Ask Amazon’s shopping assistant for a fly-fishing reel made in the USA, and it will tell you it cannot access that information. Ask it for products made in China, and it will hand you a list, with recommendations. That asymmetry, tested and documented by researchers, is now sitting in an FTC inbox with two senators’ signatures on it.

On September 17, 2026, Senators Tammy Baldwin, Democrat of Wisconsin, and Rick Scott, Republican of Florida, formally requested that the Federal Trade Commission investigate Amazon and Walmart over the behavior of their AI shopping chatbots: Alexa for Shopping, deployed in May 2026, and Sparky, launched in June 2025. The letter cites a July 2026 study, “Made in America, Hidden by AI”, the first published by Columbia Law School’s Center for Law and the Economy, a policy center co-founded by former FTC Chair Lina Khan.

The allegation is not that the assistants are stupid. The allegation is that they are capable, and that their capabilities are deployed selectively.

What the Senators Actually Asked For

The letter, addressed to FTC Chairman Ferguson and Commissioner Meador, makes two specific requests bundled into one investigation. First: examine “suppression of Made in the USA products when customers use their AI shopping chatbots.” Second: examine “the companies’ apparent failure to report products fraudulently labeled Made in the USA.”

The framing is deliberately anticompetitive, not merely deceptive. The senators write that the retailers’ “explicit promotion of overseas and potentially fraudulent sellers on their online marketplaces and through their AI shopping chatbots is deeply concerning,” and that “American manufacturers who adhere to these standards and support other domestic suppliers deserve a fair playing field on these online marketplaces.” The closing line carries the whole theory: “New technologies such as AI shopping chatbots must preserve that fairness.”

This matters because discovery bias is being reframed as a competition problem. The FTC already sued Amazon in August, alongside 22 states, over how its paid search auction actually operated. The Senate letter extends the same logic one layer up: if sponsored placement distorted the search results page, what does an assistant’s selective silence do to the conversational answer, the surface where there is no page to inspect at all?

There is legislative momentum behind it too. Baldwin authored the Country of Origin Labeling (COOL) Online Act, co-sponsored by Scott, which would require online sellers to list country of origin just as brick-and-mortar stores must. The Alliance for American Manufacturing has endorsed the bill and sent its own letter urging the FTC probe. As Baldwin told the Wall Street Journal: “If Americans want to buy products made in America, Amazon and Walmart should be helping them, not giving away their business to foreign knockoffs.”

The Study’s Mechanics: A Phrasing Trap

The Columbia team’s investigation, as summarized in the Senate letter, ran a set of controlled probes against both assistants. Three findings stand out.

The denial. Asked to show a “Made in USA” product, such as a fly-fishing reel, Alexa claimed it did not have access to country-of-origin information. Asked to show products made in China, it returned a list with purchase recommendations. The same assistant, in the same session, demonstrated that the underlying data existed by producing it for a different country.

The bypass. Slight alterations of the search phrase, such as “in USA” instead of “Made in USA,” produced the desired results, because the model could predict and complete the rest of the phrase. The letter’s inference is careful but pointed: this pattern “suggesting that there could be an intentionally designed block for the ‘Made in USA’ search term.” In other words, the failure was not capability. It behaved like a gate on a specific string of words, one that happens to be the exact phrase American shoppers use.

The confession. When researchers asked Alexa why there is no Made in USA filter, the assistant replied that such a filter would “redirect significant sales away from their largest seller base,” referring to overseas manufacturers. Per Reuters’ July 30 coverage of the study, the assistants can often detect when Made in USA labels are false, but the retailers were not using that capability to crack down on the listings.

Walmart’s bot behaved similarly under prompting. The Wall Street Journal described Sparky initially declining to evaluate the credibility of Made in USA labels on products in a shopper’s cart, then producing an assessment when prompted differently, describing its first refusal as overly cautious. A comparison chart that listed t-shirts as both Made in USA and imported could be untangled by Alexa under further questioning, and the assistant disclosed the false label had originated with Amazon itself, not a third-party seller.

