Methodology
How SiteNexis
measures AI visibility
Every score, every formula, every measurement decision — documented transparently. Including the Graph Truth Discipline that separates DOM fact from AI inference, and the Graceful Truth Layer that expresses data state with precision. Everything is explainable. Everything is reproducible. Nothing is a black box.
Architecture
Four-Layer Intelligence Stack
Every audit runs through four dependency layers. Layer 4 cannot produce meaningful output without the entity graph, trust signals, and retrieval scores from Layers 2 and 3. The architecture enforces correctness by requiring each layer to be sound before the next builds on it.
Machine Trust Layer
Retrieval simulation · Trust modeling · Temporal authority · Recommendation surfaces
AI Visibility Layer
AI Perception Graph · Citation probability · Retrieval readiness · Visibility scoring
Semantic Intelligence Layer
Entity intelligence · Schema analysis · Content quality · Machine readability
Crawl & Structure Layer
Puppeteer crawl · Chunk extraction · Link graph · SEO signals · Technical performance
Reporting layer
Machine Resource Studio
Machine Resource Studio is the explanation layer above the audit records. It does not replace the scoring engines. It reads their stored results and connects each explanation and recommendation to evidence from the crawl or an issue record.
Scores
The studio shows stored scores and labels each formula. Missing scores remain unavailable.
Narrative
The executive, technical, and AI explanations are generated from the audit state and its evidence.
Action
Recommendations include priority, effort, impact, confidence, and source evidence.
Evidence rule: A recommendation needs a source issue or crawl record. A missing agent result produces a limitation message. It does not produce a zero score or an invented recommendation.
Graph Truth Discipline
Fact Graph vs. Perception Graph
SiteNexis maintains a strict separation between what was observed in the DOM and what was inferred by AI. These two representations of a domain's graph structure are stored separately, traced separately, and never merged. Conflating extracted fact with AI inference is a category error that produces unreliable scoring.
Fact Graph
DOM-extracted only. Every node and edge in the Fact Graph is verified from crawled HTML — page URLs, schema entities, internal links with anchor text and position, external domains. Confidence is 1.0 for all fact records. Nothing is inferred.
- Source: crawled HTML + schema markup
- Confidence: 1.0 (verified extraction)
- Tables: fact_nodes, fact_edges
- Append-only — never overwritten
Perception Graph
AI-inferred only. The Perception Graph models how an AI system represents the domain internally — entities, relationships, citation readiness, disambiguation strength. Confidence is a probability estimate (0–1), never a verified fact.
- Source: AI inference from crawled content
- Confidence: 0–1 (probabilistic estimate)
- Tables: perception_nodes, perception_edges
- derivedFrom field traces each node's origin
Invariant: The Fact Graph never reads from the Perception Graph. The Perception Graph may be derived from Fact Graph data, but the relationship is one-directional. Cross-writing is architecturally prohibited. This ensures that AI inference cannot contaminate the verified DOM record.
Technical Intelligence
Can AI systems reach your content?
The foundation layer. No other dimension of AI visibility is meaningful if AI systems cannot reliably access and extract your content. Layer 1 measures every property that determines whether content enters the AI retrieval pipeline at all.
Crawl Accessibility Score
Definition
Whether AI-specific crawlers (GPTBot, ClaudeBot, Google-Extended, Bingbot) can reach and render page content.
Why it matters
AI systems cannot process content they cannot access. A page blocked by robots.txt for AI crawlers is AI-invisible regardless of content quality.
How measured
Evaluates robots.txt directives per crawler agent, JavaScript rendering completeness, response time consistency, and HTTP response codes.
How SiteNexis fixes it
SiteNexis identifies the specific crawler identifiers blocked and the specific pages affected, with corrective robots.txt directives ready to implement.
Common problems: Accidental AI crawler blocks in robots.txt, JavaScript-only content not rendered server-side, inconsistent response times causing crawler timeouts.
Technical SEO Health Score
Definition
Correctness of the crawl and indexation infrastructure: canonicals, redirects, sitemaps, meta robots.
