Where the Number Came From: How Generative Search Loses the Record, and What an Operator Can Do About It
On 11 September 2026 the same question, the share of Loudoun County's local tax revenue that comes from data centers, went to four generative search products. Back came six true fractions over different bases and years, and one product in four named its base. The brief follows two of those figures to the county's adopted budget, sets out the three moves and the stopping rule of a trace, and ends with rules an operator can staff: base and year in the same sentence as every fraction, one citation per decision memo followed to the record, and agreement among tools counted for nothing until one chain has reached a document.
An explanatory brief with one worked case, written for people who run organizations and whose staff use generative search products to produce the memos they decide on.
On 11 September 2026 the same question, the share of Loudoun County, Virginia's local tax revenue that comes from data centers, was put to four generative search products. The answers returned six true fractions over different bases and years, and one product in four named its base. The brief follows two of those figures to the county's adopted budget, sets out the three moves and the stopping rule of a trace, explains why the checks organizations run on memos pass such figures (the signals those checks read, fluent prose and an attached citation, used to be expensive and are now free; the products are tuned to defer to the premise of the question, including the premise the asker did not state; agreement among products is a fact about a shared corpus), and ends with rules an operator can staff: base and year in the same sentence as every fraction, one citation per decision memo followed to the record with the chain length recorded, and agreement among tools counted for nothing until one chain has reached a document.
The Loudoun case is offered as a demonstration of what can happen, not as a measurement of how often it happens. The brief makes no rate claim. The captured answers, traced pages, and local copies of the county records are in the Source Diversity Audit dataset (10.5281/zenodo.22839545). The larger claims about corpus-scale effects are argued in the working paper "Who Is Training Whom" (10.5281/zenodo.22839554).