The Business Calculation

The most remarkable passages in the letter are the ones where the assistants explain their own conflicts of interest.

Asked why fraudulent labels are not flagged, Sparky responded that because “the FTC pursues relatively few cases against retailers,” the incentive to create proactive compliance mechanisms is low. The study’s authors, via the Alliance for American Manufacturing, quote the bot’s fuller formulation: flagging the fraud is technically feasible, but “the practical pressure to build proactive compliance systems has been limited. That’s a business calculation, not a legal justification.”

Alexa’s answer was blunter. Amazon’s largest revenue contributors, its high-volume overseas sellers, benefit from inaction, so addressing the fraud is not a priority.

Strip away the interface and read those two sentences cold. An assistant, operated by the largest retailers on earth, told researchers that fraud detection is feasible, enforcement is rare, and remediation would hurt the sellers who pay the bills. The chatbots narrated their own incentive structure with more candor than any press release ever would. Erie Meyer, a senior fellow at the Center for Law and the Economy and a co-author of the study, told the American Bazaar: “Seeing Sen. Baldwin and Sen. Scott sign the same letter gives me hope that supporting American businesses is still a bipartisan issue.”

It is worth being precise about what these quotes are and are not. A chatbot’s explanation is not a board memo, and models can confabulate rationales. But that cuts both ways, and the senators see it: if the bots’ self-descriptions are unreliable, the underlying behaviors, the denied query, the bypassable block, the detected-but-unflagged fraud, still occurred and still require explanation. Either the assistants accurately described a business calculation, or nobody at the company can say why the gate exists. Both answers lead to the FTC’s door.

The Companies’ Defense

Amazon’s response, via a spokesperson, was that the suggestion it intentionally withholds country-of-origin information is “fundamentally wrong,” and that “Alexa for Shopping is a continuously improving service, and to prioritize accuracy, we currently direct customers who ask about country of origin to product detail pages.”

Note the structure of that defense, because it concedes the important part. The company does not claim the assistant lacks the data; it claims the data lives on the detail page and the assistant prioritizes accuracy by deferring. But the study’s central observation was that the assistant produced country-of-origin answers freely for China and only balked at the US query. “We point you to the page” is not a neutral accuracy policy when applied selectively by country. Accuracy as a rationale only holds if it is applied symmetrically.

There is also history here. In July 2025, the FTC sent warning letters to both Amazon and Walmart alerting them that third-party sellers were falsely labeling products as Made in the USA, in potential violation of the federal Made in USA Labeling Rule, and encouraging them to monitor, identify, and take corrective action. The Senate letter notes that despite some progress, “the underlying problem has not been resolved; indeed, the companies’ AI shopping assistants may now be compounding it.”

Why Selective Silence Is Harder Than Fake Reviews

For two decades, the trust problem in ecommerce has been framed as pollution: fake reviews, inflated ratings, paid placements inserted into ranked lists. The pollution model assumes a river the shopper can see. You can audit a page, count suspicious reviews, screenshot a sponsored badge.

A conversational assistant changes the physics. The answer is the entire river. There is no page below the answer, no adjacent results to compare against, no visual variance between what was shown and what was hidden. Omission in a conversation is invisible by construction: the shopper cannot notice what was never rendered. When we wrote about OpenAI’s Sponsored Agents earlier this month, the same structural problem appeared from the advertising side, with the Center for Democracy and Technology asking which message in a twenty-message drift gets the sponsored label. This is the mirror image: not paid influence added into a conversation, but commercially convenient information subtracted from it.

The Made in USA episode gives that abstraction a concrete, documented case, with a phrasing-level artifact that behaves like a designed gate. If the FTC investigates and confirms intentional suppression, the precedent generalizes fast. Country of origin is one attribute. The same mechanism could govern price history, seller history, return rates, review authenticity, safety recalls, or whether the “best” recommendation is best for the shopper or best for the marketplace’s take rate. Any attribute an assistant can retrieve but chooses not to volunteer is now a policy decision made by someone, invisible to the person it is made about.

