Two Kinds of Website Now, and You Only Control One
Personal AI agents SEO stopped being a thought experiment the moment Meta shipped Muse to everyone over 18 in the US and Canada, free for most uses. That is the part most coverage skips past. The traffic already exists, and your analytics stack almost certainly has no idea what to call it.
The web is splitting into two kinds of sites. One has a deal with the agent. The other gets a visit from an agent wearing your customer’s clothes, no name tag, and a Chromium user agent string that reads like any other shopper. In my own cPanel access logs, a session like that has all the familiar furniture: a Chrome UA, a cloud provider IP range, a viewport, a checkout POST. Nothing about it announces that no person is behind the keyboard.
StatCounter’s numbers give the direction, even if the timing is guesswork. Computers carried 98% of worldwide web traffic in January 2010 and fell under half by 2026. Human-driven browser sessions look like they are on the same curve, with agents picking up the share that browsers lost.
The contradiction that follows is what makes this awkward to operate. Your logs show a session and an order confirmation. Your revenue report shows nothing, because the buy happened inside a connector and only ever appeared in the platform’s own order records. You cannot reconcile books that were never meant to touch.
Here is the shape of what actually changes, and it is narrower than the general panic suggests. Robots.txt has no name to target, so crawl directives lose most of their leverage. Page structure, meaning headings, link text, alt text, pricing in visible text, labeled form fields, and a checkout path that does not depend on hover states, becomes the thing that decides whether the agent finishes the task or bounces. Analytics gets a visitor that is not a person, which breaks attribution quietly rather than loudly. And on Shopify, a single default setting under Sales channels files you into one tier or the other without you choosing.
Slobodan Manic covered this in detail on Search Engine Journal, working from Meta’s own documentation for Muse, and that piece (opens in new tab) is where I am pulling the vendor quotes from. He is right that the searching part of the job is familiar ground. The buying part is not, and that is the gap most SEO teams are still staffed for.
Background: How Agents Reach Websites
There are two tiers, and which one you land in is not really up to you right now.
The first is the connector. Meta launched Muse alongside what it calls “dozens of partners” with access to the entire Shopify catalogue, with Walmart, Best Buy, Sephora and Wayfair in the queue. For those initial partners, Meta says it worked closely with each provider to integrate their API. OpenAI’s version is the plugin ecosystem, which it says connects to over 4,000 apps, and any business can build one and submit it for review. The connector is the good tier. The agent gets structured data, account access, and a reliable way to complete the transaction without squinting at your theme.
The second tier is the one most of us are actually in. Without a connector, the agent arrives in a cloud browser because a person gave it a task. Meta describes Muse running “a real up-to-date Chromium-based browser” on “a dedicated secure computer.” OpenAI says Dots uses “their own cloud computer, their own browser, and the apps you’ve connected.” That is a real browser rendering your real page, clicking your real buttons.
The analogy I keep coming back to is a trade counter versus a shop floor. A connector is the trade counter: you hand over a manifest, someone behind the desk pulls the SKU, done. The browser tier is a customer walking onto the shop floor and working out your layout from scratch. If the price is only visible after a hover, or the size selector needs a JavaScript event that fires on mousemove, that customer does not complain to a manager. They leave.
Shopify merchants may not get to pick. Under Sales channels, then Agentic, the default is “Allow Shopify to manage for me,” which enrolls the store in Shopify’s AI shopping channels, and Shopify states that products are “shared with Meta by default through Shopify Catalog.” That default is a connector decision made on your behalf, quietly, in a settings pane most store owners have never opened.
If your site runs on a custom stack, there is no such pane. You are browser tier until you build and ship an API integration that a vendor accepts.
What’s Happening Now
The control everyone reaches for first, robots.txt, does not have a name to target. Meta’s page for website owners lists five crawlers, and Muse is not one of them. OpenAI’s page lists four, and Dots is not on that list either. No user agent string, no request header, no signature has been published that would let a sysadmin write a rule against these agents. If there is nothing to match on, there is nothing to disallow.
What Meta has said is more unsettling than a missing crawler entry. In its own documentation, Meta describes what a site sees when Muse visits: “When Muse browses the internet, it will appear as your activity, so if you ask Muse to buy a shirt from a clothing designer’s website, that designer might use your visit to show you an ad on Instagram.” The agent arrives wearing the customer’s session, because functionally it is the customer’s request being carried out.
The analogy that fits is a courier showing up at your warehouse with a signed collection note. The paperwork has your buyer’s name on it. You load the pallet. Nobody at the dock asks whether the person holding the clipboard is the person who signed, because the signature is what the process checks. Robots.txt has always worked the same way: it is an honour system keyed to a name, and this visitor declines to give one.
The other thing happening now, and the one I think SEO teams have thought about least, is that the agent finds your site like a human does. By brand name, or through a search. Neither Meta nor OpenAI has confirmed which search engine its agent queries, which means your rankings still function as the discovery layer for agent traffic. You cannot optimise for the arrival, but you absolutely can optimise for being in the set of pages the agent lands on.
Underneath all of this is a quiet break in the standard. The robots exclusion protocol states that its rules “are not a form of access authorization.” It never was enforcement; it was a crawler choosing to comply. Agents in browsers make that distinction visible in a way crawlers never did.
The source article by Slobodan Manic for Search Engine Journal goes further than the news angle and maps both tiers, the connector economics, and where agent sessions surface in reporting. It is worth reading in full, here (opens in new tab).
