Perception
What a language model actually perceives of a web page when nothing is ever rendered to an image — general knowledge, useful even if you never install anything.
A language model does not see a web page. Give it a screenshot and it gets pixels: it can tell you the page looks wrong, but it cannot act on anything, because nothing in an image is addressable. Give it the raw HTML and it gets everything at once — including hundreds of lines of markup no human eye would ever register.
Dinoer gives it neither. Not a smaller image, not a compressed one — no image at all, in any mode, for any reason. What it gives instead is the structure the browser itself already builds, as text.
Two views, both text, and where they differ
The accessibility tree is the structure the browser builds for screen readers — titles, roles, links, states — at a fraction of the token cost of an image. Extracted text is the page’s cleaned documentary content, noise tags stripped, for when what matters is what the page says rather than how it is structured.
Neither is a fallback for the other. The tree tells you what a page is; extracted text tells you what it says. A page can have a rich, navigable tree and thin content, or the reverse.
Reproduce it
The figures in this section are real, not staged — captured against
https://example.com with the actual tool, on 14 August 2026:
/opt/dinoer/venv/bin/python3 /opt/dinoer/shot.py \
--url https://example.com --a11y --guide-version 1.6