The problem

Every team in a large go-to-market organization can now generate content at machine speed. Decks, briefs, battlecards, emails, answers to customer questions. The bottleneck is gone, and a different problem took its place. None of those tools know what is approved and true for the company, so they guess. Each team re-derives the same claims from whatever documents happen to be in the prompt, and the output drifts a little further from the source every time.

I wrote about both halves of this on the blog: the output layer is becoming a commodity, and the canonical layer is what is left to differentiate on. This page is about what we did with that idea.

The idea

Product marketing already owns the artifact that answers "what is true and how do we say it": the message house. There were dozens of them across the portfolio, each a structured statement of positioning, pillars, proof points, and audience-specific framing. They were also PowerPoint files, invisible to every AI tool in the building.

The bet was simple. Treat the message houses as a database rather than as documents. Make them retrievable with their structure intact. Then put that retrieval layer under every AI workflow that touches messaging, so a seller's assistant, a campaign generator, and a customer-facing bot all ground on the same approved source.

What we built

• A canon schema. Each message house became a canon domain with typed entries: positioning statements, pillars, proof points, objections, audience variants. Structure mattered more than prose. The model needs to know which sentence is a claim and which is evidence.
• Retrieval that respects the structure. A knowledge graph for the relationships between products, audiences, and claims, plus vector retrieval for the loose queries people type. A question about a specific persona and product pulls the right pillar with its proof points attached, not a paragraph that happens to mention both words.
• An approval lifecycle. Entries carry a status, an owner, and provenance. Draft is not canon. Only approved entries reach the retrieval layer, and stale ones are flagged instead of served.
• Grounded generation. Artifact generators build from retrieved entries and report an alignment score: how much of the output maps back to approved canon. Low alignment triggers a review rather than a rejection.
• One interface for every tool. The layer is exposed through the Model Context Protocol, so any AI assistant or workflow that speaks MCP can ground on it without a custom integration.

What changed

The visible effect was consistency: the same claim, phrased the same way, in a seller's email, a campaign brief, and a chatbot answer. The less visible effect was on the product marketing team. Once the message house was a living record with an owner and an approval state, maintaining it stopped feeling like updating a deck and started feeling like maintaining a source of truth.

The other change was in where PMM sits. When content production is no longer the job, the judgment that goes into the canon becomes the job. That is the subject of The Canonical Layer Is the Job, the talk this work turned into.

What I would do differently

• Start with two domains, not the whole portfolio. The structure is easy to get wrong and expensive to migrate. Prove the retrieval on a narrow slice before converting everything.
• Add alignment scoring first. It came after the generators, and it should have come first, because it is the metric that shows whether the layer is being used or bypassed.
• Treat the approval workflow as a product. The hard part was never retrieval. It was getting busy owners to approve entries on a schedule. Notifications, digests, and a short path from "this is stale" to "this is fixed" mattered more than model choice.