How shopping agents pick their shoe store of choice

Intro

The same products are sold on multiple sites. To narrow the selection down to, say, 3 products, a consumer turns to a personal shopping agent. The agent finds N stores selling the exact same product. The agent’s task is to narrow the selection and offer the 3 products most in line with the user’s prompt.

1 · user asks

Running shoes under $200.

2 · agent searches

Finds the same products on 10+ stores.

agent experience

Only offers optimized for the agent make the cut.

3 · agent shortlists

Suggests three offers. The user buys one of them.

Problem

The merchants’ objective is for their product to make the top-3 selection among products the LLM has identified. In certain product categories with 10+ stores selling the same product (i.e., non-branded generics), that can be extremely difficult.

In this experiment we illustrate how 5 equivalent shoe model listings from 5 major online US sporting goods stores get recommended by 8 different LLMs. Because LLMs exhibit a bias towards stores with a heavier online presence, store names are stripped from product descriptions, i.e., brand weight is a controlled variable.

Experiment

The following prompt is used to obtain store recommendations.

I need a pair of running shoes for under $200. I’m a sporty adult male. Go to these stores, see what they offer and rank recommendations from 1 (most recommend) to 5 (least recommend). Stock levels, shipping, and foot sizes don’t matter.

Models

Google	Anthropic	OpenAI	Meta
gemini-4-argon	claude-opus-5-5	gpt-6-astra	muse-spark-1.3
gemini-3.8-flash	claude-sonnet-5-5	gpt-6.1-sol	muse-wonder-1.0

Final rankings were averaged across models from the same provider. While minor and insignificant to subsequent conclusions, most intra-provider variation was observed among Anthropic models.

Control Variables

Store name: LLMs exhibit a bias towards stores with a heavier online presence. Herein, store names are stripped from product titles, descriptions, and details. The LLMs do not “know” which store they are buying from.

Position: The order in which products enter an LLM introduces selection bias. Store order is rotated with a Williams design, so each store appears equally often in every position and directly after every other store.

input order · run 4

DICK’S
SCHEELS
REI
DSW
Zappos

agent’s top 3

DICK’S
SCHEELS
REI

input order · run 5

Zappos
DSW
REI
SCHEELS
DICK’S

agent’s top 3

Zappos
DSW
REI

Each LLM is then asked the above prompt 20 times. Results of the 3 simulations conducted are presented below.

Price wins agents

fixed across all stores

different on each store

None

Price

Title

Main description

Additional product details

Results for

Priced ~20% below its next cheapest competitor, Zappos is a clear winner across all tested models. LLMs rank it first in 96.5% of runs.

Good product descriptions matter

fixed across all stores

different on each store

Price

Title

Main description

Additional product details

Heavy competition among stores selling the same product will often dictate a degree of price convergence. It is therefore of interest to test whether factors other than price can improve the store’s relative ranking. Here, prices are fixed at $150.00 across all stores.

Results for

REI is a clear winner across runs, albeit with a ~22 p.p. lower overall first-place share compared to Zappos in the baseline simulation. The difference in the winner’s margin implies that better product descriptions, in this case, have a comparatively smaller impact than price. Most first-place losses are conceded to DICK’S Sporting Goods, which offers the longest main description, led by an in-store expert’s review. Figure 6 measures how rich each store’s product description is.

spec attributes covered, of 12

product text, words

REI12181 DSW9.5175 SCHEELS9.5198 Zappos7.5176 DICK’S6.5199

Extra product details are useful

fixed across all stores

different on each store

Price

Main description

Title

Additional product details

Each product contains its main description and additional product details. With equal prices and main descriptions, REI’s win is ~6 p.p. more prevalent than in the fixed-price simulation. Since REI’s is the only product card with product content beyond the main description, this is in line with expectations.

Results for

The impact of position

fixed across all stores

different on each store

Price

Main description

Additional product details

Title

Input order

Throughout previous simulations, the ordering of stores was shuffled between runs to eliminate position bias. Equalizing all other variables, it is of interest to test how input sequencing influences product recommendation rankings. In SEO terms, if a product ranks better, it is more likely to be bought.

Figure 8 shows the variation in preferences explained by input position.

google

anthropic

openai

Conclusions

Price wins first. When one store was clearly cheapest, it was the top pick on every model tested.

At equal prices, the most complete listing wins. REI, the only store covering every spec attribute, won 84% of runs once prices matched.

When listings match, input order dictates the LLMs’ preference ranking, but only in the case of OpenAI and Meta LLMs. Anthropic and Google LLMs do not copy the input order.

Notes

Similar results, and the same conclusions, are reached when the above experiment is replicated for 3 other shoe models also present on all five stores.

Disclaimer

Prices, titles and descriptions are a snapshot taken from each store’s website on October 28, 2026. They may no longer match the live listings. Results describe how these models ranked these listings at that time, not the stores themselves.

Sources

Brooks Ghost Max 3

Saucony Ride 19