We scanned 8 real websites. 43% of their visual assets are invisible to AI.

No cherry-picking. We ran SeenLayer's own scanner against 8 live, currently-published websites — a mix of marketing agency sites, content sites and small business sites, all real, all indexed, none staged for this post. Combined, they carry 179 images and video. Here's what an AI crawler — Perplexity, ChatGPT Search, Google's AI Overviews — actually sees when it reads them.

That's 102 of 179 assets. Not broken images. Not missing files. Images and video that render fine in a browser and ship with zero machine-readable description — present on the page, absent from every AI answer.

The full sample

Visibility Score is SeenLayer's 0–100 measure across structured data, alt text/transcript quality, citable context and entity linking. "Invisible" = no alt text, no transcript, no Schema.org for that asset. A ninth site, ghostinfluencer.com, is excluded from these totals and covered separately below — its count isn't low, it's zero, for a different reason.

The one that isn't in the table

ghostinfluencer.com scans to 0 assets found — not one photo, not one screenshot. Not because the site has no images. Because the raw HTML a crawler receives, before any JavaScript runs, is this:

That's the entire page. A single-page application rendered client-side: a visitor's browser executes JavaScript and builds the page live, but SeenLayer's crawler — like GPTBot, PerplexityBot and most of the real AI crawler stack — reads the HTML as delivered, not as rendered. There is no image tag to describe because there is no HTML at all yet. This isn't a scoring problem to fix with better alt text; it's an architecture that ships invisible to AI by default, images included.

What this actually costs

AI engines don't look at pixels while crawling. They read alt text, video transcripts and Schema.org structured data — the same signals Google's own documentation cites for what AI Overviews draw from. An image with an empty alt="" is, to every one of these systems, functionally identical to an image that was never uploaded. It cannot be described, cannot be linked to an entity, cannot be cited as a source.

The Princeton GEO study measured what happens when that layer exists instead: content with structured, factual, entity-linked markup saw up to 40% more generative visibility. The 8 sites above are sitting on the wrong side of that gap — not because the content is bad, but because 57% of it was never written for the reader that matters most to AI-era discovery: the crawler, not the browser.