Why it matters
Technical errors dilute authority signals and cause AI systems to encounter inconsistent or duplicate content, reducing per-URL trust accumulation.
How measured
Programmatic audit of canonical tag correctness, redirect chain depth, sitemap URL accuracy, meta robots consistency, and internal link structural integrity.
How SiteNexis fixes it
Full issue list with severity classification and exact corrective actions per issue type.
Common problems: Duplicate canonical targets, multi-hop redirect chains, sitemap URLs mismatching canonical tags, accidental noindex on priority pages.
Core Web Vitals Score
Definition
Page performance as measured by LCP (Largest Contentful Paint), CLS (Cumulative Layout Shift), and INP (Interaction to Next Paint).
Why it matters
Rendering performance determines whether AI crawlers receive complete content within timeout thresholds. Slow-rendering pages are frequently extracted with missing content.
How measured
Lighthouse integration on top 5 pages by PageRank. Field data from Chrome User Experience Report where available.
How SiteNexis fixes it
Per-page performance issues with specific resource-level recommendations.
Common problems: Slow LCP from unoptimised images or render-blocking resources, CLS from dynamically injected content, INP from heavy JavaScript execution.
Semantic Intelligence
Does your content mean something clear to a machine?
The meaning layer. Accessible content must also be clearly interpretable — entities defined, relationships mapped, and semantic structure coherent enough for AI systems to extract accurate knowledge. Layer 2 measures whether machine understanding is possible.
Entity Confidence Score
Definition
How clearly and consistently the primary entity is defined, attributed, and externally validated across the domain.
Why it matters
Entities are the atomic units of AI knowledge. Ambiguous entity identity prevents AI systems from confidently associating content with a specific real-world object, reducing citation probability across all pages.
How measured
Four sub-dimensions: entity detection (are primary entities explicitly named?), entity consistency (same entity described identically across all pages?), entity coverage (are key entity attributes present?), entity disambiguation (is the entity distinguishable from similarly named objects?).
How SiteNexis fixes it
Entity inconsistency report with specific attribute-level conflicts across pages and schema, with corrective content and schema snippets.
Common problems: Different entity descriptions on homepage vs. about page vs. schema. Missing founding date, category, or description attributes. No disambiguation context for common names.
Schema Completeness Score
Definition
Whether structured data is present, correct, and accurately represents page content.
Why it matters
Schema is the primary machine-readable communication channel between web content and AI systems. Incomplete or inaccurate schema forces AI systems to infer attributes from body text — introducing uncertainty that reduces trust.
How measured
Validation of schema type coverage per page type, attribute completeness per schema type, semantic accuracy (schema claims verifiable in body text), and nested entity relationship coherence.
How SiteNexis fixes it
Generated schema snippets for each detected gap, ready to implement with accuracy requirements noted.
Common problems: Missing dateModified, author without worksFor, schema type mismatched to page category, attributes asserted in schema but absent from body text.
Machine Readability Score
Definition
How much meaning survives the AI extraction pipeline from raw HTML to usable semantic chunk.
Why it matters
AI systems do not read pages — they extract chunks. A page that communicates well to humans but fragments into incoherent chunks produces no AI visibility benefit from its content quality.
How measured
Seven extraction pipeline stages: rendering fidelity, boilerplate ratio, chunk boundary quality, signal-to-noise ratio, heading hierarchy depth, reading order consistency, link anchor quality.
How SiteNexis fixes it
Stage-by-stage breakdown with specific extraction failure points identified per page.
Common problems: High boilerplate ratio from navigation and footer contaminating extracted chunks, paragraph breaks misaligned with semantic units, generic anchor text providing no chunk context.
Authority Intelligence
Does your content satisfy what AI systems are looking for?
The retrieval layer. Semantically clear content must also be structured for AI retrieval — chunk-extractable, query-answer aligned, citation-ready. Layer 3 measures whether your content competes successfully when AI systems are selecting sources.
AI Extractability Score
Definition
Four-dimension measure of how completely an AI system can derive accurate knowledge from a single chunk or small cluster.
