Four field campaigns were conducted in southern Arizona (AZ) and in northern Texas and southern Oklahoma (TX-OK) in 2003 and 2004 to evaluate the performance of the U.S. National Lightning Detection Network™ (NLDN) in detecting cloud-to-ground (CG) lightning after an upgrade in 2002 and 2003. The 2-year average flash detection efficiency (DE) in AZ was 93% (1024/1097), and the measured (first plus subsequent) stroke DE was 76% (2746/3620). The corresponding values in TX-OK were 92% (338/367) and 86% (755/882), respectively. After correcting for the time resolution of the video camera (16.7 ms), we estimate that the actual NLDN stroke DE and video multiplicities were about 68% and 3.71 in AZ and 77% and 2.80 in TX-OK. The average DE for negative first strokes (92%) was larger than the measured DE for subsequent strokes that produced a new ground contact (81%) and the DE for subsequent strokes that remained in a preexisting channel (67%). The primary cause of the NLDN missing strokes was that the peak of the radiated electromagnetic field was below the NLDN detection threshold. The average estimated peak current (Ip) of negative first strokes and the average multiplicity of negative flashes varied from storm to storm and between the two regions, but this variability did not affect the DE as long as the recording sessions had more than 60 flashes. By analyzing the NLDN locations of subsequent strokes that remained in the same channel as the first stroke we infer that the median random position error of the NLDN was 424 m in AZ and 282 m in TX-OK. An evaluation of the classification of lightning type by the NLDN (i.e., CG stroke versus cloud pulse) showed that 1.4–7% (6/420 to 6/86) of the positive NLDN reports with an Ip ≤ 10 kA in TX-OK were produced by CG strokes; 4.7–26% (5/106 to 5/19) of the positive reports with 10 kA < Ip ≤ 20 kA were CGs; and 67–95% (30/45 to 30/32) of the reports with Ip ≥ +20 kA were CG strokes. Some 50–87% (52/104 to 52/60) of the negative, single-stroke NLDN reports in AZ and TX-OK with ∣Ip∣ ≤ 10 kA were produced by CG flashes. Both the upper and lower bounds in these classification studies have observational biases.