What It Means in Practice
Can you block personal AI agents SEO-wise with robots.txt?
Not by name, and that is the entire problem. Meta’s page for website owners lists five crawlers, and Muse is not on it. OpenAI’s lists four, and Dots is not on that one either. Until somebody publishes a user agent string, a User-agent: line has nothing to match against. A wildcard Disallow throws out signed-in Chrome users along with the agent, because the agent is running a current Chromium build and behaving like a browser.
I have had this conversation more than once since the first GPTBot wave. A client asks me to just block AI traffic, we add the rules, they feel better, and the numbers do not move, because the traffic they were worried about never announced itself in the first place. Blocking a named crawler and blocking an agent driving a browser are different jobs. Only one of them has a tool.
So your realistic control sits in what the page actually contains. That is not a defeat, it is just where the work moved.
On-page work that decides whether the agent can act
An agent reads the page the way a careful stranger would, with less patience. The facts it needs have to be in the visible text: what the product is, what it costs, what size or spec it comes in, whether it is in stock. Headings, link text, alt text, and spec tables all still carry weight. So does the boring stuff of naming things properly.
Two things I check on client sites now, both of which break agents and humans alike:
- Labels and buttons that say what they do. A checkout button rendered as an icon with a
divwrapper, or form fields using placeholder text instead of a real<label>, leave an agent guessing. So do date pickers that only exist after a click. - A path through the funnel that does not depend on JavaScript state. If the cart total or the delivery estimate only renders after three chained events, the agent may never see a confirmed price.
This is front-end territory that SEO teams are inheriting by default. The person who owns the checkout markup now owns whether the agent can finish, and the person who owns the title tag owns whether the agent arrives at all. Rank work has not gone anywhere, since the agent still finds you through search or a typed brand name.
The practical reading is that blocking is not on the table, but legibility is.
What to Expect Next
My expectation is that the connector tier grows and the browser tier settles into being the fallback rather than the default way agents reach a site. The incentives push that way. A connector gives the agent structured product data, real stock counts, and a checkout that returns a clean order instead of a screenshot of a confirmation page. Every retailer that signs one makes the browser route look clumsy by comparison, and retailers notice when their competitor’s catalogue answers faster.
What I genuinely do not know is the timing, or whether OpenAI follows with the same level of public documentation. Meta has described how Muse browses, what it sees, and how connectors are supposed to improve on that. OpenAI has said far less about Dots. If that gap holds, treat Meta’s Muse pages as the closest thing to a specification for agent behaviour that exists right now, and read them even if you will never touch a Meta product. Both systems are solving the same problem with the same two tiers, so the mechanics transfer.
The part that unsettles me is who picks the tier on your behalf. On Shopify, the default setting under Sales channels, then Agentic, is “Allow Shopify to manage for me.” A merchant who has never opened that screen is already enrolled in AI shopping channels with products shared through Shopify Catalog by default. A merchant on a custom platform sits in the browser tier whether they chose it or not, unless someone builds an API integration. Platform defaults are quietly making an architectural decision that used to belong to the site owner.
Analytics is where I expect the loudest breakage. Most stacks still treat a session as a person with an intent, a device, and a journey through the funnel. An agent session resembles all three until you try to segment on it, and then it does not. I expect vendors to ship some way of flagging or filtering non-human sessions within a year, and until they do, attribution stays wrong at the top of the funnel without anyone noticing. The mismatch usually surfaces late, when channel mix stops lining up with the order records in the platform.
One mundane thing to expect: your server logs will show a browser user agent that looks like an ordinary recent Chrome build, because the agent runs a real Chromium instance. No amount of log reading will hand you a clean “MuseBot” string to filter on, at least not while Meta and OpenAI stay silent about naming themselves.
The useful posture is to assume any session you cannot identify is a customer you cannot interview, then make the page legible enough that the interview is unnecessary.
Next Steps for Site Owners and SEO Teams
The audit I would run this week is embarrassingly manual. Open your three highest-revenue pages, turn JavaScript off, and read what is left. If the price, the size options, the delivery estimate, and the button text survive that, an agent has a fair chance of acting on the page. If those facts only appear after hydration or behind a hover state, the agent gets an empty shell with your logo on it.
Then go through the forms. Every input needs a real <label for="..."> or an aria-label that names the field, not a grey placeholder that vanishes on focus. Placeholders are the most common failure I find on client sites. An input whose only clue is the word “Email” fading out on click gives an agent nothing to work with, and guessing is how a shipping address ends up in the phone number field.
Shopify merchants can settle the tier question in about ninety seconds. Admin, Settings, Sales channels, Agentic. Look at what is shared and which catalog it flows through. If you run a custom stack, treat yourself as browser tier until someone builds an integration, and budget front-end time accordingly.
On the server side, keep raw user agents and full request paths longer than your default retention period. You cannot segment what you threw away, and the day non-human sessions become a measurable slice of traffic you will want something to compare against. WAF rules that flag a burst of sequential POSTs to checkout from one address are worth writing now, even if you leave them in monitor mode for a month.
Skip the robots.txt research. There is no published user agent for the Muse or Dots browser to target, and the Search Engine Journal writeup on these agents (opens in new tab) is clear that neither company has given site owners a name to allow or block. You can spend a fortnight guessing at strings, or you can spend it on labels and headings.
Ranking work still applies, since the agent arrives through search the same way a person does. Keep publishing and keep the technical basics tidy.
Start with one product page and one checkout flow. Strip both to plain text, and fix whatever stops making sense first.