While CityGML was traditionally the dominant standard and format for 3D city models, the ecosystem now includes CityJSON and its streaming derivatives, CityJSONSeq and FlatCityBuf as other formats. Consequently, modern web-based distribution relies on three competing frameworks, namely CityJSONSeq utilizing sequential JSON Lines, FlatCityBuf providing zero-copy binary encoding, and OGC 3D Tiles 1.1 delivering pre-triangulated glTF. Despite their concurrent evolution, empirical comparisons of their data-loading and rendering efficiency are lacking. This study bridges this gap by evaluating these formats across five real-world CityJSON datasets, ranging from 853 to 145,865 building geometries. The evaluation utilizes a custom, bandwidth-controlled loading harness and an independent WebGL2 and WebGPU rendering pipeline, developed without third-party scene-graph libraries to strictly isolate the graphics API. Results demonstrate that 3D Tiles/glTF consistently yields the smallest file sizes, reducing payloads by 29 to 73 percent compared to CityJSONSeq. It is also the fastest to parse and, by eliminating client-side triangulation, reaches a fully render-ready state significantly quicker, performing up to 47.1 times faster than CityJSONSeq and 36.9 times faster than FlatCityBuf on the largest dataset. Among CityJSON-derived formats, FlatCityBuf is particularly effective for large-scale data parsing due to its zero-copy deserialization capability, substantially outperforming CityJSONSeq. When paired with an optimized decoder, it also achieves greater overall rendering readiness at scale. Nevertheless, because both CityJSON formats require client-side triangulation, 3D Tiles/glTF maintains an order-of-magnitude performance advantage. Ultimately, these empirical findings provide web-based Geographic Information System architects with definitive, scale-dependent guidance for selecting the optimal streaming format.