Methodology (Public)

Section Overview
This section explains the methodological foundations of AI LocalRank — how we gather evidence, distinguish between deterministic and reasoning layers, and calibrate for different platforms.

Evidence Categories

How We Gather Evidence

AI LocalRank collects evidence from publicly available sources — the same sources that AI platforms themselves can access.

PRIMARY SOURCES (Highest Authority)

• Google Business Profile — Hours, address, phone, categories, reviews and ratings, photos and attributes
• Official Website — HTML content and structure, Schema.org markup (JSON-LD), contact information, meta tags and structured data

SECONDARY SOURCES (Corroboration)

• Business Directories — Yelp, TripAdvisor, Apple Maps, Bing Places, industry-specific directories
• Social Media Profiles — Facebook, Instagram, Twitter/X, LinkedIn, profile presence and basic information
• Review Platforms — Aggregated presence across review sites, review count and general sentiment signals

AUTHORITY SOURCES (Entity Validation)

• Knowledge Graphs — Wikidata entity lookup, Wikipedia page presence
• Citation Networks — Where your business is mentioned online, quality and diversity of citing domains

Evidence Types

Evidence Principles

1. Publicly Available Only — We only access information that AI platforms themselves can access. No private data, no authenticated systems.

2. Multi-Source Verification — We collect from multiple sources to detect conflicts and verify consistency.

3. Point-in-Time Snapshots — Evidence is captured at audit time. Subsequent changes require re-auditing.

4. Source Attribution — Each piece of evidence is tagged with its source for transparency.

Deterministic vs Reasoning Layers

Two-Layer Architecture

AI LocalRank operates on two distinct processing layers:

DETERMINISTIC LAYER (Computation — Always Reproducible)

• Evidence extraction and normalization
• NAP consistency checking
• Schema.org parsing and quality scoring
• Directory presence detection
• D×E×A confidence calculation
• Conflict detection between sources
• Scenario-based answer projection

Properties:

• Same inputs → Same outputs
• Reproducible and auditable
• No randomness or variation

REASONING LAYER (Interpretation — Explains Findings)

• Narrative generation for findings
• Natural language explanations
• "Why this is happening" summaries
• Human-readable interpretation of data

Properties:

• Grounded in deterministic findings
• Cannot change computed scores
• Provides explanation, not computation

Why This Separation Matters

Key Invariant

The Reasoning Layer cannot modify Deterministic Layer outputs.

Explanations are generated from scores, never the reverse. This ensures:

• Scores are not influenced by narrative preferences
• Findings are reproducible regardless of explanation
• The system is auditable and trustworthy

Platform Calibration (Conceptual)

Why Calibration Exists

Each AI platform has distinct behaviors:

• Different data source priorities
• Different trust models
• Different response patterns when uncertain
• Different update frequencies

AI LocalRank calibrates for these differences.

Platform Behavior Models

CHATGPT

• Primary sources: Website, Schema.org
• Trust behavior: High trust in structured data
• Uncertainty response: May still attempt answer
• Recency sensitivity: Moderate

PERPLEXITY

• Primary sources: Citations, Multi-source agreement
• Trust behavior: Requires external corroboration
• Uncertainty response: Shows citations explicitly
• Recency sensitivity: High

GEMINI

• Primary sources: Google Business Profile, Knowledge Graph
• Trust behavior: Deep Google ecosystem integration
• Uncertainty response: Knowledge Panel style
• Recency sensitivity: Moderate-High

CLAUDE

• Primary sources: Website quality, Consistency across sources
• Trust behavior: Cross-references for validation
• Uncertainty response: Hedges and qualifies
• Recency sensitivity: Moderate

GROK

• Primary sources: X (Twitter), Real-time social signals
• Trust behavior: Social engagement weighted heavily
• Uncertainty response: Direct style with caveats
• Recency sensitivity: Very High

Calibration Methodology

AI LocalRank calibrates platform models through:

1. Platform Documentation Analysis — Review of published guidance from each platform, understanding of stated data source priorities

2. Behavioral Pattern Study — Observation of how platforms respond to various business types, identification of consistent patterns

3. Signal Weight Estimation — Determination of relative importance of different signals, adjustment based on platform-specific behaviors

4. Ongoing Refinement — Continuous monitoring of platform behavior changes, periodic recalibration as platforms evolve

Calibration Transparency

We disclose that:

• Calibration is based on research and observation
• Exact weights and thresholds are not published
• Platforms may change behavior at any time
• Our models are approximations, not perfect replicas

Scenario-Based Evaluation

What Scenarios Are

Scenarios are realistic questions users might ask about your business.

Universal Scenarios (All Business Types):

• "What are the hours for [business]?"
• "What is the phone number for [business]?"
• "Where is [business] located?"

Category-Specific Scenarios:

Restaurant:

• "Can I make a reservation at [business]?"
• "Does [business] have outdoor seating?"

Healthcare:

• "Is [business] accepting new patients?"
• "What insurance does [business] accept?"

Retail:

• "Is [business] open right now?"
• "Does [business] offer delivery?"

Why Scenario-Based

• Realistic: Matches actual user behavior
• Specific: Tests concrete information needs
• Actionable: Reveals specific gaps
• Comparable: Enables cross-platform comparison

Scenario Coverage

Each audit includes:

• 2 universal scenarios (apply to all businesses)
• 6 category-specific scenarios (tailored to business type)
• Coverage of key user intents

Conflict Detection

What Conflicts Are

Conflicts occur when different sources disagree about the same fact.

Conflict Types

Conflict Severity

Quality Scoring

How Quality Is Assessed

Different evidence types are scored for quality:

Schema Quality

• Presence of required fields
• Tier progression (Foundation → Contact → Trust → Authority)
• Validity of structured data format

GBP Quality

• Profile completeness
• Photo presence
• Review activity
• Hours and attribute coverage

Website Quality

• Accessibility (can AI crawl?)
• Content depth
• Structural clarity
• Contact information presence

Review Quality

• Volume across platforms
• Recency of reviews
• Sentiment signals

Quality Aggregation

Individual quality scores combine through the DxExA framework to produce platform-specific confidence scores.

Reproducibility

Deterministic Outputs

Given the same business at the same point in time:

• Evidence collection yields the same data
• Normalization produces the same structured output
• Scoring calculations return the same values
• Conflict detection identifies the same issues

Snapshot Model

Each audit produces a snapshot — a frozen point-in-time view:

• Versioned with timestamp
• Includes engine version
• Reproducible if re-run with same evidence

Methodological Limits

What We Cannot Measure

Honest Acknowledgments

• Our platform models are approximations
• AI platforms may behave differently than projected
• Results are snapshots, not permanent states
• Improvement is directional, not guaranteed