Why does Confluence go stale as a sales source of truth?

Because a wiki versions pages, not claims, and nothing propagates. When one fact changes on a page, the decks, templates, and AI agents that quoted it are not updated — so the page drifts out of date while still looking authoritative.

Is Notion or SharePoint any better than Confluence for sales messaging?

Not structurally. All three store commercial claims as prose in documents, with provenance limited to 'last modified by' and search that ranks keywords over current validity. The failure is the document model, not the vendor.

What replaces the sales wiki?

Not a different wiki — a claim-level layer beneath it. A typed knowledge graph stores each price, proof point, and positioning statement as a versioned, source-attributed node and propagates a change to every dependent surface. The wiki stays; the claims inside it become governed.

How does a claims graph keep AI sales agents on-message?

Agents read the same governed canon over a Model Context Protocol layer instead of improvising from a scraped page. One canonical claim flows to every human and AI surface, so they stop giving different answers to the same question.

Does adopting a Truth Graph mean ripping out Confluence?

No. The graph sits one layer above your existing tools as the substrate they read from. Confluence still hosts long-form docs; the graph governs the claims inside them so they stay accurate.

Why Confluence fails as a sales source of truth

Every Head of GTM eventually sits with the same private question: why isn’t the positioning showing up in deals? You wrote it. Somewhere between the Confluence page and the live call, it quietly stopped being what the company says.

The Commercial Truth manifesto makes a structural claim: marketing has never had infrastructure. Engineering commits code to a versioned system; finance posts every entry to a ledger; the function that decides what a company says about itself runs on documents-with-links. Confluence, Notion, and SharePoint are where that promise gets made — and where it breaks.

This is not a Confluence problem, or a Notion problem, or a SharePoint problem. It is the structural ceiling of treating commercial claims as prose. Name the pattern once and you stop re-buying it every two years.

What the wiki actually delivered

The pitch for the company wiki is centralization: one place, collaboratively edited, searchable, so the revenue team stops chasing attachments. The page count climbs. The accuracy does not.

A product marketer writes a pricing page for a launch and rarely returns to it. The competitor ships the feature your battlecard swears they lack. The case study still opens with a logo that churned last quarter.

Reps notice before anyone tells them. Quoting a stale pricing tier on a live call stalls the deal, so they learn to stop trusting the page. What looked like a knowledge base has become a content graveyard — well-organized, fully searchable, and out of date.

So they go digging. They stalk the top AE in Slack, replay last week’s call recording, and ask the same question in three channels to see whether the answers agree. We have called this information archaeology, and it is where a ramping rep loses the first quarter.

Why a document cannot be a source of truth

The failure is architectural, not behavioral. Telling people to “use the wiki more” cannot fix any of the reasons it decays.

First, a wiki versions pages, not claims. One page holds a dozen distinct facts — a price, a competitor capability, a proof point — and the system treats the whole page as a single block of text. The atomic unit of your message is the claim; the wiki has no concept of it.

Second, nothing propagates. When the pricing page changes, the deck that quoted it does not change, the proposal template does not change, and the AI SDR does not change. Edit history records that an edit happened; it does not carry that edit to the surfaces that depend on it.

Third, provenance answers the wrong question. “Last modified by” is useful for assigning blame and useless for establishing truth — it names who touched the page, not whether the claim is current. A two-year-old assertion and yesterday’s approved one render in identical type.

Fourth, search retrieves text, not validity. A query for a competitor battlecard returns every page that mentions the name, with no signal for which is live, which is deprecated, and which was a draft someone forgot to delete. The rep is left to guess, which is the exact task the wiki was bought to remove.

There is a fifth reason that only started to matter recently. An AI agent cannot read a Confluence page the way a colleague does. Fed an unstructured page it invents a confident variation of your positioning; fed nothing it stays ungrounded — and the inconsistency now scales at machine speed.

The shared drive and the messaging app fail from the other end. SharePoint hands you two competing “final” battlecards and keyword search over filenames. Slack holds the real-time answer for about a day before it sinks under standups and emojis.

What does this cost across a 200-rep org? Not yet a number we will publish as if it were market-wide ground truth. Assay’s positioning canon documents a six-to-eight-week lag before a change lands across every surface; the honest public answer on dollars is a range, not a point estimate.

What a claim-level substrate looks like

The fix is not a better wiki. It is a different primitive underneath the wiki — move the unit of record from the page to the claim, and the failures invert.

In a typed knowledge graph, each claim is a node carrying five things: a source type, a confidence score with an enforced ceiling, a version history, a cascade map of every surface that references it, and an audit entry on every change. A price is a node. A competitor capability is a node.

Now propagation is structural. Change the pricing node and the cascade flags each dependent deck, template, and agent for review — the truth gap between “updated the page” and “updated the deal” closes by construction. Assay’s canon frames the goal as compressing that six-to-eight-week drift tail to under two days.

Provenance becomes a property, not a postscript. A node carries its source type and a confidence badge — verified, stale, or quarantined — so a rep reads reliability instead of guessing it. Search returns the live claim, not every page that ever mentioned it.

The agent problem closes too. When positioning, pricing, and claims live in one governed layer, every AI surface reads the same canon over a Model Context Protocol layer instead of improvising from a scraped page. One change, all your humans and bots, traceably.

None of this retires Confluence. The wiki still hosts the long-form doc; the drive still holds the contract; Slack still hosts the conversation. The graph sits one layer above them as the substrate they read from when they need to know what the company is currently saying — and what was a documentation problem becomes infrastructure.

Closes / opens

The wiki was never going to be the source of truth, and every commercial leader who has heard the phrase for a decade already suspects it. The question is no longer which document to trust. It is whether your claims are typed, sourced, and propagating — or still decaying quietly on a page.

That capacity is measurable. The methodology Assay is developing for the Commercial Truth Index scores exactly this: whether the substance your reps and agents emit is grounded, calibrated, coherent, and auditable. A wiki sits near the floor by construction; a governed graph is what moving up the scale looks like.

Closes the predecessor-comparison cluster (LSO §F.10) as a public-facing essay. Opens the migration question: how a GTM team moves its live canon off documents and into a typed graph without a freeze.

This essay is grounded in Assay’s predecessor-comparison positioning canon (what we’re not) and the Truth Graph spec, with buyer and pillar context from the brand canon. Methodology for the Commercial Truth Index is in development.