What a typed knowledge graph does for B2B positioning

A VP of GTM ships a new positioning at the quarterly kickoff. The deck is clean, the messaging is signed off, and within a week the claim has been copied into a battlecard, paraphrased on the pricing page, and pasted into three AI agent prompts. None of those copies know about each other. When a competitor moves next month, nobody can say with confidence which surfaces still assert the old line.

The reflex in that moment is to write a better doc. The better answer is to stop storing positioning as prose at all.

The Commercial Truth manifesto argues that marketing is the last commercial function without infrastructure — engineering got version control, finance got the ledger, and the claims a company makes about itself never got a substrate (Source: Assay brand canon, §10 Truth Graph manifesto). A typed knowledge graph is what that substrate looks like when applied to positioning.

The concept, named plainly

A typed knowledge graph for B2B positioning is a graph of everything a company asserts about itself — products, pricing, capabilities, competitors, proof points, claims — where every node has a type, a confidence score, a source type, and a version, and every edge between nodes is itself typed (Source: Assay knowledge-graph concept canon). It is not a search index over documents, not a key-value store, and not a vector database (Source: Assay knowledge-graph concept canon).

The word doing the work is typed. A positioning statement, a proof point, and a competitor fact are three different kinds of object with three different evidence rules. Storing them as undifferentiated paragraphs erases that difference; typing them preserves it.

In practice that means a node is one of a fixed set of kinds — a PositioningStatement, a Pillar, a BuyerArchetype, a CompetitiveTalkTrack, a Competitor, a ProofPoint — and the system knows which it is (Source: Assay dogfood-loop case study). The claim stops being a sentence in a deck and becomes an object the rest of the stack can reason about (typed knowledge graph for marketing claims).

Why this matters now

This is not a theoretical tidiness argument. It became load-bearing because the substrate underneath GTM changed.

AI agents have proliferated across revenue teams in the last two years — an SDR agent, a CRM agent, a deal-prep agent — and each one grounds from whatever document it was last pointed at. When the source is unversioned prose, every agent inherits the ambiguity and amplifies it into live conversations faster than any human can chase (the truth gap). Five tools end up asserting five versions of the same company (5 AI tools, 5 versions of your company).

A document cannot tell an agent whether a claim is human-verified or machine-guessed. A typed node can, because the type and the source travel with the fact (the confident wrong answer). That is the difference between an agent that knows what it is allowed to say and one that confidently asserts last quarter’s positioning.

The primitives it introduces

A typed knowledge graph introduces a small set of structural primitives that prose has never carried. Each one maps directly onto a failure mode the doc model lives with.

Typed nodes. Every node declares its kind, and the kind constrains its evidence rules (Source: Assay knowledge-graph concept canon). A Competitor node and a ProofPoint node are not interchangeable paragraphs; they are different objects the system treats differently.

Typed edges. Hierarchy and relationship live in edges, not in parent_id columns, and every edge has an explicit type — HAS_CHILD, CITES, CONTRADICTS — drawn from a constrained allowlist (Source: Assay knowledge-graph concept canon). Wildcard edges are forbidden, so the graph cannot quietly connect two things that should never relate.

Intrinsic confidence. Every node carries a confidence score in the range zero to one, and the UI buckets are derived from it rather than stored (Source: Assay knowledge-graph concept canon). A claim is “what we say, with this much certainty,” not just “what we say.”

Intrinsic source type. Every node carries a source type — human-verified, second-party, third-party, or AI-extracted — and the source type sets a ceiling the confidence score cannot exceed (Source: Assay dogfood-loop case study). A node tagged AI-extracted cannot present itself as human-verified regardless of how certain a model claims to be (source-type taxonomy for marketing claims).

Intrinsic time. Every node has a valid_from and a valid_until, so point-in-time queries are first-class (Source: Assay knowledge-graph concept canon). You can ask what the positioning was in Q1, not only what it is today (git for marketing claims).

The primitive with no clean document analogue is the one that earns the word substrate: cascade. Change a node, and every dependent surface — every deck, battlecard, landing page, and agent prompt grounded in it — is flagged for refresh (Source: Assay knowledge-graph concept canon). The change propagates to everything that quoted the claim (every AI agent needs a source of truth).

A worked example

Picture the edit from the opening, but run it on a typed graph instead of a deck. A VP of GTM changes one core positioning line in response to a competitor move, editing exactly one node.

Under the doc model that single change would mean hunting through a deck, a website section, a pricing page, and a set of message variants by memory, hoping nothing was missed. Under the typed model the line is a node, and every surface that quoted it is an edge away.

Because the node carries typed edges, the cascade map flags each dependent surface — the manifesto section that cites it, the positioning canon it rolls up to, the website block that paraphrases it, the pricing copy, and the message variants seeded from it. The single decision propagates to every flagged surface in one reviewed pass, instead of leaking across weeks of half-updated copies.

The flag is automatic; the resolution stays a human-approved step, so the graph never silently rewrites a claim a person has not signed off. That boundary is deliberate, and it is the difference between propagation you can audit and propagation you have to trust.

The exact reduction in propagation lag this produces at customer scale is not yet measured publicly. What is observable already is categorical, not quantitative: positioning moves from a set of unlinked snapshots scattered across surfaces to one versioned object with a provable history (what CRM did for pipeline, commercial truth does for knowledge).

This is not only a thought experiment. The propagation behavior on a real, in-production positioning canon — node counts, the exact cascade events, the source-type distribution — is documented in the Assay dogfood-loop case study, which is the place for those specifics rather than this concept piece (Source: Assay dogfood-loop case study).

What this is not

A typed graph is not an LLM; the models are downstream consumers that read from it (Source: Assay knowledge-graph concept canon). It is not a CRM, which holds prospect data the graph merely references, and it is not a CMS, though it replaces the CMS for fact-bearing content (Source: Assay knowledge-graph concept canon).

It is also not the same as a tidy wiki or a well-kept Notion database. Those are better containers, and a container cannot enforce a source type, cannot cascade a change, and cannot prove its own history (we don’t have a content problem, we have a truth problem). The line between a document and a substrate is exactly the set of primitives above.

Closes / opens

Closes LSO §F.6 — the parallels cluster, and specifically the question of what a typed knowledge graph means when the subject is B2B positioning rather than generic enterprise knowledge.

Opens the next question: if positioning is now a graph of typed nodes with source ceilings and cascade edges, what is the minimal schema a GTM team needs to express it, and how does a VP of GTM adopt it without a full re-platforming? That is the implementation question, and it is where the infrastructure layer argument turns from concept into product.

The methodology Assay is developing for the Commercial Truth Index is meant to gauge the capacity these primitives are built to create: whether a company’s positioning is grounded, calibrated, coherent, and auditable across every surface and every agent. The Index is still in development, so treat that as the direction of travel rather than a settled score.

This essay is grounded in the Assay knowledge-graph concept canon and the dogfood-loop case study (DLC-01). Methodology for the Commercial Truth Index is in development.