Why it matters
AI retrieval is chunk-level, not page-level. Each chunk must independently pass quality evaluation. Content that requires multi-chunk context to make sense produces systematically lower retrieval scores.
How measured
Dimension 1: Entity Clarity (0–25) — named entities explicit and defined in-chunk. Dimension 2: Conversational Readiness (0–25) — H1/title match query form, FAQ structures present. Dimension 3: Chunk Extractability (0–25) — semantic self-containment. Dimension 4: Summarisability (0–25) — clear central claim, no internal contradictions.
How SiteNexis fixes it
Per-chunk extractability issues with specific content restructuring recommendations.
Common problems: Chunks with dangling pronoun references, entities mentioned but not defined, arguments that span multiple chunks without intermediate summary.
Citation Probability Score
Definition
Modelled likelihood that an AI system selects this content as a citation source for relevant queries.
Why it matters
Citation eligibility requires more than accessibility and quality — it requires the specific structural properties that AI citation selection filters evaluate.
How measured
Weighted formula: Factual density (20%) + Claim specificity (15%) + Primary entity authority (15%) + Topical authority depth (15%) + Structural citation readiness (15%) + Temporal freshness (10%) + Trust signal density (10%). Weights configurable in /config/citation-weights.json.
How SiteNexis fixes it
Factor-level breakdown showing which of the seven dimensions is suppressing citation probability, with specific content interventions per factor.
Common problems: Generic assertions without supporting specifics, claims presented without attribution or external validation, broad topic coverage without depth on any single topic.
Semantic Trust Score
Definition
Whether the content demonstrates the authorship, organisational, content, and structural trust signals that AI systems evaluate for source credibility.
Why it matters
AI systems filter sources against trust thresholds before citation. Content from domains that fail trust signal evaluation is systematically deprioritised regardless of individual page quality.
How measured
Authorship trust signals (byline presence, author entity schema, author expertise attribution), organisational trust signals (About page completeness, company schema, policy pages), content trust signals (factual attribution, claim consistency), structural trust signals (HTTPS, schema completeness, contradiction absence).
How SiteNexis fixes it
Trust signal gap report with implementation priority based on score impact per fix.
Common problems: Missing author attribution, no About page entity definition, unattributed factual claims, internal contradictions across pages.
Retrieval Readiness Score
Definition
Composite of machine readability, chunk quality, and query-answer alignment across six query type models.
Why it matters
Retrieval readiness determines whether content competes successfully when AI systems are ranking candidate chunks against a query. High machine readability alone is insufficient if content is not structured for the specific query types that generate retrieval events.
How measured
Query type alignment across: Definitional ("What is X?"), Comparative ("Is A better than B?"), Procedural ("How do I X?"), Evaluative ("Is X good?"), Factual ("When was X founded?"), Navigational ("Find X in Y location"). Per-query-type gap identification.
How SiteNexis fixes it
Query-type specific content structure recommendations with HowTo/FAQ/Organization schema implementation guidance.
Common problems: Procedural content not in numbered list form, evaluative content without schema-backed review signals, factual content without Organization schema anchor.
Machine Trust Intelligence
Does the AI ecosystem trust and recommend your brand?
The trust and recommendation layer. Layer 4 models the dynamic properties of AI trust — how it forms, how it is maintained, how it decays, and how it translates into recommendation surface presence. This is the layer most tools do not reach.
Machine Trust Score
Definition
Composite trust state of the domain from the perspective of AI citation systems — entity credibility consistency, schema trust alignment, external validation depth, contradiction absence, and trust degradation resistance.
Why it matters
Trust in AI systems is domain-level. A single trust signal failure can suppress citation probability across all pages on the domain, not just the page with the issue.
How measured
Entity Credibility Consistency (30%) + Schema Trust Alignment (20%) + External Validation Depth (25%) + Contradiction Absence Score (15%) + Trust Degradation Resistance (10%). Contradiction detection via Claude API on top 20 pages by PageRank.
