Convergent by Design
A June 2025 study measuring error correlation across major language models found that different LLMs fail in the same ways, and that this correlation is higher among more accurate models. As performance improves, models converge not just in what they get right but in what they get wrong. The researchers' core finding: different LLMs are more correlated with each other than they are with ground truth.
In product discovery, the "preferences" of different AI shopping agents are overlapping outputs of systems trained on substantially similar data, aligned through similar processes, optimizing against similar metrics. A controlled experiment in ACM proceedings found that users of a major LLM produced more homogenous ideas at the group level, with higher overlap between participants, compared to those using non-LLM tools. The Kantar study found the same pattern in commerce specifically: different AI models showed similar preference structures when evaluating the same product sets. Switching agents doesn't diversify the results much.
The Visibility Spiral
Convergent preferences land on a marketplace structure that already concentrates attention brutally. Research in the Oxford Review of Economic Policy documents that shoppers on major e-commerce platforms rarely browse past the first couple of results, and that manipulating which products appear in those positions directly shifts purchasing. Sales velocity drives ranking, ranking drives visibility, visibility drives sales.
Add an agent that pre-filters before the user ever sees a result. Products lacking structured data lose 20 to 40 percent selection probability. Products without robust review histories get deprioritized. At least one major e-commerce platform has blocked AI crawlers entirely, removing hundreds of millions of products from agent-mediated discovery.
What falls through? A regional ceramics studio selling through its own website. A specialty food importer whose inventory isn't listed on a major merchant platform. Their products exist. They might be excellent. But if the structured data isn't comprehensive, the review count is thin, and the merchant platform integration is absent, the agent never considers them. The person who would have found them through a wandering search never gets the chance.
The Recursive Narrowing
There's a longer-arc mechanism worth watching. A peer-reviewed analysis documents how LLM outputs, already favoring common patterns, begin shaping the data that future models train on. Models privilege central tendencies. Those outputs enter the training corpus. The next generation narrows further. Separate research confirms that iterative training on self-generated data degrades model performance over time, a phenomenon called model collapse.
Products outside the dominant optimization pattern get ranked lower today. Over time, they risk becoming less representable in the systems that mediate discovery, as training data increasingly reflects previous agent outputs rather than the full breadth of what's available.
Human browsing was noisy, inefficient, and the mechanism that kept the long tail visible. When agents filter with optimized consistency, the narrowing is quiet, cumulative, and self-reinforcing. The things that disappear from discovery leave no trace. They simply stop being found.