Resolving patterns of ancient and rapid diversifications is one of the most challenging tasks in evolutionary biology. These difficulties arise from confusing phylogenetic signals that are associated with the interplay of incomplete lineage sorting (ILS) and homoplasy. Phylogenomic analyses of hundreds, or even thousands, of loci offer the potential to resolve such contentious relationships. Yet, how much useful phylogenetic information these large data sets contain remains uncertain and often goes untested. Here, we assess the utility of different data filtering approaches to maximize phylogenetic information and minimize noise when reconstructing an ancient radiation of Neotropical electric knifefishes (Order Gymnotiformes) using ultraconserved elements. We found two contrasting hypotheses of gymnotiform evolutionary relationships depending on whether phylogenetic inferences were based on concatenation or coalescent methods. In the first case, all analyses inferred a previously—and commonly—proposed hypothesis, where the family Apteronotidae was found as the sister group to all other gymnotiform families. In contrast, coalescent-based analyses suggested a novel hypothesis where families producing pulse-type (viz., Gymnotidae, Hypopomidae, and Rhamphichthyidae) and wave-type electric signals (viz., Apteronotidae, Sternopygidae) were reciprocally monophyletic. Nodal support for this second hypothesis increased when analyzing loci with the highest phylogenetic information content and further increased when data were pruned using targeted filtering methods that maximized phylogenetic informativeness at the deepest nodes of the Gymnotiformes. Bayesian concordance analyses and topology tests of individual gene genealogies demonstrated that the difficulty of resolving this radiation was likely due to high gene-tree incongruences that resulted from ILS. We show that data filtering reduces gene-tree heterogeneity and increases nodal support and consistency of species trees using coalescent methods; however, we failed to observe the same effect when using concatenation methods. Furthermore, the targeted filtering strategies applied here support the use of “gene data interrogation” rather than “gene genealogy interrogation” approaches in phylogenomic analyses, to extract phylogenetic signal from intractable portions of the Tree of Life.