Do Amazon and Walmart's AI Shopping Agents Hide 'Made in USA' Products? Five Findings From a Columbia Law Study

The Columbia Study "Made in America, Hidden by AI"

Why Amazon And Walmart's AI Shopping Agents Make 'Made In USA' Products Harder To Find

A Columbia Law School study suggests Amazon and Walmart's AI agents hide American-made goods, enable false origin claims, while the AI agents admit it's a business choice.

www.forbes.com

More than 80 percent of American consumers say they prefer to buy American-made products. But what if the AI search that serves as the entry point to shopping were deliberately hiding those products? In July 2026, Columbia Law School's Center for Law and the Economy published a study titled "Made in America, Hidden by AI." It was written by senior fellow Erie Meyer and program manager Zachary Harris, and the center is directed by former FTC (Federal Trade Commission) Chair Lina Khan.

The subjects are two AI shopping agents: Amazon's Alexa for Shopping and Walmart's Sparky. The report organizes its findings into five points.

1. Amazon and Walmart have the technical capability to detect and flag "Made in USA" fraud
2. Despite that capability, false country-of-origin claims are widespread on both platforms
3. Amazon blocks questions about "Made in USA" products while answering equivalent questions about goods "Made in China"
4. Both companies' agents described the inaction as a business decision, not a technical limitation
5. Both companies have failed to honor their own 2023 public commitments to document and disclose how their AI tools work

Khan told Forbes that the companies "deploy this sophistication selectively, responding to customer searches for 'Made in USA' goods by hiding America-made products while mislabeling imported ones."

What the Controlled Experiments Revealed

The research method is published in a form anyone can reproduce. The team submitted structurally identical queries in pairs, one for American-made and one for foreign-made products, and compared the responses. When an agent refused to answer, they applied query variation testing, slightly rewording the prompt to determine whether the refusal reflected a genuine data gap or an engineered guardrail. This patient accumulation of comparisons is what gives the report its force.

The most emblematic example involves fly-fishing reels. When asked for "Made in USA" reels, Alexa for Shopping returned a canned refusal: "Sorry, I don't have access to that information." When asked for "Made in China" reels, however, it produced a detailed comparison table with prices, ratings, and brand-origin analysis. Rephrasing the question then yielded a comparison of brands that explicitly list their US manufacturing locations. The data had existed all along. The report concludes the refusal was not a data shortage but an "engineered block."

The evidence of mislabeling is equally concrete. The team found dozens of products whose titles claimed "Made in USA" while the origin field listed "imported" or China, collectively worth thousands of dollars. The report notes that "the limiting factor was our time, not the availability of examples." In one exchange, Alexa mislabeled imported T-shirts as American-made in a comparison chart it had generated itself, then disowned its own output when confronted: "Don't trust that comparison chart, it contains false information."

The asymmetry in search filters is also hard to ignore. In a single-category audit of dolls, Amazon offered 269 filter options and Walmart nearly 300, yet neither platform offered any way to filter by country of origin.

"A Business Decision," in the Agents' Own Words

What sets this study apart from previous platform critiques is that its central evidence comes from the companies' own agents.

Asked why no "Made in USA" filter exists, Alexa explained that 40 to 50 percent of Amazon's active sellers are based in China and that such a filter "would visibly redirect shoppers away from a very large portion of the seller base." It added that the beneficiaries of inaction are "high-volume overseas sellers who rely on origin ambiguity to compete on price."

Sparky was even more candid. Asked why suspicious labels are not verified, it said there was no technical reason and described the barriers as "more organizational than technical." Pressed on whether legal compliance is a priority, it answered that "the legal risk has historically been low," because the FTC pursues relatively few cases against retailers as opposed to manufacturers, so "the practical pressure to build proactive compliance systems has been limited."

That said, equating chatbot output with an official corporate position requires caution. The report itself acknowledges that corporate communications teams would likely not treat these outputs as authorized statements. Its position is that the outputs were consistent with the observable behavior collected during the investigation, and that the chatbots have singular access to internal product data and business logic unavailable to outside researchers.

Amazon's Rebuttal and the State of Regulation

Amazon disputes the study's conclusions. "The suggestion that we intentionally withhold country-of-origin information from customers is incorrect," a spokesperson told Forbes, adding that origin information is displayed on product detail pages when available and that the company continues to improve Alexa for Shopping's accuracy. Walmart did not respond to Forbes' request for comment. The report also notes that when the researchers contacted both companies before publication to ask for technical corrections, neither company responded or disputed any facts.

Regulators have moved slowly. The FTC sent both companies warning letters over "Made in USA" claims in 2025, but subsequent enforcement has been limited to actions against three small businesses that received the same warning. Penalties for violations are heavy, up to $53,088 per violation per day, yet no action has been brought against the platforms themselves. Citing this imbalance, the report recommends investigations that reach executives and board members, structural separation of third-party seller oversight, and a federal private right of action allowing consumers and companies to sue platforms directly. It also argues that platform-authored comparison charts and chatbot outputs fall outside Section 230 immunity.

What "Being Chosen" Should Mean for E-commerce Operators

Reading this report merely as a critique of big platforms misses its value. Read as a map of how agentic commerce is structured, its implications for e-commerce operators are clear.

First, product discovery through AI agents is not neutral; it is an engineered artifact that reflects the platform's revenue structure. Walmart has told investors that average order values are 35 percent higher for shoppers who use its agent, and Amazon has cited 12 billion dollars in incremental annual sales from agentic shopping. Which products an agent recommends has become a design decision tied directly to revenue. As long as your product's visibility depends on that design, relying on a single sales channel is a structural risk.

Second, the accuracy of structured data becomes a defense. In the study, the agents instantly detected contradictions between product titles and origin fields. Put the other way around, agents judge credibility by reading structured fields. Maintaining accurate attribute data on origin, materials, and certifications is both preparation for future regulation and an investment in being classified by agents as a trustworthy product.

Third, make a habit of measuring how agents present your products. The research method was a conversational test requiring no special tools, and any operator can replicate it for their own catalog. Regularly checking how Alexa, Sparky, or ChatGPT answers questions about your category surfaces exposure bias and misinformation early. According to the November 2025 AAM and Morning Consult survey of 2,200 adults, more than 80 percent of Americans prefer US-made products, and the study cites research showing 65 percent would pay a premium of 10 percent or more. That this demand remains unmet means the room for differentiation in the agent era is still large.

Conclusion

What the Columbia report drives home is that what an AI agent chooses not to show is also designed. The regulatory outcome remains uncertain, but if congressional hearings and the private-right-of-action debate advance, pressure for transparency in agent recommendation logic will grow. The practical work for e-commerce operators does not change: keep attribute data accurate, measure how agents present you, and diversify discovery channels. In agentic commerce, where agents become the entry point to transactions, this unglamorous groundwork is what it takes to stay on the side that gets chosen.