Waikay Technology: Architecture and AI Monitoring Methodology

1LLM Interrogation System Architecture

Core Technology Infrastructure

Waikay leverages existing InLinks technology foundations:

• Knowledge Graph containing 100 million entities and 10 billion relationships across 17 languages
• Proprietary Named Entity Recognition (NER) algorithm delivering 5x superior performance compared to Google NLP
• Semi-directed graph structure with unidirectional semantic relationships
• Daily synchronization with Wikipedia, DBpedia, and Google Knowledge Graph
• Automated verification and filtering processes ensuring data quality

Knowledge Graph Adaptation for AI Monitoring

Waikay applies the same Knowledge Graph technology as InLinks within a multi-instance architecture:

This multi-KG approach enables precise gap identification between brand reality and individual model perceptions.

2Multi-Model Interrogation Methodology

Target AI Models

Waikay simultaneously interrogates four primary AI models:

Dual Query Strategy

For Gemini, the system executes two distinct queries through prompt engineering:

1. Explicit query restricting responses to training data only
2. Grounded search query enabling real-time web search capabilities

For other models, R.A.G. is deployed.

3Comparative Analysis Methodology

Bidirectional Brand/Topic Analysis

Waikay implements a bidirectional analysis approach, inspired by research on brand associations within language model networks. This methodology interrogates LLMs in two directions:

1. Brand to Entities: Which entities/topics do LLMs associate with your brand?
2. Entity to Brands: Which brands do LLMs associate with specific topics?

This dual interrogation generates comprehensive data on semantic associations, enabling thorough analysis of brand positioning within the AI ecosystem. The resulting data feeds directly into Topic Reports, surfacing the AI Understanding Score and gap classification for each brand.

This approach draws inspiration from Dan Petrovic's research on bidirectional brand association analysis in language model networks.

Tripartite Analysis Process

Waikay systematically analyses:

• The primary site or brand
• Two selected competitor sites
• Comparative analysis of responses across all three entities

4Information Extraction and Gap Classification

Facts Definition

A fact represents a complete declarative statement that LLMs use to describe a brand within brand-sector context prompts. These facts enable identification of accurate, erroneous, or missing information in brand AI representation. Each fact can be traced to sources for direct correction or third-party site optimisation.

Automated Gap Classification

Facts analysis enables gap identification: discrepancies between AI perception and organisational reality. These gaps manifest through missing, incorrect, or insufficient facts. Gaps are automatically categorised based on confidential hierarchisation of named entity types missing from LLM responses, with thresholds derived directly from non-alignments between the brand Knowledge Graph and the LLM Knowledge Graphs.

The system categorises identified gaps as:

• Critical Topic Gaps (critical gaps)
• Significant Topic Gaps (important gaps)
• Moderate Topic Gaps (moderate gaps)

5Infrastructure and Performance

System Architecture

Multi-API Call Management

The system employs a microservices architecture to manage simultaneous interactions with different LLMs. Specific infrastructure details remain confidential due to a patent pending.Patent pending

Rate Limit Elimination

The system encounters no rate limitations in AI model interactions, enabling seamless and continuous analysis.

Configurable Analysis Frequencies

6Scoring Algorithms and Metrics

AI Understanding Score

The AI Understanding Score is the primary output of Topic Reports. It measures how well an AI model understands and represents a specific brand or topic, by constructing comparative Knowledge Graphs for the brand and for each AI model's perception, then calculating alignment scores using NLP metrics applied to the 100-million entity graph. The proprietary algorithm employs contextual disambiguation to analyse 10 billion semantic relationships. Full methodology details remain confidential due to pending patent status.Patent pending

The score directly triggers automated gap classification and feeds the Action Plans module.

AI Share of Voice

AI Share of Voice is an output of Prompt Tracking. Responses to commercial prompts are logged across all AI models over time. Every brand mentioned in each response is identified, normalised to parent brand level via the entity knowledge graph, and aggregated to produce an AI Share of Voice score expressed as a percentage.

The formula uses an open denominator derived directly from AI outputs:

Because the competitive set is derived from what the AI actually says rather than a manually defined competitor list, there is no artificial ceiling on competitor analysis. Brands that begin appearing in AI responses are automatically captured. For more detail on the methodology see What is AI Share of Voice.

AI Topical Presence Score

AI Topical Presence is the third output of Prompt Tracking. Responses to commercial prompts are processed to compute an AI Topical Presence score per brand. The score combines three normalised components (Depth, Breadth, Concentration) using a weighted additive formula: TP = ( α × DepthNorm + β × BreadthNorm − γ × Concentration ) × 100. The score is market-relative, normalised against the highest-scoring brand in the competitive set.

