Iris
07/07 · the observervisual observationIris looks at the pages the other six only read — rendered layout, images, diagrams — and reports what a reader sees that the text never says.
- reads the rendered page
- derived signals only
- flags what text hides
The other six read an estate the way a text crawler does: the HTML, the metadata, the words in the markup. Iris looks at the page the way a person does — the rendered layout, the screenshots, the diagrams, the numbers that live inside an image and never appear in the text. Most of an estate is written down. Some of it is only ever drawn.
That gap is where retrieval quietly fails. A pricing table shipped as an image is invisible to the agent quoting your prices; a claim that lives only in a diagram is a claim your own AI cannot cite. Iris works from derived signals and a downsampled capture — never a stored copy of the page — and its output is always a flag: what a reader can see that the text cannot, ranked by whether an agent would miss it. Iris reports; it never redesigns.
Sample findings
"Your pricing table is an image. Every agent that reads only text sees a blank page."
"Three product diagrams carry numbers that appear nowhere in the page’s text."
"Your homepage renders one headline; the HTML ships a different one underneath."
"A consent wall covers a page the text crawler reported as fully read."
Method
- 01Render captureThe page as a browser draws it — layout, images and text a plain fetch never receives.
- 02Derived signalsLayout, contrast and image content reduced to signals and a downsampled thumbnail. No copy of the page is kept.
- 03Text-versus-render diffWhat a reader sees set against what the crawler received — the claims that live only in pixels.
- 04Retrieval noteWhether an agent fetching text alone would miss it — the flag that decides severity.
depends on
Scout’s map says which pages exist; Warden’s structure says which a reader never reaches.