Entity Salience

Salience is a relative metric, not an absolute count. An entity's score is calculated by comparing its prominence against all other entities in the same document. A CEO mentioned once in a 200-word bio may have higher salience than a product mentioned ten times in a catalog.

• Mechanism: Algorithms analyze syntactic position (subject vs. object), discourse structure, and co-occurrence density.
• Key Distinction: High frequency does not guarantee high salience; a central theme mentioned sparingly can dominate the score.

Salience algorithms weigh grammatical role heavily. Entities appearing as the grammatical subject of main clauses receive higher scores than those in prepositional phrases or subordinate clauses.

• Title Weighting: Entities in the document title or H1 heading receive a multiplicative score boost.
• First Mention Advantage: The initial paragraph carries disproportionate weight, as it typically establishes the document's primary topic.

An entity's salience is amplified by its dense co-occurrence with other high-salience entities. This creates a reinforcing network effect within the document's entity graph.

• Graph Centrality: The document is modeled as a graph where nodes are entities and edges are co-occurrence links. Salience correlates with eigenvector centrality in this graph.
• Example: In an article about Tesla, 'Elon Musk' and 'Gigafactory' mutually reinforce each other's salience through repeated joint mentions.

A critical distinction: term frequency-inverse document frequency (TF-IDF) measures keyword importance, while salience measures conceptual centrality. An entity can be highly salient with a single, strategically placed mention.

• Anti-Pattern: Keyword stuffing increases frequency but often decreases per-mention salience by diluting syntactic prominence.
• Optimization: Place the target entity as the subject of the first sentence to maximize its salience score with minimal repetition.

For brand entities, AI models aggregate salience signals across an entire corpus. A brand that is consistently the primary subject across its owned web properties builds a strong, stable entity profile.

• Canonicalization: Consistent use of the exact entity name and linked references to a single Entity Home page consolidates cross-document salience.
• Risk: Inconsistent naming (e.g., 'IBM' vs. 'International Business Machines') fragments salience scores across multiple entity IDs.

In RAG and summarization tasks, the model's attention mechanism is biased toward high-salience entities. These entities are more likely to be preserved in a summary, cited in an answer, or used as anchoring concepts.

• Citation Probability: A source document where a brand has high salience is significantly more likely to be cited by the AI when answering a query about that brand.
• Sentiment Coupling: The sentiment expressed near a high-salience entity disproportionately influences the model's overall sentiment toward that entity.