Abstract

An agentic system that:

(a) measures how LLM models represent a company in prompts used by its buyers.
(b) audits that company’s online presence on dimensions that influence LLM responses.
(c) makes strategic recommendations after analyzing LLM responses and audit results.

The system runs in three phases designed around human approval checkpoints.

Research: core context, prompt, and scoring artifacts are created.
Analysis: Human-approved prompts and the company’s online presence are analyzed.
Reporting: Human-approved reccomendations and findings are compiled into a report.

The system runs as three Claude Code-harnessed agents that execute pre-determined sequences of sub-agents, deterministic scripts, and human operator-facing approval requests.

Design Principles

The system has core design principles that differentiate it from existing industry tools:

Discovery (category, unbranded) and Assessment (branded) prompts separated.
- Most tools group them, but they have different improvement levers and buyer stages.
A performance score is used to measure how a LLM response meets success criteria.
- Most tools focus on mention rate, which doesn’t measure the quality of a response.
Audit results are quantitative. A company is scored 0-5 on 28+ dimensions using rubrics.
- Most audits are qualitative assessments that require judgment to interpret.
All scores produced by the system are reproducible.
- Most systems that rely on AI judgement drift from run-to-run with the same inputs.