Know Your Merchant

1. The gap, in plain words

Agentic commerce has a buyer side and a seller side. The buyer side is getting real verification: protocols now let a merchant or a payment network confirm that the thing making the purchase is an authorized AI agent, acting for a real person, within a set spending limit. That is useful work and it is happening now.

The seller side has no equivalent. Nothing at the point where an AI recommends "buy it here" or "here's a link" routinely confirms that the merchant behind that link is a real, solvent business, or that its offer is what it claims to be. The Paypers frames this directly as a trust gap at the heart of AI commerce: the industry is verifying agents as trusted buyers, but almost no mechanism verifies merchants as trusted sellers, and the question of who vets the merchant at the point of AI recommendation remains unanswered (The Paypers, 06 Jul 2026).

2. Why it matters now

Five dated, sourced facts, each attributed to where it came from:

• The payment industry is verifying AI agents as trusted buyers, but almost no mechanism verifies merchants as trusted sellers, and who vets the merchant at the point of AI recommendation is unanswered (The Paypers, 06 Jul 2026).
• ChatGPT recommended cloned, fraudulent sites impersonating UK retailer Russell and Bromley and cited them as shopping sources (The Guardian, 07 Jun 2026).
• National Trading Standards UK has warned that poisoned AI models are being seeded with scam websites built to be recommended (National Trading Standards UK).
• A Visa PERC threat report on the agentic commerce threat landscape found a rise of more than 450 percent in dark-web posts about AI agents over six months, and a 25 percent rise in malicious bot-initiated transactions (40 percent in the US), naming fraudulent storefronts engineered to exploit AI shopping agents (Visa PERC, corporate.visa.com).
• Cornell Tech WARP research found that a roughly 13-word edit to a public page can get a fabricated entity cited in 38 to 62 percent of deep-research agent reports (Cornell Tech WARP, arXiv, 22 May 2026, as reported by Search Engine Land).

Visa's own response points at the size of the remaining gap. The Visa Agentic Directory, announced 10 Jun 2026, is a closed, single-network list that verifies a merchant is registered. It does not verify that a specific offer or claim from that merchant is truthful.

3. How to check a merchant before you trust an AI recommendation

This is the method behind the checklist. It is built to be reader-usable: run it yourself before you buy from, or trust, a shop an AI names.

1. Domain age and registrar (WHOIS). Look up how old the domain is and who it is registered through. A domain that is days or weeks old, selling under an established brand name, is a signal to slow down.
2. Official-domain match (trademark and press record). Confirm the domain matches the address the real trademark holder uses in its own press and legal record, not just a similar-looking name.
3. Company registration and insolvency status (public registries). Check a public company registry for an active registration behind the business name, and confirm it is not dissolved or in insolvency proceedings.
4. Brand-versus-domain mismatch. Check whether the logos, product names, and branding shown belong to a domain that has no real connection to that brand.
5. Match against known scam records. Check the merchant name and domain against public scam and fraud warning records, including the kind National Trading Standards UK maintains.
6. Offer plausibility. Ask whether the price is implausibly good, and whether the page is using pressure tactics: countdown timers, "only a few left," or urgency messaging pushing you to decide fast.

Each check produces a dated verdict, with the source and date behind each signal:

Verified: the signal checks out against a named source, as of a stated date. Unverified: the signal could not be confirmed one way or the other from available public records. Flagged: the signal points to a real problem (mismatch, known-scam match, registry red flag, implausible offer).

4. One worked example: the cloned Russell and Bromley sites

This is the case in miniature: an AI tool did the recommending, a real brand's name and look were used to build trust, and nothing in the recommendation step checked whether the site behind the link was the real retailer. That is the gap described in Section 1, with a date and a source attached.

5. What comes next: a receipt an AI agent can check

The checklist in Section 3 is built to be run by a person. The same method is also the basis for a machine-readable check we are building, called kym_check_merchant, so that an AI shopping agent can query a dated receipt on a merchant and its offer before it buys from, or cites, that merchant, rather than guessing from a page it has no way to verify.

The framing is calm, not a race. The goal is to be the tool an AI checks against, not another page it has to guess about. This is not live yet. It is logged here, dated, so the record shows the plan before the tool exists, the same discipline this site applies to every other claim.

6. FAQ

Sources

The Paypers, "Know Your Merchant: the trust gap at the heart of AI commerce," 06 Jul 2026: thepaypers.com | The Guardian, 07 Jun 2026: theguardian.com | National Trading Standards UK: nationaltradingstandards.uk | Visa PERC, "Agentic Commerce: the Threat Landscape": corporate.visa.com | Cornell Tech WARP research, arXiv, 22 May 2026, as reported by Search Engine Land: searchengineland.com | Visa Agentic Directory, announced 10 Jun 2026: corporate.visa.com