AI models do not search pages — they query the graph.
Five edge types decide B2B citations: category, buyer, competitor, integration, evidence. A brand that exists as a node with explicit edges gets selected before retrieval. A brand that exists only as a string in retrieved text is at the model's mercy.
From Recognition to Representation
The companion piece — Business Directories and the Entity Layer — covered how AI engines resolve which entity your brand is across the web. This piece is about what happens once the model has resolved you: it stores you as a node in a graph of entities and relationships. Understanding the shape of that graph — what a knowledge graph is, how the model queries it, and which edges matter for B2B brands — is the next layer of AEO craft. It is also the layer most marketing teams have not started to think about.
What a Knowledge Graph Actually Is
A knowledge graph is the structure AI engines use to organize entities and their relationships in a way that makes them easy to query, reason about, and rank. The vocabulary is small:
- Nodes are entities — your company, your CEO, your product, your competitors, your category, your buyers, your integrations.
- Edges are relationships between nodes — founded by, competes with, serves, integrates with, categorized as, cited by.
- Attributes are properties of nodes — funding stage, founding year, headquarters, headcount, pricing band, primary use case.
The model uses this structure for retrieval — not as a parallel to web search, but as the primary representation of what it knows about the world. When a buyer asks the model a question, the model is not searching pages. It is querying its knowledge graph and using retrieved web content to corroborate the graph's answer.
The Model Doesn't Search — It Queries the Graph
Consider a buyer asking ChatGPT: "What project management software is best for distributed teams?" That prompt does not trigger a Google-style page search. It triggers a structured graph traversal that looks roughly like this:
- Identify the category node — project management software
- Filter the entities under that category
- Apply the attribute filter — best for distributed teams
- Walk the competitor edges between surviving entities to verify they're real category peers
- Rank by evidence weight — cited by G2, Wikipedia, Crunchbase, analyst notes
- Return the top three to five entities and synthesize a sentence about each
An entity like Asana shows up in that answer because its node in the graph has all the right edges: category: project management software, best for: distributed teams, competes with: Monday.com, ClickUp, cited by: G2 reviews, Wikipedia, Crunchbase. The model is not deciding to mention Asana from the strength of its homepage copy. It is selecting Asana because the graph already knew the right facts about it.
A brand that wants to be in the answer to that query has to be in the graph with the same edges. A brand whose graph node is missing the best for distributed teams attribute will not survive step 3, no matter how much content it has published.
The Five Edge Types That Decide B2B Citations
Five relationship types do most of the work in B2B AI answers. A brand that is explicit about all five tends to dominate its category; a brand missing two or three is missing from the graph traversal that builds the shortlist.
- Category edges. The "is a" relationship. The model has to know your brand is a [category] — and if the category is contested or emerging, it has to know which name for the category the brand claims. A brand with a clean category edge gets retrieved on category queries; one without is invisible.
- Buyer edges. The "serves" relationship. Marketing teams. Finance teams. Revops leaders. CISOs. Each buyer is itself a node in the graph; an edge from your brand to that buyer is what surfaces you on "best [category] for [buyer]" queries. With 70–80% of the B2B buyer evaluation process now completed before sales is contacted (Stackmatix), the buyer edge is the lever that decides whether the brand even shows up during the research phase where the shortlist actually gets built.
- Competitor edges. The "competes with" relationship. If the graph does not know who you compete with, you will not surface on "alternatives to [rival]" queries — which are some of the highest-intent B2B prompts. Comparison pages, analyst notes, and review-site comparison grids all add to this edge.
- Integration and partner edges. The "works with" or "integrates with" relationship. For B2B SaaS especially, integration edges drive a major category of queries — "what works with Salesforce," "tools that integrate with HubSpot." A brand with integration edges into the buyer's existing stack gets surfaced in those queries; one without does not.
- Evidence edges. The "cited by" relationship. Every off-site source that names your brand adds an evidence edge — G2, Crunchbase, Wikipedia, analyst notes, podcasts. The strength of the evidence edges is the model's confidence weight when ranking competing entities.
