Agent Experiences

Agent Visualizer

The Agent Visualizer draws the path an AI agent takes through one of your pages: where it arrives, where it reads, where it gets stuck, and everything past that point it never reaches. It’s the same scan as the Agent Analyzer, shown as a path instead of a score.

A real example

Scored 97 of 100

A real scan of our own blog index, taken on October 3, 2026. Two of the steps fall short, and both are ours to fix.

  1. 1Arrive

    Nothing in the way

    Fetch the page and be allowed to keep it.

    Title
    Blog | Agent Experiences
    Language
    en

    The page is fetched and kept. Everything after this is possible because this worked.

  2. 2Read

    Nothing in the way

    Get the words out of the markup, without running a browser.

    Readable words
    923 in the page
    Inside <main>
    862 words

    The words come out of the markup without running a browser, so every agent gets the same text a person would.

  3. 3Separate

    Nothing in the way

    Tell the content apart from the navigation and the chrome.

    Landmarks
    <main>, 17 <article>, <nav>
    Outline
    h1 AXO Blog · h2 Why an Agent That Works 80% of the Time Finishes a Third of the Job · h2 The Same Agent, the Same Page, 34 Points Apart · h2 Muse and Dots Both Drive a Browser · h2 GEO, AEO, LLMO: What Do You Actually Call This? · h2 Atlas Is Dead, Agentic Browsing Isn't

    Your content is distinguishable from the navigation and the footer, so a quote comes from the article rather than from a menu.

  4. 4Understand

    Gets through, with gaps

    Work out what this page is and what question it answers.

    Declared types
    Organization, ImageObject, Person, EducationalOccupationalCredential, WebSite, Blog, BlogPosting
    Headline
    AXO Blog

    What is missing: The markup and the visible page disagree, so a machine holds two versions of the facts and may repeat the wrong one.

    What this costs you: The agent has to guess the subject from the text. It matches your page to the wrong questions and misses the right ones, which shows up as being absent from answers you should be the obvious source for.

  5. 5Attribute

    Gets through, with gaps

    Find out who wrote it, when, and on what basis.

    Author
    Alan Weibel
    Dates
    none found

    What is missing: Claims stand on their own with nothing behind them, so neither a reader nor a model can check where they came from.

    What this costs you: The agent has a claim it cannot attribute. Systems that cite sources tend to prefer one they can name, so your material gets used without credit or passed over for a weaker page that says who wrote it and when.

  6. 6Explore

    Nothing in the way

    Find the rest of the site, and tell what has changed.

    Links out of here
    55 internal, 3 external
    Within the content
    35 of those are inside the main region

    The rest of the site is reachable from here, and changes are visible, so one good page leads to the others.

  7. 7Act

    Nothing in the way

    Do something here rather than only read: a form, a tool, an API.

    Form submits to
    POST /api/subscribe
    Form fields
    email:email required

    There is something here to do, named clearly enough that an agent can do it on someone’s behalf.

An agent gets all the way through, but loses something at each marked step.

What the steps mean

There are seven steps, in the order an agent goes through them. It arrives at the page, pulls the words out of the page’s code, tells your content apart from the menus, works out what the page is, finds who wrote it, looks for the rest of the site, and finally tries to do something other than read.

If a step stops the agent completely, it never reaches the steps after it. Smaller problems make a step harder without stopping it. A single score hides that difference. A page can have flawless structured data (labels in a page’s code that state facts such as the author, date, or price) and a perfect author byline, and still be a dead end, because the agent got stuck two steps earlier and never saw any of it.

That’s why every step after the point where the agent stopped is drawn as unreached instead of passed. The agent never got the chance to try it.

The same idea, for software

The Visualizer follows an agent through a web page. An agent working inside an app goes through 7 steps of its own, from finding out what the software can do and proving it’s acting for someone, through trying the task and recovering when it fails, to handing back a result its user can check.

The agent journey map for software →