Existing research on source tracing, the task of classifying the method that generated an audio deepfake, focuses on high-quality wide-band speech, while narrow-band telephony speech has received little attention despite its importance. Telephony speech faces significant challenges such as limited bandwidth and codec variability that can degrade performance beyond laboratory conditions. This paper aims to quantify the impact of telephony transmission on audio deepfake source tracing. We separately evaluate the performance degradation due to lower bandwidth and codecs and we demonstrate that codecs are responsible for the majority of degradation. We treat telephony-transmitted data as an unseen condition and we explore data augmentation techniques such as RawBoost and random audio quantization to improve performance. We show that these strategies eliminate the need for costly real telephony data and can outperform training on it, reducing the equal error rate by a relative 20.1%.