Plumbing

The AI agent reading your site doesn't say who it is. Here's what still reaches it.

By Arnav Mukherjee, founder of TofuBofu · October 6, 2026

TLDR

Stop planning for named crawlers alone. Amazon blocked Meta's Muse in September, partly because it "doesn't identify itself".
An agent with no name gives robots.txt nothing to match. Amazon reached for its Conditions of Use.
Expect these visits to look human in your analytics. No AI referrer, no utm_source, so no engine attached.
Write pages an agent can operate. OpenAI says its Atlas agent reads ARIA tags to understand a page.
Measure the answers themselves. An answer doesn't depend on a visit you can see.

Our AI traffic tag decides whether a visit came from an AI engine in two ways. It reads the referrer, and if that isn't an AI engine it reads the utm_source parameter. If neither matches, the visit stays an ordinary pageview. That rule is right for every engine that sends a click. It has nothing to say about a reader that doesn't.

In September Amazon blocked Meta's new personal agent, Muse, and gave as one of its reasons that the agent "doesn't identify itself when it browses." An agent that browses like a person arrives with no referrer and no parameter. Our tag counts it as an ordinary visitor. So does any tool that classifies visits by referrer.

What changed in September?

Meta launched Muse on 8 September 2026. Its own announcement describes an agent that runs on "a dedicated secure computer with its own browser" and that "can open a browser, fill out forms, and negotiate" on a person's behalf. According to GeekWire's reading of the launch materials, a service with no public API gets used through a browser "the way you would."

By 20 September Amazon had cut it off. GeekWire's Todd Bishop reported Amazon's three stated reasons:

Meta didn't tell Amazon that Muse would access its store.
The agent doesn't identify itself when it browses.
It appears to capture and store customer credentials. Meta had said previously that Muse has no visibility into people's passwords or payment methods.

Look at the tool Amazon used. Its popup reads: "Continued access by an unauthorized AI agent violates Amazon's Conditions of Use, to which our customers have agreed." GeekWire gives the legal context: Amazon lost its anti-hacking injunction against Perplexity at the Ninth Circuit on 4 August, and that ruling left claims built on contracts and terms of service open. Our read is that robots.txt was never an option either, because Meta's crawler documentation names no user agent for Muse to write a rule against. GeekWire also notes that Amazon's own shopping agent, Buy for Me, identifies itself and lets brands opt out. Amazon is drawing the line at disclosure.

You don't have Amazon's legal department or its terms of service with every visitor. For a B2B firm, the useful lesson is about what this kind of reader can and can't see.

Which three assumptions does an unnamed agent break?

Not every agent hides. Amazon's Buy for Me names itself, and Meta runs a named fetcher for user-requested pages. The ones that don't announce themselves, Muse among them, break three assumptions that most AEO advice rests on, including our own guide to AI crawlers:

The reader announces itself. GPTBot, ClaudeBot and PerplexityBot publish user agents, so you can allow or block them by name. An agent browsing like a person gives you no name to write a rule for.
The reader respects robots.txt. Meta's crawler documentation says its Meta-ExternalFetcher "fetches individual links at a user's request" and "may bypass robots.txt rules." That's a named fetcher. Meta's page doesn't mention Muse at all.
The visit leaves a trail. OpenAI says ChatGPT adds utm_source=chatgpt.com to referral links from its search results, and that parameter is one of the two ways AI referral tracking works. An unnamed agent that types or follows a URL itself adds nothing, unless the link already carried one.

So the agent reads your pricing page, your service page and your contact form, then reports back to a person you never see. Your server logs show a browser. Your analytics show a direct visit, if the agent runs your JavaScript at all.

What does an agent read when it lands on your page?

OpenAI answers this for its own agent, and the answer is plain. Its publisher FAQ says ChatGPT Atlas uses ARIA tags, "the same labels and roles that support screen readers," to "interpret page structure and interactive elements." It tells developers to follow WAI-ARIA best practices "by adding descriptive roles, labels, and states to interactive elements like buttons, menus, and forms."

Read that as the spec for every agent of this kind, because the alternative is a screenshot and a guess. An agent working out how to book a call with you needs to know which element is the booking button. If your button is a styled div with no label, the screen reader gets nothing and so does the agent.

Meta hasn't published equivalent guidance for Muse. We're extending OpenAI's advice to it on the reasoning that any agent operating a browser has to read the same accessibility tree. That's our read, not Meta's statement.

See what AI engines say about you, without waiting for a visit

A free TofuBofu scan asks ChatGPT, Claude, Perplexity, Gemini and Google AI Mode the questions your buyers ask, and records whether you're named.

Run your free scan

What should a B2B firm fix this week?

Start with the pages an agent would be sent to: pricing, services, contact or booking, and any comparison page. On each one:

Use real buttons and links. A <button> that says "Book a 30-minute call" beats an icon in a clickable div.
Label every form field. A visible <label> tied to each input, never placeholder text alone.
Put prices and terms in text. A pricing table drawn as an image can't be read out of the accessibility tree.
Keep headings in order. One h1, then h2 sections that say what's in them.
State who you serve in a sentence. "Managed IT for dental practices in Ontario" is something an agent can match against its brief.

