NiubiGEO: The Audit Trail for How LLMs See Your Brand

A TypeScript GEO tool that treats AI visibility like forensic evidence, preserving raw model responses, citations, and analysis history instead of hiding behind a proprietary score.

8 min read • View on GitHub • More from Albert-Weasker

A wide forensic workbench shows raw model output on one side, a preserved evidence archive in the middle, and a polished report on the other. The scene explains that NiubiGEO treats LLM answers as evidence to be saved, traced, and replayed rather than as a score to trust blindly.
NiubiGEO’s core move is simple: preserve the raw answer first, then build the report from the preserved trail.

We want every team to understand how they appear in AI answers at a low cost, with a way to verify every conclusion.

Albert-Weasker, Project Creator / Maintainer · Albert-Weasker/niubigeo README
Key Takeaways

NiubiGEO is not trying to be the prettiest GEO dashboard in the room. Its sharper idea is that an AI visibility audit is only useful if you can replay it later, inspect the inputs, and explain exactly why a brand was mentioned, recommended, or ignored.

The real product is the receipt trail

The project’s strongest claim is also its least marketable one: it treats every LLM answer as forensic material. That means raw provider responses, raw answers, citations, field issues, and source attempts are preserved as evidence, not flattened into a single score.

A hedcut-style portrait of Albert-Weasker based on his GitHub avatar. The image identifies the creator of NiubiGEO and gives a face to the maintainer behind the evidence-first GEO tool.

That matters because GEO is not a stable measurement problem. LLMs cite differently, summarize differently, and sometimes surface answers through reasoning paths that are hard to see after the fact. NiubiGEO’s answer is to keep the record intact before interpretation can muddy it.

Why GEO needs a different kind of tool

Traditional SEO tools were built for ranking systems that left obvious traces. GEO tools have to deal with models that may recommend a brand, cite a rival, or blend both into a single response. The unit of analysis is no longer a blue link. It is a mixed bundle of answer text, citations, and provenance.

DimensionTraditional SEONiubiGEO's model
Primary signalSearch ranking and clicksModel mention, recommendation, and citation trail
EvidenceSERP snapshots and metricsRaw provider response, raw answer, and source attempts
Failure modeMissed keywords or rank dropsOpaque model output that cannot be replayed
InterpretationScore-first reportingEvidence-first audit record
OwnershipOften SaaS-controlledSelf-hosted, inspectable, and replayable

The important shift is philosophical. In a GEO workflow, the report should not be the source of truth. The source of truth should be the preserved evidence that the report summarizes.

How NiubiGEO keeps evidence intact

The architecture separates capture from interpretation. That is what makes re-analysis possible without losing the original evidence.

The save-first, parse-later discipline is the heart of the codebase. The system stores the raw provider response before parsing, so a failed extractor does not erase the evidence. That is a small implementation choice with a big consequence: you can change the parser later and rerun analysis against the same archive.

type RecognitionArchive = {
  rawProviderResponse: unknown;
  rawAnswer: string;
  fieldIssues: string[];
  sourceAttemptMap: Record<string, unknown>;
};

async function persistThenParse(response: unknown) {
  await saveArchive(response);
  return parseStructuredEvidence(response);
}
A close-up drafting table shows one audit prompt with the target brand name crossed out and a neutral keyword version rewritten beside it. In the background, a model pipeline splits into recommendation, mention, and competitor buckets, explaining how NiubiGEO avoids priming the model before measurement.
NiubiGEO treats neutrality as a measurement rule, not a cosmetic prompt tweak.

That neutrality rule is easy to miss, but it is the kind of thing that separates a useful audit from a self-fulfilling one. If you name the brand inside the prompt, you are no longer measuring discovery. You are nudging the result.

Why local re-analysis is the quiet superpower

RecognitionAnalysisRevision is the feature that makes the evidence model pay off. Once raw answers are archived, improved extraction logic can revisit the same material and recompute the analysis without making new provider calls.

What changesWhat stays fixed
Parser rules, extraction heuristics, and field mappingThe raw evidence archive
Report formatting and derived labelsThe original provider response
Analysis quality over timeThe audit trail itself

That creates a cleaner operational loop. The team can improve the analysis engine without changing the historical record, which is exactly what you want if the goal is defensible reporting rather than a one-off dashboard.

Where NiubiGEO sits in the GEO landscape

Tool familyStrengthTrade-off
Commercial GEO SaaSPolished monitoring and packaged scoringOpaque logic and limited evidence ownership
Open-source browser automation toolsCloser to the user interfaceMore brittle and harder to normalize
NiubiGEOPreserved evidence, local re-analysis, and traceable reportsLess glossy, but far easier to audit

That is the real contrast. NiubiGEO is not trying to win by layering more features onto a score. It is trying to make the score secondary to the receipts.

What the codebase says about the project

The repository looks methodical in the way the product idea suggests it should. It is TypeScript-first, split into provider abstractions, product services, UI layers, and Playwright coverage, with documentation that reads like an internal methodology manual rather than a marketing README.

That matters for trust. A tool that audits model behavior needs to be auditable itself, and NiubiGEO’s modular structure, explicit evidence model, and repeatable test coverage all point in the same direction: this is built to be inspected, not merely used.