I gave GPT-Vision a screenshot of the shinobis.com homepage. In a separate chat, I pasted the complete HTML of the same site. The same 5 questions in both cases: what it is about, who it is for, what the hierarchy is, what you would do first, and what confuses you.
It described two different sites.
From the image: "I understand what this is, but I'm not 100% sure what you want me to do first." From the HTML: "The HTML is much more explicit for an AI than it is for a conventional blog taxonomy."
My site is more legible to machines than to humans. And that was not intentional.
What the image does not communicate
From the screenshot, GPT-Vision identified the general topic quickly: a designer sharing AI tools and workflows. So far, correct. But it flagged five problems that are only visible when looking at the page as a human sees it.
The site identity is unclear. The logo appears as "XX" without a readable name. A new visitor does not know whose site this is without reading further.
There are too many competing entry points. Free tools, Claude Code, three paths, articles. GPT-Vision said: "I can understand the site, but I don't immediately know what the creator considers the #1 thing I should do."
"Start with Claude Code" is too specific without prior context. A new visitor does not know why Claude Code matters.
The generative thumbnails visually dominate the article list but communicate nothing about the content. The titles do the work. The images occupy visual space without contributing information. GPT-Vision described the thumbnails as "QR-code-looking" and noted they "don't help me understand the content."
The three paths ("Most connected / Explore / Build") sound like categories but their relationship to the rest of the site is not obvious from the visual.
What the HTML does communicate
From the HTML, without seeing anything visual, GPT built a much more complete and coherent mental model of the site.
It identified exactly what it is: "a personal technical publication and tool hub by a UX/UI designer who uses AI in real production work." That is because the HTML explicitly declares what the site is, who it is for, and the principle behind it. It does not infer it. It reads it.
It mapped the complete hierarchy: positioning, then tools, then orientation, then user intent, then curated content, then articles. It noted the page prioritizes the value proposition before the content, which is a deliberate information architecture decision that from the image is perceived as "too many entry points."
It identified the primary action without ambiguity: "Try my free tools" has the btn-primary class and sits directly below the H1. For a machine, the hierarchy is clear. For a human looking at the page, it competes with "Start with Claude Code" next to it.
And it concluded something no previous audit articulated as clearly: "The HTML is much more explicit for an AI than it is for a conventional blog. It literally declares what this is, who it's for, the principle, the user intents, and the primary tool action. That is a very strong semantic structure."
The same site, two readings
The divergences are the finding.
Identity. From the HTML, the AI knows exactly who I am, what I do, and why it matters. The nav says "Designer in the Age of AI." Meta tags establish "AI Tools and Workflows for Designers." From the image, the logo looks like "XX" and the visitor has to read the hero to understand whose site it is.
Thumbnails. From the HTML, alt text exists and is descriptive for the grid articles. Machines read that. From the image, a human sees QR-like visual patterns that dominate the article list without communicating anything about each post's content. The HTML is accessible for machines. The visual experience is not informative for humans.
Primary action. From the HTML, btn-primary unambiguously marks "Try my free tools" as the main action. From the image, it competes with "Start with Claude Code" and the three paths and the articles. The priority the code establishes does not translate visually.
Orientation. From the HTML, the "What this is / Who it's for / The principle" sections are structured content answering the three basic questions a new visitor has. From the image, they are text blocks competing with CTAs and articles for visual attention.
What this reveals about GEO
For 6 months I built infrastructure for machines to understand my content. JSON-LD with 5 connected entities. llms.txt with a site summary. Content negotiation to serve Markdown to agents. A graph search with a WebMCP endpoint. All of it works. AI models read my site and understand exactly what it is, who it is for, and how the content relates to each other.
But the visual layer, the one humans see when they arrive for the first time, does not communicate with the same clarity. The invisible infrastructure is more precise than the visible interface. The backend speaks to machines with exactness. The frontend speaks to humans with ambiguity.
It is an interesting inversion of the usual priority: I deliberately optimized the semantic structure for machines and discovered that the visual interface had not received the same level of attention.
Thumbnails as symptom
The generative images for each post were born to solve a real problem: stock photos repeat and feel dead. Each image is a unique visual pattern generated from the hash of the post title. They are fast to load (2 to 8KB in SVG), unique by definition, and require no manual image searching.
But GPT-Vision just confirmed something I suspected without data: the thumbnails do not communicate content. A human looks at the article list and sees repetitive patterns that tell them nothing about each post. The titles do all the work. The images occupy visual space that could be communicating something useful.
That does not mean generative thumbnails are a mistake. It means their function is visual identity, not content communication. If you evaluate them as communication, they fail. If you evaluate them as identity (each post has a unique non-stock image), they work. The problem is they occupy the visual space where a human expects communication.
What I am going to change
The visual layer needs to inherit the clarity of the semantic layer. The HTML already knows what to prioritize. The interface does not yet.
Thumbnails are a future project that needs exploration: how to make the generative image communicate something about the post content without losing the unique visual identity or the loading speed. It is a design problem I am not going to solve with a quick fix.
The primary action needs to dominate visually. If the HTML says btn-primary, the interface should reflect it without visual competition at the same level.
The site identity needs to be readable in the first 2 seconds without reading text. A visitor arriving from an AI citation should know whose site this is before scrolling.
The lesson
If your semantic infrastructure is clearer than your visual interface, you have a problem most blogs do not have: machines understand you better than people. And in an ecosystem where humans still decide whether your content is worth their time, that means half the work is perfect and the other half needs to catch up.
The same test that reveals the strength of your GEO reveals the weakness of your UX. And that is exactly the kind of discovery you cannot make by reading only metrics or only code. You need to see your site as two different audiences see it: the machine that reads your HTML and the human that looks at your screen.