How we run Assay on Assay: a dogfooded GTM case study

Most vendors say they dogfood. They mean their marketing team sends email through their own tool. Assay means something narrower and harder: every commercial claim the company makes about itself — every positioning line, price, competitive talktrack, and proof point — is a typed node in the same Truth Graph we sell.

The Commercial Truth manifesto argues marketing has never had infrastructure. The strongest way to argue that a substrate exists is to run your own company on it. This is the case study of doing exactly that — and of what it proves, and what it deliberately does not.

For a VP of GTM, the relevant proof is not a feature tour. It is whether the avoided question — what is actually working in our positioning — can be answered as a query instead of an anecdote, because the claims are typed and outcomes attach to them. For a founder, it is whether the company’s voice survives the next ten hires as a system rather than memory.

Why dogfooding is the credibility signal

Category-defining infrastructure has always demonstrated itself on its own substrate. Stripe runs payments through its own API; dbt’s data models are dbt models; Linear’s roadmap lives in Linear. The dogfood loop is the structural signal that separates infrastructure from a tool with a good demo.

The test is whether the company eats the schema, not just the UI. A marketing team using its own email tool proves the email tool works. A marketing team whose every claim is a typed, sourced, version-controlled node proves the substrate works.

It also changes what a buyer can verify. Because the canon is typed and audited, the figures in this very case study — the node counts, the source-type mix, the cascade event — are themselves inspectable records rather than marketing assertions. The substrate is the kind of thing that lets a vendor show its work, and the dogfood loop is the vendor doing exactly that on itself.

The graph today

As of a 2026-05-27 snapshot, the canon is roughly 700 typed nodes — about 80 in the master positioning canon, around 480 across the eight product canons, 42 competitor nodes, and the remainder across verticals, use cases, governance policy, personas, decisions, and essays. The counts are approximate; engineering regenerates exact figures from the live database before external use. Reporting a range instead of false precision is the posture the graph is built to enforce.

The breadth is the point. Positioning, pricing, competitive intelligence, vertical specifics, governance policy, and the essays themselves are all the same kind of object — a typed, sourced node — so a change in any of them propagates through the same machinery. There is no separate system for “the marketing site” versus “the sales deck” versus “the AI prompt”; there is one canon and many surfaces reading from it.

The source-type distribution is published, not asserted: roughly 60% first-party verified, about a quarter first-party unverified drafts, and small slices of second-party, third-party, and AI-derived claims. The AI-derived share is capped below human-verified by construction — the ceiling is a schema constraint, not a soft convention.

What one change looks like

When a high-level node changes, the dependents are flagged automatically. In May, a single positioning-line edit flagged 14 downstream nodes across the brand canon, product canons, pricing copy, and the expression shelves; the team cleared them in one review pass. The flagging is automatic; the resolution stays human-approved — autonomous canon-update is gated by design.

That is the loop in one event: a founder decision, propagated to every dependent surface, with a recorded trail of what changed and why. The same nodes feed the website, the sales templates, and the prompts the AI agents read on their next query.

And every step of that propagation is on the record. Each change writes a hash-chained audit entry — who decided, when, and the evidence behind it — so the question why does our positioning read this way has an answer regenerable from primary records rather than a founder’s recollection.

The loop closes on the canon itself

The graph is not only where Assay’s claims live; it is where they get tested. The expression shelf in the brand canon is the seed for the Variation Shelf — when Calibration ships its first level, AI-generated variants of those expressions become the experimental arms, and the response each earns in market attaches back to the winning node as outcome evidence.

The loop runs recursively. The same Calibration engine that will learn which positioning wins for a customer is the engine that learns which positioning wins for Assay, under the same draft-to-approved promotion protocol customers get — the outcome data attached, the change logged.

This is the part a feature tour cannot fake. A vendor can demo a dashboard; it cannot retrofit a year of its own positioning decisions, each one sourced, scored, and propagated, into a graph after the fact.

What it proves — and what it doesn’t

What it proves for the technical buyer: the Truth Graph is ship-grade, the cascade handles realistic propagation, the source-type ceiling is enforced at schema, and the audit trail is hash-chained and regenerable from primary records.

What it proves for the analyst: that Assay behaves like the infrastructure plays it claims kinship with. The Commercial Truth Index extends the same logic publicly — Assay submits to the benchmark it operates, scored by the same methodology as every other vendor, able to lose on its own test.

What it proves for the founder: that the company’s voice can become institutional memory instead of founder memory. The canon built over two years does not walk out with the next departure or get reinterpreted by the next hire — it is a typed, sourced record the tenth GTM employee onboards into, already propagating to every surface. That is the difference between a system and a set of decisions in one person’s head.

What it does not claim: that every node is perfectly maintained, that the loop is autonomous end-to-end, or that a 700-node canon is customer-scale. A customer at $50M ARR runs an estimated 5,000-20,000 nodes. The architecture scales; the operational discipline at that scale is the stated work ahead.

The honesty is the point. A benchmark you can lose on, a canon with disclosed stale cells, a node count given as a range — these are the calibrated posture the product sells, applied to the vendor selling it.

Closes / opens

The dogfood loop closes on itself. The same engine that learns what positioning wins for customers learns it for Assay; the same benchmark that scores every vendor scores Assay too, with the same audit trail and no preferential treatment.

A company willing to be measured on its own substrate is making a different kind of claim than one that is not.

Closes the dogfood cluster as an authoritative case study (LSO §F.20). Opens the question every infrastructure buyer should ask any vendor — do you run your own company on this, at the schema level, and can you show the audit trail?

This essay is grounded in Assay’s dogfood case study and source-type vocabulary, with buyer context from the brand canon. Node counts are an approximate 2026-05-27 snapshot, regenerated from primary records before publication. Methodology for the Commercial Truth Index is in development.