Knowledge graph
A knowledge graph stores knowledge as a network: people, companies, places, or products are nodes, and their relationships are named edges that machines can query directly.
A knowledge graph stores the relationships themselves as data.
A knowledge graph is a body of data shaped as a graph. Its nodes stand for things in the real world, called entities, and its edges for the relationships between them. That is the definition in a 2021 survey by Aidan Hogan and colleagues. The sentence “The supplier is based in Hamburg” becomes two nodes and an edge labeled “based in.” A schema, often called an ontology, defines which kinds of nodes and relationships exist.
The phrase has appeared in research since at least 1972. Its current meaning traces back to Google, which introduced its Knowledge Graph in 2012. Amazon, eBay, Facebook, LinkedIn, Microsoft, Uber, and others then announced knowledge graphs of their own.
In search, Google's graph helps decide whether a company shows up as a distinct entity, for example in a knowledge panel. Inside the company, a knowledge graph connects data from purchasing, engineering, and sales and gives AI applications connections that no single document contains.
Chapter 1 of 7
How a knowledge graph is built.
The smallest unit is a statement with three parts: subject, predicate, and object. The W3C's RDF standard calls this a triple, and an RDF graph is simply a set of triples. Every thing gets a globally unique identifier, an IRI. Two occurrences of the same IRI refer to the same thing, whatever source they come from.
Property graphs are the other common model: nodes and edges carry extra attributes, such as a date on a “supplies” relationship. Hogan and colleagues cover both as standard data models.
The schema gives the connections meaning. An ontology states that a supplier is a company, or that only components can meet a standard. With rules like these, a system can infer facts nobody entered explicitly. Graph query languages can follow paths across any number of edges.
Chapter 2 of 7
Google made the term famous in 2012.
On May 16, 2012, Google introduced its Knowledge Graph under the motto “things, not strings”: search should understand things, not just match character strings. According to Google, the graph then held more than 500 million objects and more than 3.5 billion facts and relationships, drawn from sources including Freebase, Wikipedia, and the CIA World Factbook.
Google's first example was things that share a name. Someone searching for “Taj Mahal” might mean the monument or the musician. The graph treats them as two entities.
The graph is most visible in the knowledge panel. Google describes knowledge panels as information boxes that appear when you search for people, places, organizations, or things in the Knowledge Graph. The information comes from various sources across the web and updates automatically as it changes there.
Chapter 3 of 7
Entity SEO makes a company a distinct entity.
Entity SEO is the work of getting search systems to recognize a company, person, or product as one distinct thing. Nobody can write directly into Google's graph, but Google documents helpful signals.
The main one is structured data using schema.org, a vocabulary founded jointly by Google, Microsoft, Yahoo, and Yandex. Google says organization markup helps it understand a company's administrative details and tell it apart from similarly named ones, and it influences the logo in the knowledge panel. The sameAs property links a website to profiles on other sites. Google recommends JSON-LD.
Wikidata is a free, multilingual knowledge base anyone can edit, with data released under CC0. Each item there has an identifier made of Q and a number. The subject of a knowledge panel can claim it and suggest changes; everyone else uses the panel's feedback link.
Chapter 4 of 7
Inside a company, a knowledge graph ties scattered data together.
Many business questions span several links, for example which products contain parts from a supplier whose certificate is expiring. The answer usually sits in several systems at once.
A knowledge graph stores the relationships themselves as data. A relational database assembles links from tables at query time, usually enough for questions spanning one or two links. A vector database finds similar text through embeddings but knows no named relationships. A graph pays off when questions regularly cross many edges and new kinds of relationships get added without restructuring tables.
The hardest part is identity. The same company may appear under an abbreviation in purchasing and under its full name in accounting. Before data goes into the graph, duplicates like these have to be merged into one entity.
Chapter 5 of 7
GraphRAG combines a knowledge graph with a language model.
Retrieval-augmented generation (RAG) means a language model gets relevant passages from a document collection before it answers. In 2024, a Microsoft Research team showed that conventional RAG fails on global questions such as “What are the main themes in the dataset?”
Their method, GraphRAG, has a language model build a knowledge graph of entities and relationships from all the documents, then summarize groups of closely related entities. Each summary yields a partial answer, and the model combines them. On datasets of around one million tokens, answers were substantially more comprehensive and diverse than with conventional RAG.
Two caveats. Building the graph itself takes language model calls across every document. And whatever the model gets wrong in the process ends up in the graph. A curated graph of verified master data is a different thing.
Chapter 6 of 7
What good implementation looks like.
It starts with the business questions the graph should answer, not with picking a graph database. The database alone isn't a knowledge graph; a schema and maintained data make it one.
The schema is documented and small enough for business teams to follow. Every entity has exactly one identifier, and every statement shows where it came from. Someone owns each area and adds new products, suppliers, or standards. General facts such as places can come from Wikidata identifiers instead.
Most of the effort goes into the schema, merging duplicates, and upkeep. Open standards such as RDF keep the data independent of any one product. Entity SEO applies the same discipline externally: the same name, location, and profiles everywhere, in machine-readable form on your own website.
Chapter 7 of 7
Where the term stops working
A knowledge graph only holds the knowledge someone models and maintains; left alone, it goes stale. A graph database merely stores it. A vector database finds similar items instead of named relationships, an ontology is the schema rather than the data, and the knowledge panel shows Google's graph, not your own.
Frequently asked questions
What is a knowledge graph?
A knowledge graph is a body of data that stores knowledge as a network of nodes and edges. Nodes stand for things such as people, companies, or products, and edges for named relationships between them. A schema defines which kinds exist.
Is a knowledge graph the same as Google’s Knowledge Graph?
Not necessarily. Lowercase, “knowledge graph” describes any body of data that stores knowledge as nodes and named relationships, inside a company or on the web. Capitalized, “Knowledge Graph” usually means Google’s graph, which made the term famous in 2012. In German, both are often called a Wissensgraph.
How is a knowledge graph different from a database?
A knowledge graph stores relationships themselves as named data, while a relational database assembles them from tables at query time. New kinds of relationships can be added without restructuring tables. The graph still lives in a database, usually a graph database.
What is an ontology in a knowledge graph?
An ontology is the schema of a knowledge graph. It defines the classes of things, the allowed relationships, and the rules, which can also be used to infer new facts.
How does a company get a Google knowledge panel?
Google shows knowledge panels for entities in its Knowledge Graph and fills them from various sources across the web. Schema.org markup, consistent details, and sameAs links help Google identify the company. An existing panel can be claimed so the company can suggest changes.
What is GraphRAG?
GraphRAG is a Microsoft Research method in which a language model first builds a knowledge graph from documents and then generates answers from it. It helps with questions about an entire document collection. Errors made while building the graph end up in it.
Sources
- Hogan et al.: Knowledge Graphs, ACM Computing Surveys 54 (4), 2021, arXiv:2003.02320
- Google: Introducing the Knowledge Graph: things, not strings, Amit Singhal, May 16, 2012
- Google: About knowledge panels, Knowledge Panel Help, as of November 2026
- W3C: RDF 1.1 Concepts and Abstract Syntax, Recommendation, February 25, 2014
- Schema.org: About Schema.org, as of November 2026
- Google Search Central: Organization structured data, documentation, as of November 2026
- Wikidata: Introduction, as of November 2026
- Edge et al. (Microsoft Research): From Local to Global: A GraphRAG Approach to Query-Focused Summarization, arXiv:2404.16130, 2024