How SiteNexis fixes it
Trust issue report with source conflict identification, sameAs link health check, and schema alignment gap analysis.
Common problems: Cross-page entity attribute conflicts, schema claiming attributes absent from body text, sameAs links resolving to 404 or wrong entities, contradictory factual claims across pages.
Retrieval Quality Score
Definition
Six-stage simulation of the AI retrieval pipeline: chunk extraction, ranking pressure, summarisation degradation, context truncation, answer formation probability, and citation eligibility filtering.
Why it matters
AI retrieval is a multi-stage process where content can succeed at early stages and fail at later ones. Understanding exactly which stage causes failure is required to fix it efficiently.
How measured
Deterministic simulation on top 30 pages by PageRank. Same content always produces same result. Parameters in /config/retrieval-simulation-model.json. Chunk Stability Index (25%) + Answer Formation Probability (25%) + Summarisation Loss Score (25%) + Citation Eligibility Score (25%).
How SiteNexis fixes it
Per-page retrieval failure report by stage, with fragile claim identification and truncation zone warnings.
Common problems: Chunks that fragment across tokenizer boundaries, facts that distort when compressed into AI summaries, claims that require multi-chunk context to be accurate.
Authority Velocity Score
Definition
Rate of AI visibility growth or decline — change in Entity Confidence, Citation Probability, and external validation signals across consecutive audit cycles.
Why it matters
Velocity distinguishes growing domains from stable or declining ones. A site with a score of 65 growing from 50 is in a fundamentally different position from one declining from 80.
How measured
Delta calculation across consecutive audits: ΔEntity Confidence (30%) + ΔCitation Probability (30%) + ΔExternal Validation (20%) + Update Frequency Score (20%). Requires minimum two audit snapshots.
How SiteNexis fixes it
Velocity trend identification with specific decay signal attribution — which signal is declining and when the decline began.
Common problems: Declining entity confidence without corresponding content changes, external validation sources going offline, update frequency classified as stale (3+ months) or abandoned (6+ months).
Recommendation Surface Score
Definition
Presence and inclusion probability across four AI recommendation surfaces: AI Overviews, chat-based AI, voice assistant retrieval, and autonomous agent discovery.
Why it matters
AI recommendations do not all happen in the same place. A brand visible in AI Overviews but absent from voice and agent surfaces has an incomplete recommendation footprint.
How measured
AI Overviews inclusion probability (30%) + Chat recommendation probability (30%) + Voice retrieval probability (20%) + Agent discovery probability (20%). All surface scores are probabilistic estimates labelled as such in the UI.
How SiteNexis fixes it
Per-surface gap analysis with surface-specific implementation requirements — speakable schema, FAQ markup, robots.txt agent directives, discovery endpoint structure.
Common problems: No speakable schema for voice, no well-known discovery endpoints for agents, FAQ schema missing for AI Overviews inclusion, entity confidence too low for chat recommendation threshold.
Score Formulas
Composite score calculations
Every composite score is the weighted sum of its sub-scores. All weights are documented. All sub-scores are individually explainable. No black boxes.
AI Visibility Score
Machine Readability (15%) + Entity Confidence (20%) + Retrieval Readiness (20%) + Citation Probability (20%) + Semantic Trust (15%) + Schema Completeness (10%)Machine Trust Intelligence Score
Retrieval Quality (20%) + Machine Trust (25%) + Authority Velocity (15%) + Recommendation Surface (20%) + Entity Authenticity Confidence (20%)Score thresholds (used consistently across all scores)
90–100
Excellent
70–89
Good
50–69
Needs Work
0–49
Critical
Principles
Universal explainability requirements
Traceability
Every point deduction maps to a named Issue record with type, severity, description, and recommendation. No score changes without a causal Issue.
Reproducibility
Same content produces the same score. Claude API calls are cached by content hash. No stochastic elements without explicit documentation.
Decomposability
Every composite score exposes its sub-score breakdown. You can inspect exactly which factor is suppressing the overall score.
Delta attribution
Score changes between audit runs map to specific changed issues. You know not just that the score changed but exactly what changed it.
