The best catalog for humans isn’t the best catalog for AI
A Cornell Tech paper models what happens when an AI assistant shortlists from your catalog before the shopper sees anything. Planning as if shoppers see everything cost 2.52% to 11.37% of revenue on average, in the model.
Most retailers plan a catalog for a shopper who browses everything. When an AI assistant sits between you and the shopper, it picks a shortlist first. A new paper from Cornell Tech asks what that does to revenue and to the best catalog.
One thing to be clear about. These are model results, not field data. No real store was measured, and the shoppers in the study are simulated.
What the paper models
Omar El Housni and Marouane Ibn Brahim posted the paper on arXiv on Oct 9, 2026. In their model, a retailer chooses a catalog to expose to an AI. The AI turns it into a smaller customer-facing shortlist. The customer sees only that shortlist and picks one product, or buys nothing.
The retailer controls what the AI can see, but not necessarily what the customer sees. Two numbers describe each setup: how many products the retailer can expose to the AI, and how many the AI shows.
Where the AI data came from
The authors prompted models from OpenAI, Anthropic and Google with real product listings from three categories: Lightning chargers, robot vacuums and pet leashes. Each model answered under three shopper personas, budget-sensitive, value-seeking and premium-seeking, which were written into the prompts. That produced 32,400 ranked recommendation lists.
They fitted a ranking model to the lists and tested it on 5,400 held-out rankings. The product the model scored highest matched the AI’s first choice in 78.8% of cases. They then combined the fitted rankings with synthetic customer preferences to estimate revenue.
What ignoring the shortlist cost
The paper compares a retailer who plans as if the customer sees the whole catalog with one who accounts for the AI. Across seven setups, the first lost 2.52% to 11.37% of revenue on average. The tighter the shortlist, the bigger the loss: 11.37% when the AI showed 3 products, and 2.52% when it showed 7, at the same catalog size.
These are averages over many simulated cases. In the worst cases the paper reports losses up to 57.08%.
The better catalog was smaller
Accounting for the AI generally led to offering fewer products. In one setup, the planner who ignored the AI chose 8.35 products on average, and the AI-aware planner chose 4.20. With a shortlist of 7, the numbers were 8.35 and 6.84.
The paper’s explanation is that the AI-aware retailer gives up extra catalog to remove the AI’s discretion over which products reach the customer. In a store, that would be a different claim. Nothing here says that cutting your catalog will raise sales.
Order mattered more than exact scores
A simpler method that used only the order in which the AI ranks products did far better than ignoring the AI. Its mean revenue loss was at most 0.53% across the setups. The takeaway the authors draw is that what you need to know is how the AI ranks your products, more than its exact scores.
What this means for a brand
Read this as a signal, not a recipe. More products do not mean more visibility once an AI picks the shortlist. What the AI shows depends on how it ranks your products for a given kind of shopper, and that can be observed by asking it.
What to do this week
Five steps.
How Shopthru.ai helps
Shopthru.ai is an AI agents team for consumer brands. It asks ChatGPT, Gemini, Google AI, Perplexity and Copilot what your shoppers ask, finds where AI gets your products, prices or policies wrong, and drafts each fix into an approval inbox. It ships only on your yes, then checks again. Instagram and Amazon are next.