Being in the Graph vs Being Mentioned
The most important distinction in AEO at this layer is the difference between two states:
| State 1: String in retrieved text | State 2: Node in the graph |
|---|---|
| Model sometimes mentions the brand when a page that mentions it gets retrieved | Model can return the brand on category, buyer, competitor, or integration queries |
| Survives or dies on the retrieval lottery for any given prompt | Surfaces deterministically when the query matches its edges |
| No relationships expressed — the model has nothing to reason with | Five edge types let the model select the brand before retrieval happens |
| Vulnerable to being summarized incorrectly | Resilient — the graph corrects what individual pages get wrong |
Most B2B brands exist in state 1. They have published content. The content gets crawled. The brand name appears as a string in retrieved text some percentage of the time. They have not made themselves into a graph node with explicit relationships. The teams that move to state 2 — and the move is concrete, not philosophical — get dramatically more citations on the queries that matter most.
How to Feed the Graph Deliberately
Knowledge graphs are not built by hope. They are built from structured signals the model can parse and trust. Five moves feed each of the edge types above directly.
- Schema.org markup with relationships, not just identity. Beyond the Organization block from the entity-layer piece, use
competitor,audience,areaServed,hasOfferCatalog, andknowsAbouton the homepage and capability pages. The schema spec exists for a reason — most B2B sites use 10% of it. - Wikidata claims expressing relationships. A Wikidata entity has a "claims" tab. Each claim is a triple: subject, property, object. Add claims like instance of: AI Answer Intelligence Platform, competes with: [rival], industry: marketing technology. The major models retrieve Wikidata claims as canonical.
- Comparison and "alternatives to" pages that name the competitor edge. The citable page playbook's section on comparison content is also the most direct way to feed competitor edges into the graph. The model reads "[Your brand] vs [Rival]" and adds the edge.
- Customer pages and capability pages that name the buyer edge. "For finance teams" makes finance team a node, and your brand's edge to that node explicit. The same logic applies to "for revops leaders" and any other named buyer.
- Press releases and earned media that state relationships explicitly. "[Your brand], an AI answer intelligence platform competing with X, today announced an integration with Y to serve Z" reads as awkward marketing copy and parses as four explicit graph edges. Earned media that names category, competitor, integration, and buyer in one sentence is graph-rich.
Common Failure Modes
Four patterns explain most of the brands that exist as strings rather than nodes:
- Isolated node with no edges. The brand has a Wikidata entry but no claims attached. A Crunchbase profile with no category. A homepage with Organization markup but no relationship properties. The node exists; the edges that would make it queryable do not.
- Wrong category edge. The graph has tagged the brand under "marketing software" when the company has spent two years pivoting into "AI answer intelligence." Until the category edge is corrected, the brand will surface on the wrong queries and be invisible on the right ones.
- No competitor edges. The brand refuses to name competitors in any owned content. The graph has no competes-with edges. The brand is therefore absent from every "alternatives to X" query — the highest-intent B2B prompt type.
- No buyer edges. The brand talks about itself in general terms — "for modern teams." There is no finance team or revops node attached. The brand is unselectable on any buyer-specific query.
How to Know You're in the Graph
The diagnostic is simple. Run three query types against the five major answer engines monthly:
- Category query — "best [category] for B2B." If the brand surfaces, the category edge is live.
- Buyer query — "best [category] for [named buyer]." If the brand surfaces, the buyer edge is live.
- Competitor query — "alternatives to [rival]." If the brand surfaces, the competitor edge is live.
A brand that surfaces on direct brand queries but on none of the three above is a string, not a node. The fix is the five moves above — and the fix compounds, because once the graph has the edges, every adjacent query starts surfacing the brand too.
How This Pairs With the Earlier Pieces
The entity-layer piece described how the model resolves which entity you are. This piece describes the structure the model stores you in after resolution. The off-site corpus piece described the sources the model retrieves from to populate that structure. The citable page playbook described how to make individual pages retrievable. The team operating model described who in the team ships each piece.
The graph is the connective tissue. A clean entity layer with no graph edges is a node with no relationships — recognized but unselectable. A noisy graph with strong off-site corroboration is a brand the model mentions in passing but cannot reason about. The pieces compound when they are run together, and the graph is where the compounding actually shows up in the answer.
The Standard to Hold
A B2B marketing team running AEO seriously should be able to answer five questions about its graph state at any moment: what category is our brand tagged under in the model's representation, which named buyers do we have explicit edges to, which competitors do we have competitor edges to, which integrations are expressed as relationships, and what evidence sources are weighting our node. If the answer to any of those is "we don't think about it that way," the brand is still living as a string in the model's retrieved text — at the mercy of the retrieval lottery, instead of being selected by the graph query that decides the shortlist.
The model does not search the way Google searches. It queries the graph. The brands that win the next era of search are the ones that get into the graph deliberately — and the only way to do that is to feed it the relationships, on purpose, on every surface.
