Building a Knowledge Graph for AI: How to Connect Entities for LLMs

Google has run the best-known knowledge graph since May 2012, when it launched with more than 500 million objects and 3.5 billion facts. By 2020 it held over 500 billion facts about five billion entities. In March 2025 Google said its AI Mode can "tap into fresh, real-time sources like the Knowledge Graph."

Be clear about what markup does, though. Google's own AI optimization guide says structured data "isn't required for generative AI search, and there's no special schema.org markup you need to add." So a brand graph isn't a ranking switch. It removes guesswork: it states who founded you, what you sell and what it costs, in a form any parser can read, and it ties your site to the profiles that confirm it.

This guide covers where vector embeddings and graphs meet inside generative AI, a complete JSON-LD graph for a fictional company, how the pieces connect, a checklist and how to validate it. If you're new to the term, start with what a knowledge graph in SEO is.

Three Meanings of "Knowledge Graph" (and the One You Build)

People use the same phrase for three different things. Mixing them up is why so much advice on this topic promises more than markup can deliver.

Three quick answers, since each has its own guide:

• What is a knowledge graph in SEO? It's a database of entities (people, companies, products, places) and the facts that link them, which search engines use to understand what a query is about. Google's is the best known. Our glossary entry covers panels and how to check yours.
• What is the Knowledge Graph Search API? A free, read-only Google API that returns matching entities with an ID, types and a description. It finds entities; it can't add or edit them.
• What is an SEO knowledge graph? The connected schema.org markup across your own site, where every page points at the same entity IDs.

This guide is about the third meaning and how your brand graph feeds it. The markup is the same whichever system reads it.

Introducing the Knowledge Graph

How Vector Embeddings and Knowledge Graphs Meet in Generative AI

Most articles on this topic are either academic papers or knowledge panel tutorials. The part that matters for a brand sits between them: how systems that search by meaning use graphs of explicit facts.

Two ways to store what a machine knows, plus a third

A graph stores facts as triples: subject, relationship, object. "Plannora, founder, Maya Okafor" is one triple. Graphs are exact and easy to audit, but somebody has to build them.

A vector store turns text into embeddings, long lists of numbers where texts with similar meanings sit close together. It can be built from any pile of documents and it matches meaning well. It's weaker at exact facts and at chains like "who founded the company that makes this app?"

The third store is the model's weights, the knowledge it absorbed in training. The original RAG paper (Lewis et al., 2020) calls this "parametric" memory and pairs it with "non-parametric memory": a "dense vector index of Wikipedia, accessed with a pre-trained neural retriever."

Retrieval-augmented generation starts with vectors

Retrieval-augmented generation (RAG) fetches relevant text at question time and hands it to the model. Our explainer on how RAG reduces hallucinations covers what that fixes and what it doesn't. Microsoft's GraphRAG documentation notes that "the majority of RAG approaches use vector similarity as the search technique," which it calls baseline RAG. It also names the weakness: baseline RAG "struggles to connect the dots" when an answer needs facts spread across documents.

Rankbox's own similarity experiment shows what vector matching rewards at the passage level: per percentage point of density, a definition sentence moved cosine similarity 3.2 times as much as a keyword mention (median of 8 open embedding models). Clear statements of what a thing is carry more weight than repeated names.

GraphRAG: text in, graph out, answers on top

In "From Local to Global" (Edge et al., Microsoft Research, April 2024, revised February 2025), a language model reads a document collection and builds the graph itself. It works in two stages: first "derive an entity knowledge graph from the source documents," then write summaries for clusters of closely related entities. The indexing steps are to extract "all entities, relationships, and key claims" from chunks of text, cluster the graph, and summarize each cluster.

At question time the two stores meet. GraphRAG's local search embeds the question, finds entities "semantically-related to the user input," and uses them as "access points into the knowledge graph." It then pulls connected entities, relationships and the source text chunks tied to them. Vectors find the door; the graph shows the rooms behind it.

The results are specific to the test. The paper used two collections of about a million tokens each, podcast transcripts and news articles, with a language model judging pairs of answers. In its head-to-head chart, GraphRAG's answers beat plain vector RAG on comprehensiveness 72–79% of the time on both collections, and on diversity 75–81% (podcasts) and 62–71% (news). Plain vector RAG still gave the most direct answers. At the top level of the graph, GraphRAG needed over 97% fewer tokens than summarizing the source text.

