Madison Genthry 
The promise of algorithmic recommendation was, at its most ambitious, the end of search. Feed the system enough data about who you are and what you have chosen before, and it will converge, eventually, on what you actually want — eliminating the friction of discovery and replacing the unreliable expertise of human recommendation with something more scalable and more consistent. A decade and a half into the serious deployment of these systems across consumer retail, the results are instructive.
The Recommendation Engine and Its Ceiling
Recommendation algorithms are genuinely good at a specific thing: predicting what a consumer who has bought X is likely to buy next, given the behaviour of consumers who previously bought X. This is useful. It is not, however, the same as understanding what a consumer needs, and the gap between those two things is where algorithmic curation consistently fails.
The system optimises for signal. What has been purchased, clicked, lingered over, returned. What it cannot capture is the absence of a purchase because nothing in the catalogue was right — the consumer who searched, found nothing adequate, and left. This is a significant failure mode in categories where consumer preference is specific and the range of available products is wide but shallow. The algorithm reads silence as satisfaction. It rarely is.
The consumers who are most valuable to specialist retailers — engaged, knowledgeable, with well-developed preferences — are precisely the consumers whose behaviour is most likely to confuse a recommendation engine trained on mass market data. Their choices are outliers. The system is not built for them.
Human Curation as a Technology Problem
There is a tendency in technology circles to frame human expertise as a transitional phenomenon — something useful while the machines learn, but ultimately destined to be superseded once the data sets are large enough and the models sophisticated enough. Applied to retail curation, this framing has not aged well.
The expertise that makes a specialist retailer valuable is not primarily information retrieval. It is judgment — the accumulated understanding of a category that allows someone to make decisions about what belongs in a range, how products should be described, what a consumer in a given situation actually needs. This judgment is developed through engagement with the category and its consumers over time. It is not a database problem.
Consumers navigating unfamiliar territory in a specialist category find their way to the right product through a combination of structured navigation and implicit trust that the range has been assembled with their kind of requirement in mind. That trust is earned through consistency of expertise, not through the sophistication of the search infrastructure. The consumer who finds what they were looking for at specialist retailers is responding to curation, not computation.
The Paradox of More Data
One of the counterintuitive findings of the algorithmic retail era is that more data has not straightforwardly produced better discovery. The volume of available product information has grown enormously. The ability of the average consumer to navigate it effectively has not kept pace. The result, in many categories, is a discovery experience that feels worse despite being, in raw informational terms, richer.
This is a known problem in information science — the point at which additional data degrades rather than improves decision quality by increasing cognitive load beyond the consumer’s capacity to process it. Specialist retail solves this problem not by providing more information but by providing less — a range that has already been filtered by expertise, reducing the decision space to something navigable.
The AI systems being deployed in retail at scale are, in many cases, attempting to solve this problem computationally — by personalising the reduction of choice rather than curating it. It is an interesting approach with real technical ambition. Whether it produces an experience that consumers find as useful as genuine expert curation is, at this point, genuinely open. The evidence from categories where both approaches coexist is not conclusively in the algorithm’s favour.
What This Suggests About the Next Decade
The retail technology story of the next decade is likely to be more nuanced than either the algorithmic triumphalism of the 2010s or the nostalgic defence of human expertise that emerged in response to it. The more plausible outcome is a differentiation — some categories and some consumer segments served well by sophisticated recommendation systems, others where the depth and specificity of consumer preference continues to reward human curation over computational approximation.
The categories that will remain in the latter group are those where consumer preference is genuinely specific, where the consequences of a mismatch are felt clearly, and where the consumer is knowledgeable enough to notice the difference between a product that was right for them and one that was statistically probable. In those categories, the specialist who has built real expertise is not competing with the algorithm. They are doing something the algorithm has not yet learned to replicate.
