What a truth graph for marketing actually is

A truth graph for marketing is a typed, versioned, source-attributed record of every commercial claim a company makes. Positioning, pricing, competitor reframes, proof points, persona insights — each becomes a node, not a sentence buried in a deck. Each node carries a source, a numeric confidence score, and a date past which it is considered stale.

The buyer’s moment that makes this concept matter is mundane and constant. A rep quotes a price the deal desk retired two quarters ago; a battlecard cites a competitor capability that shipped since; an AI agent answers a prospect from a prompt forked from a meeting nobody can find. The Commercial Truth manifesto names the underlying condition: marketing is the only commercial function that has never had infrastructure for what it produces.

Engineering has version control. Finance has the ledger. The claims a company makes about itself live in documents that nobody can confidently call current.

A truth graph is the structure that replaces those documents with something a machine can reason over.

Why a graph, and why now

The shift is from storing claims as text to storing them as structured objects. A document tool assumes the unit of truth is a page; a truth graph assumes the unit of truth is a claim (Source: Truth Graph product canon, COMP-01). That single change is what lets confidence, provenance, and dependency attach to each statement instead of dissolving into prose.

The timing is not arbitrary. Enterprise revenue teams now run several AI agents in parallel — sequence generation, account scoring, conversation analysis, content production — each operating from its own approximation of the company’s positioning, and none of them agree (Source: Commercial Truth manifesto, “What’s broken”).

The same problem is detailed in 5 AI tools, 5 versions of your company. Every agent you add multiplies the number of places a claim can drift.

Treating that as a discipline problem — hire better marketers, refresh the battlecards quarterly — is the answer engineering teams gave in 1995 for lost edits in shared code (Source: Commercial Truth manifesto). It did not work then. The structural fix was infrastructure, and the infrastructure layer makes the parallel explicit.

The primitives a truth graph introduces

The concept is only as strong as the primitives underneath it. Four of them do the work, all drawn from the Truth Graph product canon.

First, the typed node. Every claim has a type — PositioningStatement, Pillar, Competitor, ProofPoint, PricingTier — and each type carries constraints (Source: Truth Graph product canon, §3 and §10). A knowledge base searches text; a typed graph reasons over structure, because the system knows what kind of thing each node is.

Second, the source-type ceiling. Every node records where it came from, and that origin caps how confident the claim is allowed to be. A human-verified fact can reach the top of the scale; an AI-extracted claim is capped at its source’s ceiling; an inference with no direct source is held at 0.50 until a human verifies it (Source: Commercial Truth manifesto, “The four obligations”).

This is the mechanism behind the confident wrong answer problem — confidence is governed by provenance, not asserted.

Third, the confidence score and staleness date. Each node renders with a numeric confidence and a last-verified recency, so “stale” becomes a computed state rather than a feeling in the room (Source: Truth Graph product canon, §10 Confidence Badge). The decay of an unmaintained claim is treated as a property of the node, a dynamic mapped in the knowledge decay curve.

Fourth, the edge. Edges encode hierarchy, dependency, evidence, and conflict between claims (Source: Truth Graph product canon, §1 Description). When a parent claim changes, the graph can trace which downstream claims depend on it — the basis for propagation rather than manual search-and-replace.

A worked example

Take one claim: “We integrate natively with Salesforce.” In a document world that sentence lives in a sales deck, a landing page, a battlecard, and the prompt of an AI agent answering inbound questions. Four copies, no link between them, no record of who is right.

In a truth graph it is one ProofPoint node. It carries a source — say, the engineering release note — which sets its confidence ceiling. It carries a confidence score and a last-verified date.

Edges connect it to the PricingTier where the integration is gated and to the Competitor node it differentiates against.

Now the integration moves from beta to general availability. You edit the node once. Because the deck, the landing page, the battlecard, and the agent prompt are all linked to that node rather than holding their own copies, the change propagates to every surface — the mechanism canon calls Cascade (Source: Truth Graph product canon, §2 FEA-01).

What was an exercise in remembering all four places becomes a single governed edit. This is the difference Git for marketing claims describes between editing a copy and committing to a source.

The opposite case is just as instructive. If that release note never existed, the claim would be an inference with no direct source — and the graph would hold its confidence low and flag it before a rep ever put it in front of a prospect.

Where this lives in the substrate

A truth graph is not a feature bolted onto a marketing tool. It is the substrate: the layer every other engine reads from and that propagation, calibration, and audit all depend on (Source: Truth Graph product canon, §11 Cross-product relationships). Stripe’s payment object is not a product either — it is the thing everything else is built on (Source: Truth Graph product canon, OBJ-03).

The same idea, viewed from the operational side, is what CRM did for pipeline: it gave a function a system of record so that “what is true right now” stopped being a matter of opinion. A truth graph does that for what a company says about itself.

The methodology Assay is developing for the Commercial Truth Index measures exactly the properties this concept makes visible — whether a claim is grounded, calibrated, coherent, and auditable, including for Assay’s own graph under the same rules as every other vendor.

Closes / opens

Closes the Bank C substrate-concept cluster: it gives the plain-language definition of a truth graph for marketing, names its four primitives, and shows the worked claim-to-cascade example that the category brief leaves implicit.

Opens the obvious next question — once you accept the graph, how do you populate it from the decks, pages, and call transcripts you already have without modeling it by hand? That is an extraction question, not a definition one, and it deserves its own treatment.

This essay is grounded in the Truth Graph product canon and the Commercial Truth manifesto. Methodology for the Commercial Truth Index is in development.