Hybrid retrieval: no single winner

Combining both stores doesn't win everywhere either. In HybridRAG (BlackRock and NVIDIA, August 2024), tested on earnings call transcripts, the hybrid scored highest for answer relevancy (0.96, against 0.91 for vector and 0.89 for graph). The graph alone had the best context precision (0.96 against 0.84 for vector). A survey of graph RAG methods splits the whole family into three stages: graph-based indexing, graph-guided retrieval and graph-enhanced generation.

Databases have moved the same way. Neo4j's vector indexes store embeddings as properties of graph nodes, so one query can match by meaning and then follow relationships. A 2024 roadmap paper on unifying language models and knowledge graphs describes the pairing: models "often fall short of capturing and accessing factual knowledge," while graphs "explicitly store rich factual knowledge."

What the engines actually document

Here is the honest limit. Google documents that AI Mode can use its Knowledge Graph, and that its AI features rest on RAG over its core Search ranking systems. OpenAI says ChatGPT search "sometimes partners with other search providers" and rewrites your question into targeted queries; its help page names Microsoft and Shopify. As of September 2026, no AI search engine's public documentation describes running GraphRAG over a site's schema.org markup.

So don't build for an imagined pipeline. Build a graph that's unambiguous on every path: explicit in markup for systems that parse it, and plain in text for systems that extract entities from prose.

Retrieval changes fast; weights don't

When you publish a page or fix a fact, search-grounded answers in ChatGPT, Perplexity, Google's AI Overviews and Copilot can pick it up once the page is re-crawled and re-indexed, usually within days to weeks. A model's trained weights don't change. What it "remembers" without searching updates only when the vendor trains a new version on newer data. Clear pages and a clean graph help the first path and, over time, add consistent facts to the text future training runs may see. Neither can overwrite the second. If your product has changed direction, our guide to semantic drift covers the full cleanup.

The Seven-Node Brand Graph

This is Rankbox's model for a software company's minimum graph. Seven nodes cover the four things buyers and machines ask about most: who runs the company, what it makes, where its code lives, and what it costs. Each node has a job, a home page on your site, and edges to the others.

Two rules make the model work. Every edge has to be true and visible on your site, because Google's structured data guidelines say not to mark up information that isn't visible to users. And every node needs a stable @id, so each page can point at the same thing.

The Complete JSON-LD Graph for Plannora

Plannora is a made-up project management tool. Every URL, handle and ID below is a placeholder. In particular, Q000000000 is not a real Wikidata item, and the plannora-hq and example- profiles are invented. Swap in your own, and delete any line you can't back up.

{
  "@context": "https://schema.org",
  "@graph": [
    {
      "@type": "Organization",
      "@id": "https://plannora.io/#org",
      "name": "Plannora",
      "legalName": "Plannora Labs, Inc.",
      "url": "https://plannora.io/",
      "logo": "https://plannora.io/brand/plannora-logo.png",
      "description": "Project management software for creative agencies.",
      "foundingDate": "2021-04-12",
      "founder": [
        { "@id": "https://plannora.io/about#maya-okafor" },
        { "@id": "https://plannora.io/about#tom-lindqvist" }
      ],
      "sameAs": [
        "https://www.wikidata.org/wiki/Q000000000",
        "https://www.crunchbase.com/organization/plannora-hq",
        "https://www.linkedin.com/company/plannora-hq",
        "https://github.com/plannora-hq",
        "https://x.com/plannora_hq"
      ]
    },
    {
      "@type": "Person",
      "@id": "https://plannora.io/about#maya-okafor",
      "name": "Maya Okafor",
      "jobTitle": "Co-founder and CEO",
      "worksFor": { "@id": "https://plannora.io/#org" },
      "sameAs": [
        "https://www.linkedin.com/in/example-maya-okafor",
        "https://x.com/example_mokafor"
      ]
    },
    {
      "@type": "Person",
      "@id": "https://plannora.io/about#tom-lindqvist",
      "name": "Tom Lindqvist",
      "jobTitle": "Co-founder and CTO",
      "worksFor": { "@id": "https://plannora.io/#org" },
      "sameAs": [
        "https://github.com/example-tlindqvist",
        "https://www.linkedin.com/in/example-tom-lindqvist"
      ]
    },
    {
      "@type": ["SoftwareApplication", "Product"],
      "@id": "https://plannora.io/#app",
      "name": "Plannora",
      "url": "https://plannora.io/",
      "applicationCategory": "BusinessApplication",
      "operatingSystem": "Web, iOS, Android",
      "brand": { "@id": "https://plannora.io/#org" },
      "publisher": { "@id": "https://plannora.io/#org" },
      "offers": [
        { "@id": "https://plannora.io/pricing#free" },
        { "@id": "https://plannora.io/pricing#team" }
      ]
    },
    {
      "@type": "Offer",
      "@id": "https://plannora.io/pricing#free",
      "name": "Free plan",
      "price": 0,
      "priceCurrency": "USD",
      "url": "https://plannora.io/pricing",
      "itemOffered": { "@id": "https://plannora.io/#app" },
      "offeredBy": { "@id": "https://plannora.io/#org" }
    },
    {
      "@type": "Offer",
      "@id": "https://plannora.io/pricing#team",
      "name": "Team plan",
      "price": 10,
      "priceCurrency": "USD",
      "priceSpecification": {
        "@type": "UnitPriceSpecification",
        "price": 10,
        "priceCurrency": "USD",
        "referenceQuantity": {
          "@type": "QuantitativeValue",
          "value": 1,
          "unitText": "user"
        },
        "billingDuration": "P1M"
      },
      "url": "https://plannora.io/pricing",
      "itemOffered": { "@id": "https://plannora.io/#app" },
      "offeredBy": { "@id": "https://plannora.io/#org" }
    },
    {
      "@type": "SoftwareSourceCode",
      "@id": "https://plannora.io/#sdk",
      "name": "plannora-js",
      "description": "Official JavaScript SDK for the Plannora API.",
      "codeRepository": "https://github.com/plannora-hq/plannora-js",
      "programmingLanguage": "TypeScript",
      "license": "https://opensource.org/licenses/MIT",
      "targetProduct": { "@id": "https://plannora.io/#app" },
      "maintainer": { "@id": "https://plannora.io/#org" }
    }
  ]
}