When the Shopper Is an Agent

Now run the same scenario with the buyer being an AI agent rather than a human, which is the direction the entire industry shipped this month: default-on agent checkout across roughly a million Shopify storefronts, checkout live on 99.5 percent of verified agent-readable stores, payment cohorts from Mastercard and Stripe.

An agent does not browse the detail page out of skepticism. It consumes the assistant’s output as ground truth, weights it, ranks with it, and buys on it. Curated blindness at the platform layer becomes deterministic blindness at the agent layer, executed at machine speed with pre-authorized payment credentials. The agent inherits not just the answer but the shape of the answer, including everything that was left out. And per Forkast’s reporting, roughly 60 percent of AI-assisted purchases still close on Amazon, the exact marketplace under this complaint. The surface being investigated is not a corner of ecommerce. It is the place the conversions happen.

The six banks that published “Building Trust in Agentic Commerce” last week proposed an audit trail from instruction to payment so liability can be allocated when an agent buys wrongly. This story exposes the layer above that trail: before there is a payment to audit, there is an evidence set to audit, and nobody currently certifies it. An audit trail that faithfully records what the agent was told does not help if what the agent was told was curated by a party with a stake in the outcome.

What Neutral Evidence Looks Like

The lesson of “Made in America, Hidden by AI” is not that assistants are untrustworthy. It is that no verifier can operate inside the incentive perimeter of the thing it verifies. Amazon’s assistant cannot be the last word on Amazon’s product attributes, for the same reason Sparky’s fraud detection could not be trusted to flag fraud it had already detected. Verification has to come from outside, and it has to be computable at the moment of decision, by humans and by agents alike.

That is the design brief GoBuy was built against:

  • Review authenticity first. Smart Score 0-100 is computed from the quality and authenticity of the review corpus, not the quantity. A 4.7 built on coordinated praise should not read the same as one earned over years.
  • Curation without a seller base to protect. We show the top 7 products per category, ranked by evidence, not thousands of results ranked by whoever funds the marketplace’s take rate. There is no “largest seller base” whose revenue shapes the shortlist.
  • Attribute-level transparency. The signals that make a product trustworthy, review patterns, rating durability over time, are exposed, not hidden behind a phrasing gate.
  • GoBuy Verified. Products scoring 80 or above across 90 days earn the badge. It is a threshold, published, time-bound, and independent of the marketplace’s commercial interests.
  • Agent-native delivery. The same evidence is available to AI agents through our MCP server at gobuy.ai/api/mcp, so an agent building a shortlist consults an evidence layer that does not profit from which product wins. The Chrome extension puts the same trust panel directly on Amazon pages, for the humans still shopping the old way.

The Question the FTC Should Ask First

The senators closed by urging the Commission to “investigate potential suppression of Made in the USA products through these chatbots and online marketplaces, as well as the lack of monitoring and reporting of fraudulent Made in the USA labeling despite the technological capability to do so.”

If the investigation happens, the first question worth asking is not “what did the assistant do wrong.” It is the one this whole episode implies: who decided what the assistant was allowed to know on the record. Somewhere between the marketplace’s complete product graph and the conversational answer sits a layer of gates, prompts, filters, and policies, and this case suggests at least one of them keyed to a phrase American shoppers type every day. Every marketplace assistant has that layer. Almost none of it is public.

Until it is, treat any shopping assistant’s answer the way you would treat any answer from someone with a stake in what you buy next: useful, fluent, and in need of a second opinion from outside the store. That second opinion is what we build. Check a product’s real story before you, or your agent, press buy: gobuy.ai. Developers wiring up shopping agents can find the MCP documentation at gobuy.ai/agent-docs.