Knowledge Graph

Definition

A knowledge graph stores facts as a network of entities (nodes) connected by typed relationships (edges). Instead of unstructured text, you get queryable assertions: "Citare is a product", "Citare is built by Ravi Pillai", "Citare competes with Profound, Ahrefs, Semrush".

Three knowledge graphs matter for AI search:

1. Google Knowledge Graph — the entity database Google has built since 2012; powers the right-rail "knowledge panels" on SERPs and feeds entity context to both classic ranking and AIO/AI Mode.
2. Wikidata — the open knowledge graph that backs Wikipedia and is consumed by every major LLM during training. Strong Wikidata presence = strong default LLM knowledge of your entity.
3. Per-LLM internal graphs — Anthropic, OpenAI, Google, Perplexity all maintain proprietary entity-resolution layers on top of their training data. These aren't queryable from outside but their behavior is observable through prompt outputs.

For a brand, the practical question is: when an LLM gets a query that could refer to multiple namesake entities (e.g. "Citare" — the AI search platform vs. a literary journal vs. an Italian verb), which one does the model pick? The answer is determined by the model's training-time exposure to your entity, which correlates strongly with Wikidata + Wikipedia + structured-data signals on your site.

Why it matters for AI search

Three concrete reasons brands invest in knowledge-graph presence:

1. Disambiguation — common-word or namesake brand names get confused with other entities in LLM outputs. A strong Wikidata entry with sameAs links to your domain, LinkedIn, X, GitHub, etc. tells the model "this is the canonical 'Citare' you want."
2. Citation grounding — when AIO or Perplexity composes an answer, it draws on cited sources for facts and a knowledge graph for the entity skeleton. Brands with strong knowledge-graph presence get cited more confidently because the model can ground claims in entity assertions.
3. Default-knowledge surface — even without web search, an LLM can answer questions about well-known entities from its training data alone. If your brand is in Wikidata + Wikipedia, the model has default knowledge of you that doesn't depend on a search retrieval succeeding.

The Citare Brand Radar use case

The Citare Brand Radar pipeline begins with a per-brand knowledge graph (Stage 1 of the 5-stage pipeline). Before any LLM dispatch, the pipeline constructs a structured representation of:

• The brand entity (name, domain, founded date, founder)
• Product entities (each major product, its category, its competitors)
• Persona entities (3-4 buyer ICPs)
• Query topics (the JTBD questions buyers ask)
• Competitor entities (direct + adjacent)

This knowledge graph is what gets passed into the Sonnet dispatcher to generate persona-anchored queries. Without it, you get generic queries that don't match real buyer language; with it, queries land where your buyer actually asks.

Schema for knowledge-graph presence

The structured-data signals that feed the Google Knowledge Graph (and indirectly the LLM graphs):

• Organization JSON-LD with @id, name, alternateName, url, logo, sameAs (array of authoritative profile URLs)
• Person JSON-LD for founders + authors with similar sameAs structure
• Product or SoftwareApplication JSON-LD per product
• Consistent @id across your site (use canonical absolute URLs)

The single highest-leverage move for a new brand is: get a Wikidata entry (free, manual submission), then add Organization JSON-LD with sameAs linking to your Wikidata page + LinkedIn + X + GitHub.

Common pitfalls

• Treating knowledge graph as a "Google thing." Wikidata matters more for LLM default-knowledge than Google does — Google's graph informs SERPs, but Wikidata informs every LLM training pipeline.
• Inconsistent entity assertions across pages. If your About page says you were founded in 2024 and your LinkedIn says 2023, the model can't pick an authoritative answer and may decline to assert either.
• Schema without sameAs. Organization JSON-LD without a sameAs array is local entity data with no external linkage; it doesn't bridge to the broader knowledge graph.