Entity‑Conditioned Probing with Resampling: Validity and Reliability for Measuring LLM Brand/Site Recommendations

We introduce entity-conditioned probing with resampling, a simple, reproducible method to measure intrinsic brand/site associations in large language models. A single schema-constrained prompt produces top-N lists for each category–locale cell; we collect k independent samples per cell and aggregate with a Plackett–Luce (PL) model to obtain latent worth scores and ranks with 95% bootstrap confidence intervals. In a study of 52 categories × 4 locales (US/GB/DE/JP) totaling 15,600 prompt iterations, we report PL scores alongside frequency baselines (@1/@3) and find strong split-half stability at top-3 (median Spearman = 1.00; mean = 0.876, 95% CI 0.806–0.932; overlap@3 mean = 0.962, 95% CI 0.936–0.985). The method is model-agnostic, emphasizes structured outputs and alias canonicalization, and separates intrinsic association from first-turn prompting effects; when forecasting first-turn outcomes is required, a small stratified panel can be used for monotonic calibration. Code and processed aggregates are openly available (see Related Works). (Preprint v0.6.1.)