Topical Presence and Share of Voice are plotted together as a quadrant map, with SOV on one axis and Topical Presence on the other. This reveals a dimension that SOV alone cannot: a brand can appear frequently in AI responses (high SOV) without being discussed in depth across a broad range of relevant topics (low Topical Presence). The quadrant makes that distinction visible - a brand with high SOV and low Topical Presence is appearing often but shallowly, while a brand with high Topical Presence and lower SOV has built strong topic associations but not yet achieved reach. Each quadrant points to a different strategic response.

Score tracking

The AI Understanding Score (from Topic Reports) and the three Prompt Tracking scores (AI Visibility, AI Share of Voice, AI Topical Presence) are all tracked over time per model, providing Latest Score, Average Score, and Score History. Precise model attribution with retrieval timestamps enables relative performance comparison between brands across any time window.

7Outputs and Technical Features

Automated Action Plan Generation

Gap-to-Recommendation Transformation

The system analyses identified gaps and automatically generates structured action plans classified by priority and impact. These GEO Action Plans transform analytical data into concrete content optimisation steps.

Source Traceability

Comprehensive URL Mapping

• Complete source identification for all URLs used by each LLM in response generation
• Recurrence classification (usage frequency)
• Model-specific breakdown (e.g. 9 total citations: 4 from ChatGPT, 2 from Perplexity, 1 from Claude, 1 from Gemini)
• Domain-organised structure with precise URLs

Granular Validation

Hallucination Management

Hallucination detection within facts relies on client business expertise, as only domain experts can validate information accuracy regarding their specific brand, products, and services. The system provides all necessary validation elements (sources, context, history) through the validation interface, while final decisions appropriately remain with human expertise.

Fact-by-Fact Validation System

Each extracted statement can be:

Progress Tracking: Interface displays reviewed facts count (e.g. 2/77 facts processed).

8Navigation Architecture

Modular Organisation

The Waikay interface organises into five modules:

• Topic Reports: subject-specific analysis showing how strongly AI models associate a brand with each commercial topic in its market. Outputs the AI Understanding Score and identifies Critical, Significant, and Moderate gaps.
• Prompt Tracking: monitors whether and how often a brand appears in AI responses to specific queries across all tracked models. Produces AI Share of Voice and AI Topical Presence scores.
• Facts: extracts and validates individual declarative statements AI models make about a brand. Each fact is source-traced and can be checked, flagged, or deleted.
• Sources: identifies every URL cited by each AI model in brand-related responses, with recurrence classification and model-specific breakdown.
• Action Plans: automatically generates prioritised GEO/AIO content recommendations derived from gap analysis across Topic Reports and Facts.

Usage Workflow

1. Topic Reports - run an analysis for your brand and up to two competitors to establish baseline AI Understanding Scores and surface topic gaps
2. Prompt Tracking - set up queries to monitor brand appearance frequency, Share of Voice, and Topical Presence over time
3. Facts - review AI statements about your brand, validate correct facts, and flag errors for correction
4. Sources - identify which domains AI models are citing and prioritise outreach accordingly
5. Action Plans - receive auto-generated content recommendations based on identified gaps and begin execution

10Technical Specifications

Data Formats

Import/Export Capabilities

• CSV format for tabular data
• JSON format for complex structures
• Looker Studio integration (under development)

API Development Pipeline

APIs currently in development for:

• Action Plans
• Topic Reports
• Prompt Tracking

Capabilities and Limitations

Unlimited Capacities

• Analysed topics: unlimited
• Analysis frequency: configurable (daily / weekly / monthly / manual)

Processing Performance and Comparative Metrics

Testing across 1+ million web pages demonstrates average performance 5x superior to Google NLP in entity detection, with sector variations (4x factor in real estate/tourism, 12x factor in appliances/tools).

Language Coverage

The system supports 47 languages across hundreds of countries. Performance is optimised for English and French as native languages, with the unified architecture enabling deployment across all supported languages.

Precision Validation

Facts extraction precision exceeds 99%, validated through internal analysis using proprietary comparison tools.

11Competitive Positioning

Technology Differentiation

Unique Market Capabilities

Waikay delivers several capabilities without market equivalent:

Prompt Tracking Differentiation

Waikay's Prompt Tracking is designed for commercial prompts and produces three distinct outputs: AI Visibility (how often a brand is cited across LLMs), AI Share of Voice (brand mentions as a percentage of all brand mentions, using an open denominator derived from AI outputs rather than a manually defined competitor list), and AI Topical Presence (a weighted score across depth, breadth, and concentration of topic coverage). The combination of all three in a single prompt tracking module, with no limit on competitor analysis, is without market equivalent.