Share of Model Voice

The foundational metric of SOMV, measuring how often a brand entity is explicitly named, referenced, or recommended by an AI model across a statistically significant set of relevant queries.

• Calculation: (Brand citations / Total citations across all competitors) × 100
• Granularity: Tracked per prompt category, model version, and time period
• Key distinction: Unlike web scraping for SERP position, this requires systematic querying of black-box AI endpoints
• Example: A cloud provider cited in 340 of 1,000 infrastructure-related generative responses holds a 34% citation share

Not all citations are equal. Positional prominence applies a decay function to citation value based on where a brand appears within the AI's generated response structure.

• First-mention advantage: Brands cited in the opening summary carry disproportionate influence on user perception
• Recency bias: Citations in concluding recommendations often benefit from short-term memory effects in transformer attention mechanisms
• Weighting schema: Position 1 = 1.0x, Position 2 = 0.8x, Position 3+ = 0.5x (descending logarithmic scale)
• Buried mentions: Brands relegated to tangential asides or parentheticals receive minimal effective voice share

Raw citation volume is meaningless without sentiment context. SOMV incorporates a polarity coefficient that adjusts voice share based on whether the AI model portrays the brand positively, neutrally, or negatively.

• Positive multiplier: +1.2x for endorsements, recommendations, or favorable comparisons
• Neutral baseline: 1.0x for factual, non-evaluative mentions
• Negative penalty: 0.5x for criticisms, warnings, or unfavorable comparisons—these citations actively harm brand equity
• Nuance detection: Modern sentiment classifiers must distinguish between constructive critique and outright dismissal
• Example: A cybersecurity firm mentioned 50 times with 80% positive sentiment may outperform a competitor cited 80 times with mixed polarity

SOMV is only as valid as the prompt set used to measure it. Rigorous corpus design ensures the metric reflects genuine market positioning rather than cherry-picked queries.

• Category coverage: Prompts must span awareness, consideration, and decision-stage queries
• Long-tail representation: Include specific, niche queries alongside high-volume head terms
• Competitor normalization: Ensure prompts are brand-agnostic and do not inadvertently favor any single entity
• Temporal stability: Maintain a consistent core corpus while rotating a percentage of queries to capture emerging trends
• Volume threshold: Minimum 500-1,000 prompts per measurement cycle for statistical significance

A brand's SOMV can vary dramatically across different AI models due to training data composition, fine-tuning, and retrieval-augmented generation (RAG) configurations.

• Model fragmentation: GPT-4, Claude, Gemini, and open-source models each produce distinct citation patterns
• RAG influence: Models with live search access may cite different sources than those relying solely on training data
• Geographic variance: Region-specific model deployments may favor local brands due to training data imbalances
• Tracking imperative: Enterprise SOMV dashboards must segment by model to identify where optimization efforts are needed
• Strategic insight: A brand dominant in one model ecosystem but absent in another reveals targeted injection opportunities

SOMV is not a static metric. AI models exhibit temporal bias, favoring recently published or updated content. Voice share must be measured continuously to detect drift.

• Freshness bonus: Content published or significantly updated within 30 days often receives preferential citation
• Decay curves: Citations for stagnant content decline following a power-law distribution over time
• Event sensitivity: Major product launches, controversies, or news cycles can cause abrupt SOMV shifts
• Monitoring cadence: Weekly measurement for competitive categories; monthly for stable niches
• Countermeasure: Continuous entity optimization and fresh content publication maintain citation velocity

A scoring metric that quantifies the contextual importance and prominence of a specific named entity within a given document relative to all other entities mentioned. High salience signals to an AI model that the brand is the primary subject, not just a passing reference.

• Calculated using TF-IDF, neural attention weights, or graph centrality
• Directly impacts whether a model cites a brand in a summary
• Optimized by placing the brand entity in titles, headers, and opening paragraphs

The technical strategies for ensuring AI models correctly attribute sourced information to establish provenance and authority. A citation signal is the digital fingerprint that links a generated claim back to a specific brand source.

• Includes implementing clear authorship markup and publication dates
• Requires consistent use of sameAs linking across the web graph
• Strong signals reduce the likelihood of a competitor receiving undue credit for your data

A high-dimensional vector representation of a brand entity, learned from textual and structural data, that encodes its semantic attributes, associations, and position within a neural network's latent space. This is the mathematical reality of a brand inside a model.

• Proximity to query vectors determines retrieval likelihood
• Can be visualized and audited using dimensionality reduction techniques
• Shaped by the totality of training data mentions, not just a single website

The frequency with which two entities or terms appear together within a defined context, used by search engines to establish semantic relationships and associative authority without direct hyperlinks. If your brand consistently co-occurs with a high-authority topic, the model learns that association.

• A foundational signal for building knowledge graph edges
• Measured across the entire web corpus, not just your owned properties
• Strategic PR and unlinked brand mentions are primary drivers

The technical strategies and feedback mechanisms used to identify and rectify instances where an AI model generates factually incorrect or fabricated information about a brand entity. Uncorrected hallucinations directly erode Share of Model Voice by replacing accurate citations with confident falsehoods.

• Requires systematic output monitoring for brand-specific queries
• Mitigation involves reinforcing the knowledge graph with authoritative, contradictory data
• Techniques include retrieval-augmented generation (RAG) guardrails and fine-tuning on verified corpora

The use of natural language processing to computationally identify and categorize the emotional polarity (positive, negative, neutral) expressed in text mentions about a brand entity. Share of Model Voice must be qualified by sentiment; a high volume of negative citations is destructive.

• Models learn sentiment from the language surrounding a brand mention
• A high negative sentiment score can cause an AI to deprioritize or warn against a brand
• Requires active monitoring of reviews, news, and social discourse at scale