Treewalker repeatedly queries a model's ungrounded memory while analyzing token-level logprobs to capture parametric associations with a brand. By branching into alternative token paths at decision thresholds, it identifies latent semantic alternatives that the model considers during generation, and by aggregating appearance rates across multiple runs, it measures the stability and confidence of each identified concept. This process maps a brand's associative landscape and quantifies how reliably core topics are surfaced from model memory.