Why product-origin transparency matters

A generation ago, "Made in X" was information you could see on the box before you paid. Online, most of that context evaporates. On a typical online listing you see a brand name, a photo, star ratings, and a price. The Country of Origin field is present maybe half the time; the seller's actual country of registration is buried in a separate page most shoppers never open.

Consumer-safety regulators are increasingly candid about what happens next. The U.S. Consumer Product Safety Commission issued three unilateral product-safety warnings in 2020 — the ones it publishes when a manufacturer won't cooperate with a recall. In 2024 it issued 64. Of those 64, 42 involved products manufactured in China, and 61 — over 95% — were sold online.¹ This is not a claim about any one country's factories. It is a claim about a specific commerce pattern: goods that arrive with minimum inspection, from sellers with minimum accountability, through platforms that are legally not the seller of record.

The CPSC's own guidance notes that the de minimis loophole — under which parcels below a low customs threshold enter the country with almost no inspection — grew from roughly 150 million shipments a year in 2015 to over one billion a decade later.¹ That is the pipe through which most drop-shipped generic goods arrive. Knowing whether a listing rides that pipe or whether it's stocked in a domestic warehouse from a brand with a return address is not paranoia. It is the same signal your grandmother read off the tag on a toaster.

None of this means a product manufactured overseas is bad, or that a product made domestically is good. It means opacity is the risk factor. A listing where every origin signal is visible and consistent behaves very differently, statistically, from a listing where every origin signal is missing or contradicted. We surface both.

The counterfeit problem online is not a rounding error and it is not shrinking. The OECD and the EU Intellectual Property Office estimated the global trade in counterfeit and pirated goods at USD 464 billion in 2019, rising to USD 467 billion in their 2021 dataset — roughly 2.3% of world imports.² A 2016 report commissioned by the International Trademark Association projected the global counterfeit-and-piracy total could reach USD 2.8 trillion by 2022 once the digital and services share is included.³ Those are macro numbers; the micro version is a listing on your screen right now that describes itself as something it isn't.

Amazon publishes its own account of the same problem. In 2023 the company said it invested USD 1.2 billion and employed more than 15,000 people on brand protection; it identified and destroyed 7 million counterfeit units and blocked more than 700,000 attempts to open bad-actor selling accounts.⁴ Amazon is putting real resources against the problem. But the raw numbers also tell you the base rate: the store defends itself against seven-figure counterfeit attempts every year, and no defence is perfect.

Counterfeiting is only the sharp end of a wider authenticity problem. The subtler cases are misrepresentation: a US-sounding brand name registered as a shell entity, then licensed to unbranded overseas contractors; a "Swiss" or "Italian" heritage story attached to a product with no substantive connection to either country; a Delaware LLC filing a US trademark for goods that are designed, manufactured, warehoused, and monetised entirely elsewhere. These are not counterfeits in the trademark-law sense. They are what happens when a shopper cannot see who is really behind a brand — and they are exactly the kind of pattern our four-indicator score is built to expose. When the retailer country and the "money goes to" country disagree, we say so.

The fourth reason is the simplest and, we suspect, the one most shoppers already feel without articulating.

You have preferences about the world. Perhaps you prefer not to fund a supply chain with a documented pattern of Uyghur forced labour — the Australian Strategic Policy Institute's 2020 report Uyghurs for Sale identified at least 80,000 Uyghurs transferred out of Xinjiang between 2017 and 2019 into factories in the supply chains of 83 well-known global brands, working under conditions the authors argued strongly suggested coercion.¹² Perhaps you prefer to buy from a small domestic manufacturer over a shell brand whose margin flows out of the country. Perhaps you'd rather your children's toys come from a jurisdiction with a functional consumer-safety regulator. Perhaps, on any given day, you have none of those preferences and simply want a cheap dress. All of those choices are legitimate.

What is not legitimate — not on the part of shoppers, but on the part of a marketplace — is a design in which the information required to make the choice is out of reach. There is nothing radical about labels. The organic seal, the fair-trade certification, the dolphin-safe tuna icon, the "Product of…" line on a wine bottle: each of those was, at some point, an obscure activist demand that became a normal part of a shelf. Country-of-origin and beneficial-owner transparency for online goods is the same argument, extended to the store where most consumer purchases now happen. It is consumer sovereignty applied consistently.

We do not tell you what to buy. We show you the four things a typical listing quietly leaves out — where it's made, where it ships from, who the retailer really is, and where the money ultimately goes — with a confidence band and a source list. What you do with that is your business.

The scale of the problem eliminates most of the obvious solutions. A team of researchers hand-maintaining a database of every SKU on Amazon would need to update it faster than Amazon adds new listings — and the answer for a given SKU can change quarterly (brand acquired, factory relocated, new marketplace seller). Regulators are moving on transparency (see the FTC's 2024 review of Made-in-USA claims and the EU's Digital Product Passport rollout), but consumer-facing labels for online goods are, in 2026, still not required at point-of-sale in the way food labels are.

What has changed in the last two years is that a frontier-tier AI model — the same class of system that reads legal briefs, medical literature, and technical filings — can now read a product page in full, weigh it against structured evidence (trademark filings, customs records, brand websites, corporate registries), and produce a country-and-confidence answer for each of the four indicators. That's real work an AI can do well. It's why the tool exists at all.

Which model, specifically. We use Claude (Anthropic's frontier model). Claude reads every product page, ranks the evidence, and writes the score. Two design decisions keep the AI honest:

• Every claim cites a source. Claude is required to name the specific evidence it relied on for each indicator — the specs table entry, the seller's registered address, the customs record. If it can't cite, it can't score. The score card links every citation to the real source so you can check the AI's homework.
• World knowledge is committed before the model sees the paperwork. Before Claude reads any structured evidence about a brand, we ask it to write down what it already knows — "Bosch is German," "TCL is Chinese," "Hydro Flask is American." That protects against paper trails: a Chinese seller registering "BRAND INC" in Delaware would otherwise fool a naive trademark lookup into reporting "United States." When world knowledge and paperwork disagree, world knowledge wins.

Honest limits: AI is one input, not the entire pipeline. Confirmed human corrections override AI scores. Every source we consult (USPTO trademark index, WHOIS registry, customs records, seller storefronts, brand websites, cross-ASIN priors) is a real database or a real page — the AI orchestrates and reasons; the evidence is not invented. When evidence is thin, the confidence percentage tells you so, and the answer can be "Not specified" rather than a guess. Every score is our best evaluative inference from the disclosed evidence, not a factual assertion.