APC Geometry Computation

This document records how the AI Page Content (APC) agent derives geometry for each node today. The implementation lives in third_party/blink/renderer/modules/content_extraction/ai_page_content_agent.cc.

Geometry Outputs

• outer_bounding_box (viewport space) Computed from LayoutObject::AbsoluteBoundingBoxRect() using viewport mapping flags. When a local clip-path is present, the clip-path bounds are mapped to viewport space and replace the standard box outright.

• visible_bounding_box (viewport space)

  • This uses LocalBoundingBoxRectForAccessibility().

• fragment_visible_bounding_boxes (viewport space) Derived from the layout fragments produced by LayoutObject::QuadsInAncestor(). Each quad is mapped to the viewport, intersected with visible_bounding_box, and empty results are dropped. We only emit the vector when more than one non-empty fragment survives, so single-fragment cases do not duplicate the primary box.

All rectangles are integer enclosures of the underlying floating-point quads.

Coordinate Space and Units

All geometry is expressed in visual-viewport coordinates in BlinkSpace (device pixels), not CSS pixels/DIPs.

Terminology:

• Layout viewport: the viewport used for layout/scrolling.
• Visual viewport: what the user actually sees; it can be offset from the layout viewport (e.g. when the mobile toolbar shows/hides during scrolling, or during pinch-zoom).

Browser-side consumers that start from a DIP coordinate (for example actor.mojom.ToolTarget.coordinate_dip or JavaScript getBoundingClientRect()) must convert into APC geometry coordinates before comparing or hit-testing.

For the canonical coordinate space contract, see AIPageContentGeometry in:

• third_party/blink/public/mojom/content_extraction/ai_page_content.mojom

Validation

In debug builds ValidateBoundingBoxes() asserts a few invariants:

• visible_bounding_box must be expressed in viewport coordinates with a non-negative origin.
• visible_bounding_box must lie within outer_bounding_box (within a small tolerance) because both derive from the same fragment set.
• When a clip-path exists, visible_bounding_box must also be contained inside the clip-path’s viewport-space bounds.

These checks help catch regressions where upstream mapping or clipping becomes inconsistent.