
A Three-Product Shortlist Can Halve the Catalogue Worth Offering
Cornell Tech researchers fitted ranking models to three AI models and, with synthetic shoppers, found that a three-product shortlist roughly halves the catalogue worth offering when a retailer has room for ten. Ignoring the assistant cost 2.52% to 11.37% of revenue at that capacity, depending on shortlist length. Planning around the order in which an assistant ranks the range recovered nearly all of that loss.
Neritus Vale
When an AI assistant shows only three products, a retailer’s range becomes a contest for three places, and every product added is another contender. In AI-Mediation in Retail Markets, posted to arXiv on 9 October, Omar El Housni and Marouane Ibn Brahim of Cornell Tech find that accounting for such contests shrinks the revenue-maximising catalogue on average. The shortlist length is a scenario parameter: the study sets it at three, five or seven and does not observe what real assistants display. A product the assistant never shows earns nothing, and one it shows in place of a better earner costs money. Beyond its own demand estimates, a retailer needs little more than the order in which each assistant ranks its range.
Two vendor pages treat being left out as the danger, and both firms sell a fix for it. A Yotpo guide says a catalogue hidden from AI crawlers “drops out of chat answers for good”. The programme abstract for a session at eCommerce Expo, run by Semantico, which sells an AI-native product-information system, says a product with missing data is “excluded completely”. The paper prices the opposite risk: being shown can cost money when it takes another product’s place on the shortlist.
In the model, the retailer picks what the assistant reads, a catalogue the authors call “machine-facing”, without knowing which shopper or assistant it is serving. The assistant shows its top few products, the shortlist length K. Under a multinomial logit model, the customer then buys one or walks away, with each product’s chance of sale rising with its appeal. Every sale earns the retailer the list price, so the losses reported below are in revenue, not margin.
The authors had three models rank products for a stated shopper: OpenAI’s gpt-5.4-mini, Anthropic’s Claude Haiku 4.5 and Google’s Gemini 3.6 Flash. These are the mini, Haiku and Flash tiers, queried through an API with a fixed persona prompt, not inside a shopping product. Each ranked its top five of ten real listings, drawn from Amazon Reviews 2023 data on Lightning chargers, robot vacuums and pet leashes. Prompts described the shopper as budget-sensitive, value-seeking or premium-seeking, and gave each product a title, price, brand and a few attributes.
A Plackett-Luce model, fitted per assistant, persona and category, treats a ranking as a series of weighted draws. Each product’s chance of first place is proportional to its weight; second place is drawn the same way from the rest. Heavily weighted products usually win, but not always, which suits assistants that vary between runs. On offer sets the fit had never seen, the top-weighted product was the assistant’s first choice 78.8% of the time.
With room for ten products and a shortlist of three, the catalogue worth offering is about half the size the classical method would pick. The classical method plans as if customers saw everything the retailer stocks, and it offers 8.35 products on average. Planning for the assistant cuts the offer to 4.20.
The cut shrinks as the shortlist grows. Table 3 of the paper puts the assistant-aware offer at 5.88 products for a shortlist of five and 6.84 for seven, cuts of 30% and 18% from the classical 8.35.
Ignoring the assistant costs most when the shortlist is short. With room for ten and a shortlist of three, the classical plan lost 11.37% of the best plan’s revenue on average, the paper’s capacity table shows. A short list gives the assistant more chances to cut a good product. With a shortlist of seven, the loss was 2.52%. At every shortlist length, more room to stock made the classical plan lose more.
Often the best plan leaves the assistant nothing to choose. In 44% to 52% of cases, depending on shortlist length, the assistant-aware catalogue fitted inside the shortlist where the classical one did not. The authors describe a retailer “willing to forgo additional assortment capacity in order to remove the AI’s discretion over which products reach the customer.”
The randomness in the assistants’ answers mattered far less than their order.
The authors ranked products by fitted weight, fixed that order and optimised against it. Scored under the full random model, that plan lost at most 0.53% on average in any configuration. Averages hide single cases. The worst, with a shortlist of three, lost 9.02%.
What ignoring the assistant costs depends on how far it agrees with the retailer’s customers. The authors’ synthetic shoppers mix the assistants’ average preferences for their persona with random taste, in a ratio set by an alignment dial. With the dial at zero, ignoring the assistant cost 5.75% to 24.06% of revenue on average.
The strongest objection is that assistants exist to find what shoppers want, so better ones would bring the broad catalogue back. For that to hold, assistants would have to rank as customers choose. The paper does not test better assistants; its dial moves synthetic customers’ tastes toward the models’ average ranking. Near full alignment, the retailer let the assistant filter in 44.4% to 66.7% of cases, so agreement did not make the assistant free. At full alignment, ignoring it still cost 0.59% to 5.74% of revenue, a figure built from one instance for each of the three categories. The paper does not explain that residual; one reading is that an assistant ranks for the shopper, not the till. Either way, the assistant constrains the catalogue’s width.
The study is small, and its limits bite hardest in clothing. Each category holds 22 to 25 products, 72 in all, so the catalogue worth offering means choosing up to ten from about 25. Each headline average pools 633 instances: three categories, seven alignment settings from 0 to 0.99 with 30 runs each, and the full-alignment case once per category. The shoppers are synthetic, and the models were never asked for more than five of ten. Clothing sells on photographs, fit and reviews, none of which the prompts carried. With one retailer and no rivals, the model cannot say whether an extra product displaces a sibling or a competitor.
An apparel retailer could rerun the study on its own range, replacing what the authors had to invent. It would prompt the assistants its customers use with sampled sets of its styles and pair the fitted weights with demand estimated from its own sales. The paper fitted its ranking models on 27,000 training rankings, so the prompting is the costly step. OpenAI’s product feed specification asks for “one row per purchasable item or variant” and links variant rows through group_id and listing_has_variations, so the retailer should test whether an assistant ranks each colourway as a separate contender. The specification does not say how ranking treats variants.
The same specification holds an analogue of the paper’s lever, set by default to the whole range. In OpenAI’s format, each row is eligible for ChatGPT search unless the merchant switches the is_eligible_search flag off, and eligibility “does not guarantee display”. The flag is not the paper’s capacity cap and says nothing about shortlist length. A merchant that never touches it has still drawn its machine-facing catalogue. It can plan as if customers saw everything, the assumption behind the classical plan, or pay for prompts that reveal an assistant’s ranking and draw the catalogue itself.