How do recommendations change across assistants?

Exploratory study outline · In preparation · October 26, 2026

The same shopping request can be put to different assistants. This study asks how their first recommendations compare, and what changes when the request mentions a habitual brand.

The question

We compare constructed running-shoe requests in a US shopping context. Each request has a neutral version and a version that names a habitual brand.

What we observe

• The first recommended brand and the recommendation’s stated reasons.
• Agreement and disagreement across the assistants under the same conditions.
• Whether the habitual brand remains recommended when it is named in the request.

How the comparison is defined

The protocol sets out the requests, brand rotation, observation order and exclusions before collection. It covers ChatGPT, Gemini, Perplexity and Copilot, and records the interface and access conditions of each.

We report empty, null and negative findings alongside any differences, with the uncertainty of each comparison.

Limits

• A constructed request is not a person, and a recommendation is not a purchase.
• The results will describe the tested requests and assistants, not a representative market. On their own, they will not validate the simulator.

No observations have been collected yet. We will publish the results here, including null and negative findings.