Parametric memory auditing inspects how artificial intelligence models perceive and represent a brand purely through their internal weights, without the aid of live search tools, retrieval-augmented generation (RAG), or real-time URL browsing. The audit engine sends ungrounded prompt batches across models from Google, OpenAI, and Anthropic over multiple sampling rounds (typically between 1 and 20 cycles) to query what products, services, and corporate attributes are directly associated with the brand in parametric memory.

The auditing process operates across several distinct stages:

• Stochastic Sampling & Logprob Exploration: Probes elicit exhaustive lists of remembered statements. For deeper diagnostic runs, Treewalker token-level branching analyzes Vertex AI log probabilities to measure token-level confidence scores and test alternative token paths whenever a competing candidate exceeds a probability threshold.
• Ranked Fact Extraction: Raw model text is cleaned to strip chain-of-thought artifacts, markdown fences, and bullet formatting. Each output line is parsed as an individual assertion and assigned a position-based confidence score that scales down from 1.0 for top-ranked facts to 1/N for items lower in the response.
• Automated Curation & Deduplication: Extracted lines are passed to a curation engine backed by strict JSON schemas. This stage cross-references new statements against the canonical fact ledger, resolving duplicate concepts, updating claim revisions, and calculating cross-platform consensus metrics like total mention counts and platform counts.
• Perceptual Narratives & Verification: The system synthesizes cross-model summaries alongside dedicated per-platform narratives detailing where models agree, where they are uncertain, and where hallucinations occur. Finally, human editorial rulings can be logged directly against individual claims as verified or incorrect, locking factual benchmarks against future automated rewrites.