How to tell which product claims are actually backed by data

A product claim is backed by data when a buyer, seller, marketer, or AI agent can trace it to a source, see how much confidence it deserves, and know where it is allowed to appear. If the answer is “it came from a deck,” “someone said it in Slack,” or “the model inferred it,” the claim may still be useful, but it is not yet verified.

The Commercial Truth manifesto names the underlying problem: go-to-market teams do not lack words; they lack infrastructure for deciding which words are true enough to repeat. Most companies can list their assets. Far fewer can list the actual claims those assets contain, the source behind each one, and the surfaces where those claims have spread.

Start with the claim, not the asset

The usual audit starts by opening the website, the sales deck, the latest one-pager, and the enablement folder. That is useful, but it hides the real unit of risk. A page is not true or false. A claim is.

Pull the sentence-level assertions out of the asset first: what the product does, who it is for, which workflow it replaces, which result it supports, what proof exists, and what a buyer can reasonably expect. Then ask one question of each assertion: if a prospect asked “how do you know,” could the team answer without improvising?

That move changes the work. You are no longer reviewing copy for polish. You are building an inventory of commercial truth: each statement your market-facing systems are about to repeat.

Sort claims by source type

The next step is not a binary true-or-false label. It is a source-type label. A claim drawn from a signed customer quote, a shipped product capability, a pricing policy, and an AI-generated draft should not carry the same weight.

A source-type taxonomy gives each claim a ceiling. Human-approved product facts can sit high. Customer proof can sit high when the scope is clear. Rep language from a call note might be a useful signal, but it should stay low until someone verifies it. AI output is useful for drafting and discovery, not as an authority by itself.

The practical rule is simple: the source decides the confidence ceiling. If the source is thin, the claim stays thin. If the source is missing, the claim says “not yet.”

Look for inherited claims

The most dangerous claims are often the ones nobody remembers writing. They appear first in a launch deck, then in outbound copy, then in a chatbot answer, then in a board slide. By the time someone challenges them, they feel canonical because they are everywhere.

Inherited claims usually have three tells. They use confident language without a source. They appear across many surfaces with small wording changes. They survive product or pricing changes because no one knows who owns them.

Those claims should not be deleted immediately. First, quarantine them. Give each one an owner, ask for the source, and mark it as unverified until the owner can attach evidence. If the evidence appears, the claim can graduate. If it does not, the claim becomes a useful warning: the organization has been repeating a story it cannot support.

Build the backed-by-data test

For every meaningful claim, ask four questions:

• What is the source?
• Who approved the source for external use?
• What confidence should the source allow?
• Which surfaces already repeat the claim?

If those answers exist, the claim is backed by data in the only way that matters operationally: someone can inspect it, govern it, and update it. If any answer is missing, the claim should be treated as provisional.

This is where a truth graph becomes useful. The graph stores the claim as a typed record, attaches the source, records the confidence ceiling, and maps every dependent surface. A change to the source can then become a worklist rather than a scavenger hunt.

What this changes for AI agents

AI agents make the backed-by-data problem harder because they scale repetition. A human may repeat an unsupported claim in one call. An agent can repeat it across many conversations, drafts, and workflows before anyone notices.

The answer is not to ban agents from speaking. It is to make them read from governed claims. Before an AI SDR, chatbot, or content assistant asserts something about the product, it should retrieve the claim, its source, and its confidence. If the graph only has a weak source, the agent should hedge. If the graph has no source, the right answer is “not yet.”

That behavior is boring by design. It is also what keeps a useful assistant from becoming a confident amplifier of old positioning.

Where this leaves you

The point of a claim audit is not to shame the team for having unsupported language. Every growing company has some. The point is to stop treating all claims as equal just because they made it into a document.

Once claims are typed, sourced, confidence-scored, and mapped to the surfaces that repeat them, the team can finally answer a buyer’s simplest question: “how do you know?” That is also the operating logic behind the Commercial Truth Index: a company is healthier when its claims are grounded, calibrated, coherent, and auditable before humans and agents carry them into the market.