designing one detail in: when it is time to pay, the agent stops and asks the person to confirm. ¹ This new reader is the subject of the SEO for AI silo.

The checklist, item by item

Almost everything here gets checked in the same place. Open the page in Chrome and open DevTools. In the Elements panel, go to the Accessibility tab and turn on Show accessibility tree. The tree of the whole page appears where the code used to be. ⁷

If you do not work on the site yourself, that is fine. Copy the ready-made version and send it to whoever does.

What gets clicked and what gets filled in

1. Every action is a real element.

• What to check: a button is a button, a link is a link, a dropdown is a dropdown. In code, that means <button>, <a href>, and <select>.
• How to check it in a minute: in the tree, select the Buy button. It has to show up with the button role. If it shows up as a generic block, the site drew something that only looks like a button.
• Why: Chrome’s documentation lists Tree integrity among the critical checks for agents, and that group verifies that roles are valid. ⁹

2. Every button and every link has a name that says what it does.

• What to check: icon-only buttons included, like the heart for saving a favorite or the magnifying glass for search.
• How to check it in a minute: in the tree, every button and every link shows up with a name. “Buy now” is a name. Empty is not.
• Why: the first group of Chrome’s checks for agents is Names and labels. ⁹ And the failure is common: WebAIM found empty links on 46.3% of the home pages it analyzed, and empty buttons on 30.6%. ¹³

3. Every form field has a real label.

• What to check: name, email, ZIP code, quantity. Each field needs its label tied to it in the code.
• How to check it in a minute: in the tree, the field shows up with a name. A text box with no name is a field whose purpose nobody knows.
• Why: Muse can “fill out forms”, in Meta’s own words. ² Chrome points out that a missing label blocks people with visual disabilities and blocks the agent. ¹⁰ In the WebAIM data, 51% of home pages have a form field with no label. ¹³

Notice what the three numbers have in common. Empty links, empty buttons, and unlabeled fields are exactly what the tree lacks when someone needs to know what to click. WebAIM measures accessibility, and no one has yet measured how many sites Muse fails to understand. Even so, the gap is the same.

What the page says, in text

4. Every image that carries information has alt text that describes it.

• What to check: product photos, charts, badges, maps.
• How to check it in a minute: in the tree, the image item has a name. “Stridelight 3 running shoe, orange” is a name. “IMG_4032.jpg” is not.
• Why: in the tree, an image is only as useful as its name, and that name comes from the alt text. ⁶ WebAIM found images with no alt text on 53.1% of home pages. ¹³

5. Price, shipping, delivery time, and terms are written as text.

• What to check: none of that information is locked inside an image or a banner.
• How to check it in a minute: look for the price in the tree. If it appears as text, you are fine. If only the banner image appears, the price is locked inside it, as in the table above.
• Why: the tree drops the visual layer. ⁶ Whatever is drawn stays out of the plan.

6. One main heading says what the page is.

• What to check: one H1 with the name of the product or service, and the subheadings in order right below it.
• How to check it in a minute: in the tree, look for the level 1 heading. There has to be one, and it has to state what the page is about.
• Why: headings enter the tree with their role and their level. ⁶ Large type with no heading markup goes in as plain text. To see why the machine needs to know what the page is about, read how Google understands entities.

What sits underneath the page

7. Nothing clickable is hidden from the tree.

• What to check: a button or link that works on screen but was marked to disappear from the tree.
• How to check it in a minute: Lighthouse flags this in the Agentic Browsing category. The step-by-step is further down.
• Why: Visibility is the third group of Chrome’s checks for agents. Nothing interactive can be hidden from the tree. ⁹

8. Nothing moves while the page loads.

• What to check: a banner that pushes the content down, an image with no reserved size, a notice that pops in on top.
• How to check it in a minute: Lighthouse measures this. The metric is called CLS.
• Why: the agent identifies the button and only then clicks. If the button moved in between, the click lands on something else. Chrome names ads, images without dimensions, and injected content as causes. ⁹

9. The page language is declared.

• What to check: the opening tag of the code states the language. For a US English site, that is <html lang="en-US">.
• How to check it in a minute: right-click the page, choose View page source, and look for the <html> tag at the top.
• Why: it is one of the six most common failures on the web, present on 13.5% of home pages. ¹³ One line fixes it.

10. ARIA only where HTML falls short.

There is a way to patch the plan by hand: sticky notes a developer puts in the code to say what an element is. This is called ARIA.

