Short answer: Plan GLB geometry, texture, material and node budgets around target devices, viewing distance, object queries and measured runtime performance.
Start with devices and visible scope
A dedicated desktop GPU, an office laptop and a mobile device have very different budgets. Test the weakest target device and the heaviest representative view.
Triangles are only one cost
Nodes, draw calls, material switches, transparency, skinning, animation, decoding and shadows also affect performance. A low-poly model can still be slow.
Count texture transfer and GPU memory
Compressed network size is not decoded GPU memory. Set resolution by viewing distance, reuse materials, combine channels and prepare smaller mobile variants.
Object queries set batching boundaries
Batching reduces draw calls but can limit selection, visibility and live updates. Define operational object granularity before merging or instancing.
Use tiered budgets
Separate initial, near-view and on-demand assets, then record transfer size, nodes, materials, textures, peak memory, loading time and frame rate for each tier.
Accept by measured scenarios
Use fixed devices, network conditions, camera paths and interactions. Optimization is complete only when visual error and runtime targets are both satisfied.