Results Addendum — The External-Text Study (reco-study-v1): How Language Models Select Products for Recommendation

Results addendum to the pre-registered External-Text Study (reco-study-v1), reporting confirmatory outcomes exactly as registered: H1 PASS (matched-pair supporting-density difference, one-tailed Wilcoxon p = .002, n = 507 pairs); H2 FAIL (registered falsification rule fired: model familiarity r = +.176 exceeded supporting density r = +.077; density rejected as primary mechanism, reported as committed); H3 PASS with a stated fame-matching caveat (Mann-Whitney U = 3,684.5, p = .0078). Includes registered exploratory analyses, labeled post-hoc exploratory findings, deviations and operationalizations, and one withheld finding that failed reproduction. Study: 46 shopping queries, 1,679 products, 99,436 scored texts, outcomes from GPT-5 and Gemini 2.5.

Pre-registration: https://osf.io/a8zuh (embargoed). Public results presentation: https://scope.telesuite.io/results