A structured representation of knowledge as a network of entities and the relationships between them, enabling AI systems to reason about the connections between concepts, people, organizations, products, and events. Knowledge graphs are used in search, recommendation, product catalog organization, and retrieval-augmented generation systems to provide structured context that supplements the unstructured knowledge of language models.
Also known as entity graph, semantic graph, knowledge network
A knowledge graph represents information as a collection of nodes, which represent entities such as products, people, organizations, concepts, or events, and edges, which represent named relationships between those entities such as “is manufactured by,” “belongs to category,” “was mentioned in,” or “is similar to.” Unlike a relational database where schema and relationships are fixed at design time, a knowledge graph can represent heterogeneous entities and relationships in the same structure without requiring a fixed schema. Triplets of the form (subject, predicate, object) are the fundamental data unit: (product A, belongs to category, outdoor gear) and (outdoor gear, is subcategory of, sports equipment) are two triplets that together express a hierarchical product taxonomy relationship.
Knowledge graphs are used in several distinct ways in AI systems. As a structured retrieval source, a knowledge graph enables precise lookup of entity relationships that would be difficult to retrieve reliably from unstructured document search: “which products are in the same category as product X and were launched in the last 6 months” is a structured query over a product knowledge graph that would require complex information extraction from unstructured text to answer otherwise. As training data for embedding models, the relationship structure of a knowledge graph provides supervision signal for learning entity embeddings that encode the relational context of each entity. As a factual grounding layer for language models, a knowledge graph provides structured factual claims that can be injected into prompts or verified against generated claims to reduce hallucination of specific entity relationships.
Link prediction is the machine learning task most directly applied to knowledge graphs: given a partial knowledge graph with some relationships missing, predict which relationships are likely to be true. This is used to complete product catalogs, identify missing entity relationships, and extend recommendation graphs to cover items without sufficient interaction history. Graph neural networks applied to knowledge graphs learn entity representations that incorporate relational context, enabling link prediction that considers both the direct properties of entities and their neighborhood structure in the graph.
Product taxonomies, campaign knowledge structures, audience segment relationships, and competitive landscapes all have graph-structured relational properties that flat data representations cannot adequately capture. A working ad agency that uses knowledge graphs to represent these structures can power more precise retrieval, more structured reasoning, and richer contextual AI assistance than agencies that rely on unstructured document repositories alone.
Product knowledge graphs improve recommendation precision for e-commerce clients. A product recommendation system that knows about product category relationships, complementary product pairs, and brand portfolio structure can make recommendations that respect these structural relationships: recommending complementary products from the same brand tier, avoiding recommendations of directly competing products from the same catalog, and surfacing products from related categories rather than only the current category. These structural constraints produce recommendations that are more strategically aligned with the client’s business logic than pure collaborative filtering, which has no mechanism for incorporating relational product knowledge.
Campaign knowledge graphs capture the structure of complex multi-campaign programs. An agency managing dozens of campaigns for a major client across multiple brands, markets, and objectives can represent this structure as a knowledge graph: campaigns relate to brands, brands relate to products, products relate to categories, categories relate to audience segments, audience segments relate to channels. This structured representation enables AI-assisted campaign management tools to reason about the relational context of any individual campaign rather than treating each campaign as an independent entity.
Audience relationship graphs extend lookalike modeling to relational audience structures. A knowledge graph of audience entities, such as brands, interests, media properties, and behavioral categories, connected by co-occurrence and affinity relationships, enables more structured audience planning than treating audience attributes as independent features. An audience planner who knows that interest A frequently co-occurs with interest B and that both are bridged by media property C can make more informed targeting decisions than one who sees only the marginal distributions of each audience characteristic independently.
An agency manages search and content marketing for a consumer electronics client with 2,400 active SKUs across 14 product categories. The current SEO and content strategy treats each product independently, producing product pages and supporting content that are not connected by internal linking or topical authority structures. The agency builds a product knowledge graph with nodes for each SKU, category, use case, compatible product, and content topic, connected by edges representing category membership, compatibility, use-case applicability, and topical relevance. The knowledge graph is used to generate three structured outputs: an internal linking map that connects product pages through category hub pages and compatible product recommendations; a topical cluster plan for supporting content that maps to the knowledge graph’s use-case and topic nodes; and a FAQ schema for structured data markup that uses the knowledge graph to identify the most common question patterns associated with each product node. Six months after implementation, organic search impressions for the client increase by 38% and average ranking position for category-level keywords improves by 4.2 positions. The knowledge graph’s relational structure provided the content architecture for building topical authority that individual disconnected product pages could not establish.
The automations and agents module covers how to build and apply knowledge graphs for client AI applications, including the product taxonomy, audience relationship, and campaign structure representations that improve AI system precision beyond what unstructured document retrieval can provide.