• What to check: every attribute in the code that starts with aria-. Would a native element, like <button> or <select>, do the job on its own?
• How to check it in a minute: ask whoever runs the site for the list of places where ARIA was used, and why.
• Why: in the WebAIM data, pages with ARIA average 59.1 errors, against 42 on pages without ARIA. ¹³ The number shows a correlation, and nobody has proven cause. Even so, both technical sources point the same way: semantic HTML first, correct ARIA only where there is a real gap. ^{9 6}

Catalog and structured data

11. The structured data repeats what the page shows.

• What to check: if you sell through Shopify, the product is complete in the catalog. If you do not, the structured data on the page carries the same name and the same price the person sees.
• How to check it in a minute: put what is in the catalog, or in the code, side by side with what is on screen. They have to match.
• Why: the Shopify catalog is one of the two product routes found in Muse’s code. ⁴ And what you tell the machine has to be what you show the person. That is a house rule, and it holds for any machine.

Precision matters here, because the subject is easy to confuse. The structured route someone found in Muse’s code is the Shopify catalog. No source says Muse reads the structured data on your page, so this item makes no such promise. It protects the consistency between what the machine reads and what the person sees. Structured data as a topic belongs to the technical SEO silo.

The checklist, ready to copy

Paste it into your task tracker, or send it to whoever runs your site.

CHECKLIST: SITE READABLE BY AI AGENTS
SEO Genome · seogenome.com/en/seo-for-ai/ai-agents/ · September 10, 2026

[ ] 1. Every action is a real element: a button is <button>, a link is <a href>, a dropdown is <select>
[ ] 2. Every button and every link has a name that says what it does, icon-only buttons included
[ ] 3. Every form field has a real label
[ ] 4. Every image that carries information has alt text that describes it
[ ] 5. Price, shipping, delivery time, and terms are in text, and not inside an image or banner
[ ] 6. One main heading says what the page is, and the subheadings follow in order
[ ] 7. Nothing clickable is hidden from the accessibility tree
[ ] 8. Nothing moves while the page loads
[ ] 9. The page language is declared
[ ] 10. ARIA only where HTML falls short
[ ] 11. Selling on Shopify: product complete in the catalog. Not on Shopify: the structured data repeats what the page shows

TEST
[ ] See the tree: Chrome DevTools > Elements > Accessibility > Show accessibility tree
[ ] Run Lighthouse, Agentic Browsing category (experimental, Chrome 150 or later)

How to test your site today

Start with your best-selling page: the product page if you run a store, the service page if you do not.

The first test is the tree. In Chrome DevTools, in the Elements panel, open the Accessibility tab and turn on Show accessibility tree. The tree of the whole page takes the place of the code, and you can expand each item to see its role and its name. ⁷ Run all 11 items on that page, one at a time.

The second test is Lighthouse. It now has a category called Agentic Browsing. The category groups the accessibility checks that matter most to machines into three: Names and labels, Tree integrity, and Visibility. It also measures layout stability. ⁹

Before you trust the result, note three caveats from the documentation itself. The category is experimental. It requires Chrome 150 or later. And it does not give a score from 0 to 100: it shows the fraction of checks the site passed. ⁹

For technical readers: the same report checks whether an llms.txt file exists at the domain root, and it covers WebMCP. Both were left out of the checklist. No source says Muse reads llms.txt, and WebMCP still requires signing up for a Chrome origin trial. ⁹ If you want to go past the test, the next step is the technical SEO silo.

What is still unknown

A three-day-old launch leaves more questions than answers. These are the ones the sources consulted do not settle.

1. Whether Muse reads the structured data on your page. No source says so, and Meta’s announcement names no protocol at all. ⁴
2. Whether the code reading holds up. The two product routes come from a third party’s analysis, reported by one publication. The original post was not accessed. ⁴
3. How many people use Muse. The number has not been published. The 3.6 billion figure belongs to Meta’s family of apps. ⁵
4. How much of the web the agent can read. No one has measured it. WebAIM measures accessibility, which is a different yardstick. ¹³
5. Whether the absence of ads will last. Today it is a stated policy. ¹²
6. When Muse reaches other countries. The sources consulted mention the United States and give no date for any other market. ¹
7. What criteria the agent uses to choose between two sites. Skift shows that it controls where it searches and where it books. ¹¹ The sources consulted do not describe the criteria.

From now on, when you open your product page, ask two questions. Does the person understand what they are looking at? And does the plan the agent reads say the same thing? The first step is hands-on: open your best-selling page, turn on the tree in DevTools, and check item 1.