On 28 September 2026 this block passed the Schema Markup Validator with 0 errors and 0 warnings, and every type and property checked out against the current schema.org vocabulary. The validator also stitched the seven nodes into one connected object, which is the point: no node floats alone.

Five choices in the example, explained

1. The app has two types. SoftwareApplication says what it is. Adding Product allows the brand property, which schema.org defines for products, not software. Multiple types on one node are valid JSON-LD.
2. No ratings. Google's software app rich result requires aggregateRating or review. Add them only if the page shows real reviews. Without them, the Rich Results Test will say the app isn't eligible for that result, which is fine: the goal here is a clear entity, not a star rating.
3. The paid price has a unit. A bare "10" invites a guess. UnitPriceSpecification says it's $10 per one user (referenceQuantity) per month (billingDuration of P1M, the ISO 8601 code for one month). Why a clear, published price matters so much to AI answers is the subject of our pricing page guide.
4. IDs are URLs with fragments. https://plannora.io/about#maya-okafor points at the page where Maya's facts appear. JSON-LD's @id exists "to uniquely identify node objects," and a reference that holds only an @id can point at a node defined elsewhere.
5. Only real profiles go in sameAs. List pages about this exact entity that you control or that describe you: your LinkedIn page, your GitHub organization, a Wikidata item if one legitimately exists.

How the Pieces Connect

@id links inside, sameAs links out

The two properties do different jobs. @id is your own name tag for a node, so the founder of node 1 and the person in node 2 are provably the same thing. sameAs says "this node is the same entity as that page elsewhere." Schema.org defines it as the URL of a page "that unambiguously indicates the item's identity," such as a Wikipedia page, a Wikidata entry or an official website.

Put simply, @id builds the graph and sameAs docks it to everyone else's. Google's Organization documentation says this markup helps it "disambiguate your organization in search results," and that you can list several sameAs URLs.

State every edge in plain text too

Markup helps systems that parse it. Many retrieval systems read prose instead, and graph builders like GraphRAG extract "entities, relationships, and key claims" from text. So write the relationships as sentences on the page: "Plannora was founded in 2021 by Maya Okafor and Tom Lindqvist." "The Team plan costs $10 per user per month." "The official JavaScript SDK lives at github.com/plannora-hq/plannora-js."

Use one name per entity everywhere. If the About page says "Plannora Labs," the pricing page "Plannora" and LinkedIn "Plannora HQ," an extractor may create three entities where you meant one. The entity SEO glossary entry covers consistent naming in more depth.

Split the graph across the pages that show each fact

You don't have to put all seven nodes on one page. Google recommends Organization markup on the home page or an About page, and says you "don't need to include it on every page."

Google doesn't document whether it joins @id references across pages. Popular tools play safe: Yoast's schema approach has every page output all the pieces it references. Our SEO knowledge graph guide walks through that per-page pattern.

