Libro

Descripción

Large language models now stand between customers and the companies they might choose, and an industry has formed around measuring which companies get named. The figures that industry reports are not comparable with one another, because the conditions under which a citation is counted are almost never stated. This book specifies BRGEO-1, an open protocol that declares six of those conditions, and reports what happened when one of them was measured instead of assumed. Reading a model's whole answer rather than its first 200 characters raised the citation rate by 23 to 56 percentage points across five engines, on the same observations, with no answer losing a citation. One engine that appeared to name almost no companies turned out to name them past the point where the instrument had stopped reading. Alongside the specification, the book gives the field record behind it: five months of daily collection against six commercial engines, 86,543 observations, a 59-day interruption declared rather than hidden, and a series of defects that passed every automated check the pipeline ran on itself. It is written for anyone who buys, sells, or publishes a number about visibility in generative search, and for researchers who need a measurement standard before the comparisons start.