During parametric memory auditing, the probe engine targets the primary owned brand and its aliases with standardized ungrounded prompt templates across configured rounds—such as batches of 1, 3, 5, 10, or 20 iterations. These runs query models across Google, OpenAI, and Anthropic in parallel using concurrent provider pools, saving each raw generation to the database and isolating failures through transactional savepoints.

The parser then strips boilerplate formatting while preserving complete sentence assertions, assigning each extracted fact a normalized, position-based confidence score ranging from 1.0 at the top of the response down to 1/N at the bottom. This multi-round sampling quantifies the frequency, consistency, and rank variance of individual recall statements to evaluate the brand’s footprint before any curation occurs.