How an AI Reads Your Homepage — and Why It Disagrees With You

When you look at your own homepage, you don’t actually see it. You see the version in your head: the one where the tagline is clever, the layout is obvious, and the button everyone should click glows with self-evident importance. You built the site, or paid for it, or at least approved it, and so every time you load it your brain fills in the intent behind each element. A visitor gets none of that. A visitor gets pixels and words, cold.

This is why I built GazeSite to read a homepage the way a stranger does. You give it one URL. It opens the page in a real browser, takes a screenshot on desktop and again on mobile, and grabs the HTML underneath. Then six reviewers, each focused on one thing — accessibility, clarity, conversion, readability, SEO, technical health — look at that evidence and write down what they find. Each finding gets a score, a one-line headline, a description of the issue, and a prescribed remedy. That’s the whole mechanism. No account of your intentions is admitted into evidence.

The interesting part is not that an AI can do this. The interesting part is how often it disagrees with the owner, and why.

The most common disagreement is about clarity. Owners believe their homepage says what the company does. Usually it says what the company aspires to be. “Empowering teams to move faster” is not a description; it’s a mood. The AI, reading the page with no prior knowledge, will report honestly that it cannot tell what the product is. Owners find this verdict insulting until they watch a real stranger have the same reaction. The AI is not being obtuse. It is being exactly as informed as your next visitor: not at all.

The second disagreement is about the mobile version. Almost every owner reviews their site on a large monitor, because that’s where they work. But the screenshot the AI takes at phone width is frequently a different site: the headline wraps into an awkward orphan, the hero image pushes the call to action below the fold, a cookie banner eats a third of the screen. When the report scores the mobile experience lower than the desktop one, owners often assume the tool is confused. It isn’t. It just looked at something they hadn’t.

The third disagreement is the deepest one, and it comes from reading the HTML rather than the picture. A page can look fine and be broken underneath. The heading structure skips from an h1 to an h4, which matters to screen readers and to search engines even though no sighted visitor notices. The images have no alt text. The button that looks like a button is actually a styled div that a keyboard can’t reach. Owners judge their sites by appearance because appearance is what they can see. Half of what determines whether a site works is invisible, and that half is where the machine has an honest advantage.

There’s a general principle here that goes beyond websites. Feedback is valuable in proportion to how different the reviewer’s perspective is from yours, and how honest they can afford to be. Your colleagues share too much of your context to notice what a stranger would miss, and they like you too much to say the tagline is empty. An AI reviewer has neither problem. It shares none of your context, and it has no relationship with you to protect. It is, in a sense, the world’s most available stranger.

That doesn’t make it infallible. Sometimes it flags something you did deliberately, for a reason it can’t know — a sparse page that’s sparse on purpose, a wall of text that your particular audience genuinely wants. When that happens you should overrule it, and you’ll overrule it with confidence, because now you’ve had to articulate the reason. That articulation is itself worth something. Most design decisions on most websites were never decisions at all; they were defaults nobody questioned. A disagreement forces the question.

So when the report comes back and part of it strikes you as wrong, resist the urge to dismiss the whole thing. Sort the findings into three piles: things you knew and hadn’t fixed, things you didn’t know, and things you disagree with. The first pile is a to-do list. The second pile is the reason you ran the audit. And the third pile, the disagreements, is where you’ll do your best thinking — because for each one you’ll either discover the machine saw something you couldn’t, or you’ll finally know why your page is the way it is.

Either way, you come out ahead. The one thing you can’t do anymore is claim the page is obvious. Obvious to whom? You’ve now met a reader with no memory of your intentions, and it told you what it saw.