BRGEO-1: A longitudinal audit of entity-mention measurement in generative engines

A measurement of entity presence can change when the answer remains fixed but the retained text, labels or denominator changes. This study retrospectively audits 91,392 persisted responses from a practitioner-built generative-engine collection, using 83,486 historical records from 50 observed dates for the principal analyses. A frozen extractor identifies whether a response contains at least one eligible lexicon entry. In 7,436 historical Perplexity responses, restricting the longer retained text to 200 Unicode code points reduced detection from 75.73% to 51.95%, a paired difference of -23.78 percentage points. Separate query- and date-cluster resampling showed different sensitivities. A recent analysis of 272 common query/date cells across five services preserved negative window contrasts in every service. Further audits found that a semantic-sounding absorption field duplicated the mention flag, a composite numerator included observations excluded from its denominator, and lexical decoy hits could coexist with refusal markers. Exploratory sequence analysis showed that assigning perfect similarity to pairs with no detected entities increased mean Jaccard overlap from 0.7423 to 0.9460. Overlapping lexicon labels changed effective diversity without changing binary detection. Repetition and incomplete query batteries further constrained temporal interpretation. These results support reporting the observed text, unit, lexicon, transformation, comparison population and uncertainty assumptions together. They establish consequences of measurement choices in this archive, rather than causal effects of GEO interventions or general provider rankings.

Corrected empirical revision of the BRGEO-1 manuscript submitted to SSRN (Abstract ID 7446519; 183 pages; https://ssrn.com/abstract=7446519). The two manuscripts are versions of the same research, not independent studies or independent replications. Appendix D documents the reconciliation of data snapshots, analytical units, weighting and interpretation. The existing SSRN submission is retained unchanged. This version has not undergone journal peer review.

AI disclosure: OpenAI Codex and Anthropic Claude assisted with research code, drafting and internal review; Perplexity Sonar Pro assisted with literature discovery. Reported results were computed from archived data by the analysis scripts. These tools were not authors, human annotators or external peer reviewers.