Meta’s Muse and OpenAI’s Dots are personal AI agents that, at a user’s request, can research products, fill out forms and, in some cases, even add items to a cart and complete a purchase on the web. Muse is available to people 18 and older in the United States and Canada, and is said to be free for most uses. Businesses are entering a phase in which it is no longer enough to be found in search: their sites must also be easy for agents to navigate.
AI Is Starting to Do More Than Read Websites—it’s Starting to Use Them
Jumping between 10 tabs to compare options, make a reservation and reorder something: tasks people have long done manually are about to be replaced by a single prompt to AI. Both Muse and Dots are agents that carry out requested tasks on the web. Meta has disclosed details about how Muse works, while OpenAI has shared little about Dots.
It was not so long ago that computers accounted for nearly all web browsing. The arrival of smartphones changed the balance of power. According to StatCounter, desktop computers’ share of global traffic fell from 98% in January 2010 to below 50% in 2026. Could the same thing happen between human-operated and agent-operated browsers? That possibility is the starting point for this article.
There Are Two Ways to Reach Businesses
Agents can reach businesses in two broad ways. One is through integrations, also known as connectors. Meta says it lined up dozens of partners when it launched Muse and gave it access to Shopify’s entire product catalog. Walmart, Best Buy, Sephora and Wayfair, among others, are expected to join in the future. OpenAI says its plugin system connects to more than 4,000 apps, and businesses can also build their own and submit them for review.
The other route is through a browser, which can reach the vast majority of sites that don’t have connectors. According to Meta, Muse runs a Chromium-based browser on a dedicated secure computer. Dots is also said to work with its own cloud computer and browser. In other words, an agent asked to “buy a shirt from this designer” visits the site relying only on what appears on the page.
For businesses using Shopify, platform settings may determine where they stand. Shopify says the Agentic setting under sales channels defaults to “Let Shopify manage it,” and product information is shared with Meta by default through Shopify Catalog. Sites built on their own infrastructure fall into the browser-based category unless they create a connector.
Robots.txt Doesn’t Work on Visitors Who Don’t Identify Themselves
This is where traditional bot defenses start to falter. Robots.txt files and crawler allowlists are designed for visitors that identify themselves. But neither Meta nor OpenAI has disclosed a user agent, request headers or signature for Muse’s or Dots’ browser. Meta lists five crawlers for website operators, and OpenAI lists four; neither list includes these agents. There is no name to add in the first place.
Meta says that when Muse browses the web, its activity appears to come from the user. In the shirt-buying example, the seller might even use the visit to serve the user ads on Instagram. It brings to mind the protagonist of the movie *Catch Me If You Can*, walking through an airport in a pilot’s uniform. If the uniform looks authentic, no one checks whether he’s qualified to fly. Muse, in effect, arrives wearing the customer’s uniform.
Robots.txt is a gentlemen’s agreement to begin with, and the standard itself explicitly says it is “not a means of access authorization.” Amazon’s robots.txt file lists 99 agent names, but neither Muse nor Dots is among them. Amazon blocked Muse in September, using its terms of service. According to GeekWire, there were three reasons: Meta had not told Amazon that Muse would access its store; Muse does not identify itself when browsing; and it appeared to obtain and store customers’ login credentials.
The decision point is shifting from robots.txt to bot-mitigation systems. From the perspective of firewalls and platform bot filters, what they see is a cloud-based Chromium browser. Whether to let it through as a human or block it may come down to someone’s settings, without the SEO team’s knowledge. HUMAN Security says its bot-detection product can identify Muse. One option is to ask your bot-mitigation vendor whether it can distinguish between human-operated Chromium and an AI agent.
Dots also has another route. OpenAI's documentation says that if a site blocks its cloud browser, users should try it on their connected computer. In that case, it explains, approved clicks, form submissions, and actions performed while logged in may be treated as activity from the user's own account.
Agents read pages through the accessibility tree
So how do agents that arrive through a browser see a page? According to Meta, Muse's browsing subagent sees a snapshot of the accessibility tree, not the page's raw DOM. It also cannot execute JavaScript on the page. The page's own scripts run in the real browser, but Muse cannot add anything of its own—that is how the setup works.
The accessibility tree contains information used by screen readers. Browsers generate it from the DOM, giving each element a type (such as a button, link, or heading), a name, and a state. It does not include the design. OpenAI has not disclosed how the Dots browser reads pages. Google says browser agents may use screenshots, DOM structure, accessibility trees, and other information.
### How each element appears
heading_placeholder
Comments