A knowledge graph is a way of storing information not as plain text or rows in a table, but as a web of things and the connections between them. Each thing, a person, a company, a product, a place, is an entity, and each link records a relationship: “Marie Curie won the Nobel Prize,” “Paris is the capital of France,” “this part belongs to that machine.” Drawn out, these entities and links form a graph, hence the name. An ontology defines the allowed types of things and relationships, so the data stays consistent.
The term entered mainstream use through Google’s own 2012 announcement, “Introducing the Knowledge Graph: things, not strings,” written by engineering executive Amit Singhal. The post explained Google’s shift from matching keywords to understanding real-world entities, so that a search for “Taj Mahal” could distinguish the monument from the musician and show facts about each directly. That framing, treating search as reasoning about things rather than matching strings, captured why structured knowledge matters.
Why business readers should care: most enterprises sit on a sprawl of disconnected systems, a customer in one database, their orders in another, their support tickets in a third. A knowledge graph stitches these into a single connected view, which is why graphs underpin recommendation engines, fraud detection, supply-chain mapping, and master data management. More recently, knowledge graphs have become a favored way to ground large language models: rather than letting a model answer from memory and risk a confident fabrication, a system can retrieve relevant facts from a curated graph and feed them in, a structured cousin of retrieval-augmented generation that reduces hallucination.
The honest limits are cost and upkeep. Building a good graph means deciding what the entities and relationships are, extracting them cleanly from messy source data, and keeping everything current as the world changes. Graphs that are not maintained drift out of date and quietly mislead. They also do not, by themselves, handle fuzzy similarity well, which is why many modern systems pair a knowledge graph with vector embeddings to get both precise facts and flexible matching.