FAQ
Methodology questions
Are SiteNexis scores reproducible?
Yes. Reproducibility is a hard requirement. The same content always produces the same score. The only exception is Claude API-based analysis (contradiction detection, entity extraction) which is cached by content hash — the same content hash always returns the same cached result.
How does SiteNexis differ from a technical SEO audit?
Technical SEO audits address Layers 1 and 2 — crawl accessibility and on-page structure. SiteNexis extends to Layers 3 and 4: AI retrievability, citation probability, machine trust signal analysis, temporal authority modeling, and recommendation surface coverage. Technical audits are a subset of what SiteNexis measures.
Are AI provider scores exact or estimated?
Estimated. SiteNexis does not make live queries to AI systems and cannot access proprietary model internals. All provider-specific scores are modelled estimates based on measurable content signals, labelled as estimates in the UI. The methodology is transparent and based on documented behaviour patterns.
Why does Layer 4 analysis require Pro or Agency tier?
Layer 4 analysis (Retrieval Simulation, Machine Trust, Temporal Authority, Recommendation Surface Mapping, Synthetic Entity Detection) is computationally expensive — it runs 5 additional agents, makes external validation probe requests, uses Claude API for contradiction detection, and requires historical audit data for velocity calculations. The compute cost requires revenue to sustain.
What does "temporal authority requires two audits" mean?
Authority velocity and semantic drift are inherently comparative — they measure change between two states, not the state at any single point. On the first audit, SiteNexis establishes a baseline with velocity: null and status: baseline_established. Subsequent audits compute the delta from that baseline.
What is the Information Gain Engine?
The Information Gain Engine measures what your content adds beyond what already exists in the top SERP results. It uses real-time SERP data from Serper.dev to perform entity gap analysis, question gap detection, and evidence uniqueness scoring. Content with high information gain has the highest probability of being cited by AI systems seeking novel information.
What is the Investigation Layer?
The Investigation Layer is a dual-agent architecture (Judge + Detective) that separates scoring from explanation. The Judge produces scores. The Detective explains why scores are what they are — root cause analysis with evidence chains and confidence intervals. This provides investigation-grade diagnostics, not just issue lists.
System Architecture
Graceful Truth Layer (GTL)
Every API response from SiteNexis is wrapped in a GTL envelope — a structured container that distinguishes between complete data, partial data, and the honest absence of data. This design prevents the system from presenting empty states as zeros, partial results as complete, or loading states as failures.
type GTLEnvelope<T> = {
state: 'complete' | 'partial' | 'empty'
timestamp: Date
data: T | null
}'complete'
All agents completed. All scores are present and verified. The dashboard renders full data.
'partial'
Some agents completed. Scores present are real — not placeholders. Missing scores are shown as absent, not zero.
'empty'
No data yet — first audit not run, or audit failed before any agent completed. The UI shows the honest empty state, not fabricated scores.
GTL is not a loading spinner pattern — it is a data integrity guarantee. A score that hasn't been computed yet is not a zero. A partial audit is not a failed audit. The system expresses what it knows with precision, and nothing more.
Audit Execution
Mission Control: Real-Time Audit Progress
Every audit is orchestrated by 16 autonomous agents across 6 execution stages. Mission Control provides real-time visibility into each stage via a hexagonal agent wheel, circular progress ring, and particle field animation that visualises agent activity as it happens.
6-Stage Agent Wheel
The hexagonal display shows all 16 agents grouped into six execution phases. Each phase lights up as its agents activate, providing instant visual feedback on audit progress without requiring users to interpret log streams.
Stream Reliability Score (SRS)
The SRS measures audit execution health in real time using the CIDER framework: Continuity (no dropped events), Integrity (data passes validation), Completion (all stages finish), Error stability (errors do not cascade), and Recovery (failed sub-tasks retry gracefully).
SERP Intelligence
Information Gain Engine
AI systems prefer content that adds marginal value beyond what already exists in their training data and retrieval corpus. The Information Gain Engine measures what your content contributes that the current SERP cohort does not — using real-time search data from Serper.dev.
