AI agents are ranking your claims. Nothing is checking them.

Something changed in how buying decisions get made this year, and the part everyone is building is not the part that will hurt.

What is being built

The trust infrastructure for agentic commerce is arriving quickly, and it is good work. It answers three questions. Was this agent actually authorised by the human it claims to represent? Is the payment accountable and reversible? Does this merchant have credible reviews, a clear return policy, reliable fulfilment?

Authorisation, accountability, reputation. Those are the right three questions and serious companies are answering them properly.

The fourth question

None of it asks whether what the merchant said is true.

Agents rank on structured data quality and stated claims. That means marketing copy — the sentences on a product page, written by the seller, checked by nobody — is now a ranking input and a purchase input. An assistant reads "99.2% accurate across every document type", finds it well-structured and confidently phrased, and passes it on.

AI agents read marketing pages and act on them. The trust layer being built for agentic commerce verifies that the agent was authorised and the payment is accountable — not that the merchant's claims are supported. So a page with evidence behind it and a page with nothing behind it look identical to the machine deciding what to recommend.

Why that is worse than it sounds

The obvious worry is that a machine repeats something false. The real damage is quieter and lands on the wrong party.

A merchant who measured carefully, published the methodology, and wrote a claim narrow enough to defend looks exactly the same to an agent as one who wrote whatever sounded strongest. Both are well-formed sentences in clean markup. Rigour and confidence are indistinguishable to a machine reading text.

So the careful merchant is not merely unrewarded. They are outcompeted — because the other one can claim anything, it costs nothing, and no layer in the stack can tell the difference.

The distinction the whole thing turns on

Evidence present is not evidence sufficient. A claim with a citation attached is treated as substantiated, and a claim without one is treated as unsubstantiated. Both readings skip the question that matters.

That is the first of seven distinctions in a method we publish free, and it is the one that matters most here. A claim with a link beside it is not the same as a claim whose linked source can carry it. A machine cannot currently tell those apart, and neither can most humans skimming a page.

Breadth claimed is not breadth tested. A capability demonstrated in one context is described in language that covers every context, and the gap between the two is invisible in the sentence itself. An agent reads the sentence, not the test behind it.

What a fix actually looks like

An independent, dated record of every claim on a page and the evidence published beside it, machine-readable, with each cited source re-checked over time. It does not certify that a claim is true. It makes 'said it' and 'said it and showed the evidence' two different things.

Concretely: an independent reading of the page, each claim recorded verbatim, each cited source checked and re-checked over time, every reading dated. Published so a human can read it and structured so a machine can. The useful output is not a score. It is the difference between two sentences that currently look identical:

"Merchant claims 99.2% accuracy."

"Merchant claims 99.2% accuracy, published a source alongside it, independently read on a stated date, source still resolving, claim unchanged across fourteen dated readings."

What it must not do

It must not certify that a claim is true. A verifier that promised truth would be making exactly the unfalsifiable claim it exists to flag, and it would deserve the scepticism that followed.

This records what was published and whether its cited support still resolves. It does not determine whether a claim is accurate, adequate, or lawful, and it is not legal advice.

Who this lands on first

Merchants already spending on AI visibility. They have already decided this problem is real and are already paying to be read correctly by machines. Everyone in that market sells them 'get mentioned by AI'. Nobody sells them 'be verifiable when you are', which is the same anxiety and the half that survives scrutiny.

E-commerce platforms and app ecosystems. One integration reaches every merchant on the platform. They are also the party a regulator eventually asks about claims made by sellers using their tooling.

Marketplace operators. They carry exposure for claims made by sellers they do not employ, across listings they do not write — the same structure as the affiliate and franchise problems, at larger scale.

Assistant and shopping-agent builders. They inherit the liability for what their agent repeats. A recommendation built on an unsupported claim is their output, not the merchant's — and the early court decisions on chatbot statements have landed on the deploying company.

We ran it on ourselves

knightbyrd.com returned 2 claims of our own that would face the same question, and we published them rather than quietly editing the page. The method is free to apply, including by people who never buy anything from us, and including by competitors — The KnightByrd Claim-Evidence Method v1.0 at https://kema.knightbyrd.com/veris/standard.

If you want to check your own page

Scan one page free at nexus.knightbyrd.com/veris. No account, no install, nothing on your site — a URL is the whole setup. You will get counts: how many claims are on the page, how many would face a substantiation question, how many cite a source that cannot carry them.

That is the same reading an agent is doing right now, except this one shows its working.