None of this is new, and that's the good news. It's the same work that makes your site usable with a screen reader and readable by a headless browser. SEO is still the floor. This is a layer on top of it, and you can win it whatever your Google rank.

What we'd skip: chasing these agents in robots.txt. Write rules for crawlers that name themselves. For the rest, the page itself is the only thing you control.

Are B2B buyers sending agents yet?

We don't know, and we won't pretend to. The Amazon fight is about shopping carts. Meta did launch Muse for Small Business on 29 September, a version of Muse for people running a business, but it helps an owner run their own operations. It isn't evidence of buyers sending agents to shortlist vendors. We've seen no measurement of a B2B buyer sending an agent to shortlist vendors, from us or anyone else.

Our position, stated as a position: the mechanism doesn't care what's being bought. A founder who already asks ChatGPT for an MSP shortlist will hand the follow-up to an agent the day one does it well. By our estimate, the fixes above cost a developer a day. Waiting for proof costs you the period when nobody else has done them.

It also changes how you should judge any AI visibility number, including ours. A traffic report misses this reader by design, as referral tracking in GA4 already misses a share of chatbot clicks. What doesn't move is the answer itself: ask the engines your buyers use, record whether they name you, and repeat. That's what a TofuBofu scan does, and no visit has to reach your site for it to count.

Frequently asked questions

What is a personal AI agent, and how is it different from an AI crawler?

A crawler such as GPTBot or OAI-SearchBot fetches pages in bulk for training or for a search index, under a published user agent you can name in robots.txt. A personal agent does a task for one person: it opens a browser, reads pages, fills out forms and reports back. Meta describes Muse as running on a secure virtual machine with its own browser that can "open a browser, fill out forms, and negotiate" on a person's behalf.

Why did Amazon block Meta's Muse agent?

GeekWire reported on 20 September 2026 that Amazon cut Muse off from shopping on Amazon.com. Amazon gave three reasons: Meta didn't tell Amazon that Muse would access its store, the agent "doesn't identify itself when it browses", and it appears to capture and store customer credentials. Meta had said previously that Muse has no visibility into people's passwords or payment methods. Amazon's block message cites its Conditions of Use.

Can I block an AI agent with robots.txt?

Only one that names itself and honours the file. robots.txt works by user agent string, so an agent that browses like a person gives you nothing to write a rule against. Even named fetchers can sit outside it: Meta's crawler documentation says Meta-ExternalFetcher fetches individual links at a user's request and "may bypass robots.txt rules." Treat robots.txt as a control for crawlers that name themselves.

Will my analytics show visits from AI agents?

Not as AI visits. Analytics tools, ours included, recognise an AI visit by the referrer header or by a utm_source parameter such as the utm_source=chatgpt.com that OpenAI says ChatGPT adds to referral links from its search results. An agent that types or follows a URL inside its own browser carries neither, so the visit lands as an ordinary pageview with no engine attached. Our own AI traffic tag counts it that way too.

How do I make my website readable for AI agents?

Build the page the way accessibility guidelines already ask. OpenAI's publisher FAQ says ChatGPT Atlas uses ARIA tags, the same labels and roles that support screen readers, to interpret page structure and interactive elements, and recommends WAI-ARIA best practices: descriptive roles, labels and states on buttons, menus and forms. In practice: real button and link text, labelled form fields, headings in order, and prices and service details written as text rather than baked into images.

Are B2B buyers using personal agents to research vendors?

We've seen no measurement showing it, from us or anyone else. The Amazon fight is about consumer shopping, and Meta's Muse for Small Business, launched 29 September, helps owners run their own operations rather than shortlist vendors. What transfers to B2B is the mechanism: an agent reading your pricing or service page in its own browser, with no name and no referrer. Our read, labelled as a read, is that research agents for business buying follow the consumer ones, and the fixes cost little now.

Does this change how I should measure AI visibility?

It weakens visit-based measurement and leaves answer-based measurement intact. If agents read your site without leaving a trace, a traffic report undercounts AI interest by design. Asking the engines your buyers use directly, and recording whether they name you, doesn't depend on any visit reaching your site, which is why we measure answers rather than sessions.

Sources and further reading

Meta: Introducing Muse, 8 September 2026, read at source 6 October 2026: the agent's own browser, and what it does on a person's behalf
GeekWire: Amazon blocks Meta's Muse AI assistant in new standoff over agentic shopping, by Todd Bishop, 20 September 2026: Amazon's stated reasons and its block message
OpenAI: Publishers and Developers FAQ, read at source 6 October 2026: Atlas reads ARIA tags; ChatGPT adds utm_source=chatgpt.com to referral links
Meta for Developers: Web crawlers, read at source 6 October 2026: Meta-ExternalFetcher fetches at a user's request and may bypass robots.txt
Meta: Introducing Muse for Small Business, 29 September 2026, read at source 6 October 2026: Muse for people running a business
W3C: ARIA Authoring Practices Guide: the WAI-ARIA practices OpenAI points developers to