Connecting sameAs to Wikidata, Crunchbase and Official Profiles

The plan most guides sell is "create a Wikidata item, add sameAs, done." The sources' own rules are stricter than that.

Wikidata: only if you're notable

Wikidata's notability policy accepts an item if it has a Wikimedia sitelink, or if it describes "a clearly identifiable conceptual or material entity that can be described using serious and publicly available references," or if it fills a structural need. Its self-promotion essay goes further: "Creating an item about yourself, your organisation, or your work is a form of self-promotion and is strongly discouraged," and such items are "likely to be deleted." Paid editors must disclose who pays them, under Wikidata's paid editing policy.

If an item about your company already exists, you may help keep it accurate. Wikidata's autobiography page says editing the item about yourself is generally allowed, as long as statements are sourced, and your own website can only source facts about yourself. These properties mirror the seven-node graph:

Crunchbase: verified employees edit locked fields

Crunchbase lets registered users edit profiles. When a company profile has locked fields, only verified employees can change them. Crunchbase verifies you through an account whose email domain matches your company's, and verified employees get "the exclusive ability to edit the Overview section." Some edits to historical data need Crunchbase staff.

Official profiles: make the links two-way

For LinkedIn, GitHub, X and YouTube, use the exact name and one-line description from your About page, and link each profile back to your homepage. A sameAs link from your site plus a website link from the profile gives any system two matching claims instead of one. Once a knowledge panel appears for your brand, claim it so you can suggest corrections; Google notes that not every panel is claimable yet.

How to Validate Your Knowledge Graph

Validation has two halves: is the markup correct, and do machines now describe you correctly? Run them in this order.

1. Check the JSON. One missing comma breaks the whole block. Paste it into any JSON linter or straight into the validator below.
2. Run the Schema Markup Validator. Google describes it as the tool to "test all types of schema.org markup, without Google-specific validation." Aim for zero errors, and check that it shows one connected object rather than seven loose ones.
3. Run the Rich Results Test. Google's Rich Results Test shows "which Google rich results can be generated" from the page. It only reports types tied to Google features, so some nodes won't appear. If it says the app lacks a rating or review, that's the software app result's requirement, not a fault in your graph.
4. Inspect the live URL. After you deploy, use URL Inspection in Search Console to confirm Google can fetch the page and sees the markup, then request indexing.
5. Look yourself up in Google's graph. Query your brand in the Knowledge Graph Search API. No result is common for young companies and isn't an error in your markup.
6. Ask the AI engines your facts. Run questions like "Who founded Plannora?" and "How much is Plannora's Team plan?" several times in each engine, and score the answers against your fact sheet. The share of answers that get your facts right is the Accuracy Rate in our GEO metrics framework. Recheck two to four weeks after re-indexing, and remember that answers written from training data won't move until a model is retrained.

The Brand Graph Checklist

Use this before you ship, and again each quarter.

Entities

• Every node in the seven-node model that applies to you exists, and none you can't back up.
• Each entity uses one name, spelled the same on every page and profile.
• The product node carries the most specific type that fits (SoftwareApplication, WebApplication, Product).

IDs and edges

• Every node has an @id built from a canonical URL plus a fragment, and it never changes.
• founder, worksFor, publisher, offers, itemOffered and targetProduct all point at @ids that exist.
• Prices carry a currency and a unit (per user, per month).

Links out

• sameAs lists only profiles about this exact entity.
• Each profile links back to your site and uses the same name and description.
• A Wikidata link appears only if an item legitimately exists under Wikidata's rules.

Visibility

• Every marked-up fact is visible on the page that carries it.
• The About page states the key relationships as plain sentences.

Validation

• Schema Markup Validator: zero errors, one connected object.
• Rich Results Test and URL Inspection: page fetchable, markup detected.
• Accuracy Rate recorded for your fact questions, with a date.

Where Rankbox Fits

A graph only works if the pages behind it state the facts clearly. That's where Rankbox helps. Answer-Space Research maps the questions buyers ask ChatGPT, Perplexity and Google in your category and scores them for volume, difficulty and intent. The Citation-Ready Writer then writes source-backed articles, such as a facts page or a comparison, that reach your site through Rankbox's API.

Rankbox doesn't build your JSON-LD graph, and it doesn't track citations or monitor what AI engines say about you. For single blocks of markup, the free schema generator covers Organization, Person, Product and SoftwareApplication; you add the @id links by hand. The Business plan is $49.50 a month with a 7-day trial. See pricing.