Entity Gap Analysis
Compares entities present in your content against entities covered by the top SERP results. Identifies entity gaps your content fills that competitors do not.
Question Gap Detection
Extracts questions answered by SERP competitors and identifies questions your content answers that no competitor addresses — unique retrieval value for AI systems.
Evidence Uniqueness Scoring
Scores claims and evidence in your content by uniqueness relative to the SERP cohort. High-uniqueness evidence has the highest citation probability in AI-generated responses.
Data source: Real-time SERP data via Serper.dev API. The SERP cohort is collected fresh for each audit, ensuring Information Gain scores reflect the current competitive landscape, not cached historical data.
Retrieval Simulation
6-Stage Retrieval Pipeline Simulation
The Retrieval Simulation Engine models the complete decision chain an AI retrieval system makes when processing your content. All simulation is deterministic and algorithmic — no live AI retrieval calls are made. The same content always produces the same simulation result.
Chunk Extraction
Models which chunks would be selected for a given query type. Measures chunk boundary quality and identifies chunks likely to be split mid-thought by fixed-size tokenizers.
Ranking Pressure
Models competitive retrieval position under pressure from high-authority sources. Estimates embedding similarity range and identifies chunks that would be consistently outranked.
Summarisation Degradation
Models meaning loss when retrieved chunks are compressed into an AI-generated answer. Detects facts likely to be dropped, distorted, or hallucinated under compression.
Context Truncation
Models which content falls beyond the retrieval context window for long pages. Identifies critical facts positioned in truncation zones.
Answer Formation
Estimates probability that a chunk becomes part of an AI-generated answer. Combines retrieval rank, chunk stability, and summarisation loss factor.
Citation Eligibility
Models which retrieved chunks pass the AI system citation eligibility filter. Factors: specificity, authority signal density, factual verifiability.
Simulation runs on the top 30 pages by PageRank to control compute cost. Parameters are configurable in /config/retrieval-simulation-model.json. Chunk stability requires 3 independent chunking passes with different boundary strategies to measure variance.
Future Architecture
Investigation Layer: Judge + Detective
The Investigation Layer introduces a dual-agent architecture that separates scoring from explanation. The Judge (SiteNexis scoring engine) produces scores. The Detective (investigation agents) explains WHY scores are what they are — root cause analysis with evidence and confidence framing.
The Judge
Deterministic scoring engine. Produces 12 intelligence scores from measurable content signals. Every deduction maps to a named Issue. No interpretation — pure measurement.
The Detective
Investigation agents that explain why scores are what they are. Traces root causes, gathers evidence across pages, and produces investigation-grade diagnostics with confidence intervals — not just what is wrong, but why it is wrong and how certain we are.
Evidence framework: Every investigation finding includes: root cause identification, evidence chain (which pages and signals contributed), confidence level (how certain the diagnosis is), and actionable fix with expected impact. This is investigation-grade diagnostics — not pattern matching.
Architecture Decision
Machine Trust is Not AI Visibility
A highly visible but low-trust site is a fundamentally different problem from a low-visibility, high-trust one. This is why Machine Trust is scored as a separate tier — not folded into the AI Visibility Score. Trust is formed, maintained, and lost through different mechanisms than visibility.
30%
Entity Credibility
Same entity described consistently across all pages and schema
20%
Schema Alignment
Schema claims verifiable from body text, no over-claiming
25%
External Validation
sameAs links resolve, external sources confirm key attributes
15%
Contradiction Absence
No conflicting facts across pages, no schema-body conflicts
10%
Degradation Resistance
Trust signals stable over time, no sudden unexplained changes
Built by
The mind behind the methodology
Ekeleme David Kelechi
Founder & CEO · Kellyxyhub
Every score formula, every simulation stage, and every detection rule in this methodology was designed from first principles — starting with the question: how does an AI system actually decide to trust and cite a piece of content? The answer is this four-layer stack.
See it in practice
Run a free audit and see every score, every sub-score, and every issue explained.