The screen stops being a rectangle you look at and becomes a room you are standing in. Everything that follows is a consequence of adding one axis.
Definition
Spatial UI is the design language of headsets — visionOS, Quest's Horizon OS, Android XR — where windows are objects placed in your actual room rather than layers on a display. It brings back everything flat design threw away, but for a hard reason rather than a decorative one: depth is now real. A shadow is not a metaphor for elevation, it is a shadow. Parallax is not an effect, it is what happens when you move your head.
The interaction model is the other half. Input is gaze plus pinch: you look at a thing, it lights up, you tap two fingers together. There is no cursor, no hover in the mouse sense, and no way to know exactly where someone is looking except by lighting up whatever they are looking at.
The rules
Panels are translucent and adapt to the light in the physical space. An opaque window in a real room reads as a hole cut in reality, and it is genuinely unpleasant.
Elements that belong together share a Z plane. A menu is not "above" its parent in the stacking-context sense — it is measurably in front of it, and it casts onto it.
Eye tracking has a genuine error radius, roughly a degree. Sixty points minimum, generous spacing, and nothing important within a thumb's width of anything else.
Gaze feedback fires constantly as the eye moves. A colour fill on every glance is exhausting; a soft rim glow reads as acknowledgement rather than selection.
Content sits within about a sixty degree field, slightly below eye level, at arm's length or beyond. Anything that requires turning your head is a design failure someone will feel in their neck.
The scene stays anchored to the room, not to the user. Motion that is not caused by the person wearing the device is the fastest known route to nausea.
Specimen
Safari on visionOS renders ordinary pages as floating panels, so most sites already "work". What breaks is anything that assumed a mouse: hover menus that need a cursor path, tiny hit targets, and fixed overlays that collide with system chrome. The cheapest real win is bigger targets.
This page fakes it with CSS perspective and pointer parallax, which is a flat-screen impression of the language rather than the language itself. Depth you can walk around, occlusion by your own hand, and shadows cast onto a real sofa are not things a browser on a monitor can give you.