What is a knowledge graph, in simple terms?

A structured record of the things an organisation knows about (products, plans, features, customers, documents) and how they relate, such as which plan includes which feature. Each fact is stored once and linked to the document it came from.

How is a knowledge graph different from a folder of documents?

A folder holds prose in which facts are scattered and sometimes contradictory. A graph holds each fact once, with its relationships and source made explicit. A system searching documents has to guess which of five mentions of a price is current; a graph has one price with one source.

Do I need a knowledge graph to use AI for marketing?

No, but without some form of grounding every factual claim in an AI draft needs checking by someone who knows the product. A graph, or at least retrieval from authoritative documents, is what lets a draft carry its own sources.

What should go into a marketing knowledge graph?

Current product and plan facts, capabilities and integrations, the claims you make with the evidence behind each, customer results that customers have agreed to be named for, your positioning language, and the source documents themselves with dates.

How does Veritas use a knowledge graph?

Veritas builds a graph from the documents you upload and drafts from it. Every knowledge-based claim carries a numbered citation to its source, and claims that cannot be traced to your material are left out rather than invented.

Knowledge Graphs for Marketing Teams, in Plain English

The one-sentence definition

A knowledge graph stores what an organisation knows as things and the relationships between them. Instead of a folder of documents that each mention your product, price and customers somewhere in their prose, a graph has a node for the product, a node for the price, and a link between them, with a note of which document the fact came from.

That is the whole idea. The rest of this post is about why it matters for anyone publishing content with AI. The glossary entry has the short version.

Why Google built one

In May 2012, Google introduced its Knowledge Graph, describing a model that "understands real-world entities and their relationships to one another: things, not strings." At launch it contained "more than 500 million objects, as well as more than 3.5 billion facts about and relationships between these different objects."

The problem it solved is familiar to anyone who has searched for a word with two meanings. A string is ambiguous; an entity is not. Once a search engine knows a query is about a particular company, and knows facts about that company, it can answer rather than match. The same shift is why entity recognition matters so much for AI search today.

Documents versus a graph

Picture a small software company with a pricing page, three product sheets, a sales deck and a dozen blog posts. Its current price appears in five of those documents, and two of them are out of date.

A person asked what the Pro plan costs might find the right answer by checking dates. A system that searches documents finds five passages and has to decide which to trust. A graph has one node for the Pro plan's price, linked to the pricing page it came from. It is either right, or visibly stale.

The same holds for features, integrations, customer claims and product names. A graph makes each fact exist once, with a source, and makes relationships explicit: which plan includes which feature, which case study supports which claim.

Why this matters for AI content

A general language model writes what is statistically likely about a company like yours. When it does not know a fact, it produces a plausible one, which is how AI content ends up with invented features and wrong prices.

The standard fix is to give the model the facts before it writes. Retrieval-augmented generation, introduced by Lewis et al. in 2020, pairs a model with a store of documents it can retrieve from. The authors found such models "generate more specific, diverse and factual language" than a model relying only on what it absorbed in training.

Retrieval from documents is a large improvement. Retrieval from a graph goes further in two ways that matter for marketing:

Consistency. One fact, one place. If every draft reads the price from the same node, every draft states the same price.
Citation. Because every node carries its source, every claim in a draft can point at the paragraph it came from. A reviewer checks the citation instead of rewriting the sentence.

What research says graphs add

Plain retrieval is good at narrow questions answered by one passage. It struggles with broad questions whose answer is spread across many documents, such as what the main themes in a year of customer interviews are.

Microsoft researchers took on exactly that problem in GraphRAG (Edge et al., 2024), which uses a language model to build a knowledge graph over a collection before answering questions about it. For "global sensemaking questions over datasets in the 1 million token range," they report "substantial improvements over a conventional RAG baseline for both the comprehensiveness and diversity of generated answers."

For a marketing team, the broad questions are often the useful ones: which proof points support this claim, which features come up most in objections, what we have said about a competitor. A graph makes those answerable with sources attached.

What to put in a marketing knowledge graph

Products and plans: names, what each includes, prices and limits, taken from the pricing page and product documentation rather than old decks.
Capabilities and integrations: what the product actually does, linked to the document that says so.
Claims and proof: each claim your marketing makes, linked to its evidence, or marked as unsupported.
Customers and results: only what customers have agreed to be named for, with the permission recorded.
Positioning and language: how you describe the category, the problems you solve, and the words customers use.
Sources: every document, dated, so stale facts can be found and replaced.

What a graph does not fix

Wrong inputs. A graph built from an outdated deck faithfully reproduces the outdated deck. Start from current, authoritative documents.
Maintenance. Facts change, and someone has to own the update when a price or feature does.
Missing knowledge. A grounded system should decline to invent what the graph does not contain. Drafts on thinly documented topics come out shorter, which is correct and occasionally inconvenient.
Review. Citations make review fast. They do not make it optional.

How Veritas uses one

Veritas builds a knowledge graph from the documents you upload and drafts from it. Every knowledge-based claim carries a numbered citation that opens to its source, and when a claim cannot be traced to your material it is left out rather than filled in. That is the standard described in Citation-First AI, and a knowledge graph is the structure that makes it practical.