Respiratory motion artifacts degrade PET/computed tomography (PET/CT) image quality. Data-driven gated (DDG) PET/CT addresses this issue by extracting respiratory signals directly from PET data, eliminating the need for external monitoring devices. This study investigated the effects of data extraction percentage (%count) and deep learning-based reconstruction [Advanced Intelligent Clear-IQ Engine-integrated (AiCE-i)] on image quality in DDG-PET under different respiratory conditions using a phantom model. A body phantom containing six spheres (10-37 mm) was imaged using a silicon photomultiplier-based PET/CT system. Four respiratory waveforms (no-motion, sinusoidal, representative patient, and baseline shift) and four %count levels (20, 30, 40, and 50%) were evaluated using AiCE-i reconstruction. Image quality was assessed using background variability (N10 mm), percentage contrast (QH,10 mm), contrast-to-noise ratio (QH,10 mm/N10 mm), and recovery coefficient. Increasing %count consistently reduced N10 mm across all respiratory waveforms. QH,10 mm and QH,10 mm/N10 mm generally increased with increasing %count; however, at the 50% threshold, significant reductions were observed in the baseline shift and sinusoidal waveforms compared with the no-motion condition. Recovery coefficient analysis demonstrated the partial volume effect in smaller spheres and showed that quantitative performance was maintained across the evaluated gating conditions. The combination of DDG and AiCE-i maintained stable image quality across the respiratory conditions evaluated. Under the conditions of this phantom study, a %count range of 30-40% provided a favorable balance between image noise and the effects of respiratory motion for standard 120-s acquisitions.