Knowledge graph: definition and how it works

A knowledge graph is a structured store of facts about things. It has three parts:

• Entities (nodes): a company, a person, a product, a city.
• Relationships (edges): named links between entities, such as "founded by", "headquartered in" or "subsidiary of".
• Properties: values attached to an entity, such as a founding date or a website.

One fact is usually written as a triple: subject, relationship, object. "Company X, headquartered in, Austin" is one triple.

You'll also see the terms entity graph and knowledge base. An entity graph is a knowledge graph described from the entity's point of view. A knowledge base is the broader category: any organised store of facts. A knowledge graph is a knowledge base built specifically from linked entities. Wikidata calls itself a free, collaborative knowledge base, and its data is structured as a graph of items and statements.

Every entity gets a stable identifier. That identifier, not the name, is what the graph links to. It's how a graph keeps a software company called Mercury separate from the planet and the element.

Facts enter a graph in a few ways:

1. People edit it directly, as on Wikidata.
2. The operator extracts facts from documents it has crawled.
3. The operator imports other structured datasets.

Software then queries the graph by following edges. "Who founded the company that makes product Y?" becomes two hops: product to maker, maker to founder. That is the main advantage over plain text. The facts are explicit, so they can be joined, checked and counted.

The weakness is the same as any database. A graph is only as accurate as what went in, and a wrong or outdated fact keeps spreading until someone fixes it.

Your brand is an entity whether or not you manage it. The question is what each system does with the facts about it.

It helps to split AI answers into stages:

• Parametric: what the model learned in training, frozen at its knowledge cutoff. Public facts about your company may have been in that training text. Nobody can change this until the next model.
• Retrieval: when search is on, the engine fetches pages at answer time. See grounding and search on vs search off answers.
• Generation: the model writes from whatever it has.

No engine documents, with measurement, how much a knowledge graph shapes any of these stages. So treat claims that a graph entry "controls" your AI answers as unproven.

What holds up is plainer. AI platforms do get specific facts wrong about specific brands. Stating each fact (price, founding year, headquarters) clearly, on a page crawlers can fetch, gives an engine a sentence it can quote. Keeping that fact identical on your site and in public sources removes contradictions. For the practical side, see how to manage your brand as an entity and whether your Wikidata entry affects AI answers.

"Schema markup puts me in the knowledge graph and gets me cited"

Google says structured data isn't required for its generative AI features, and no special optimization is needed for AI Overviews or AI Mode. Ahrefs' matched test of 1,885 pages found AI Overview citations fell 4.6% after schema was added. Every page in that study was already heavily cited, so effects for uncited pages are untested. Existing markup does no harm. It just isn't a lever for AI answers.

"A language model is a knowledge graph"

It isn't. A model stores patterns in its weights, not explicit triples you can look up or edit. That's one reason it can state a fact confidently and wrongly.

"I need a Wikipedia page to exist as an entity"

Entities exist in many sources. Whether a Wikipedia page specifically changes AI answers is covered in do I need a Wikipedia page.

• Wikidata: the largest open, editable knowledge base.
• Wikidata hygiene for your company item: a step-by-step check of your own entry.
• Grounding: how an AI answer is tied to retrieved sources.
• Knowledge cutoff: the date after which a model learned nothing in training.