Telling an LLM who made it changes which vendor it recommends 1 month ago
Author here. Wanted to common on the brand salience component since a few friends have mentioned it to me. I ran a follow-up experiment where I replaced the vendors with non-existent companies. With n=~500 runs, the base scenario (depersonalized) simply reduced into positional bias. In this case, the models picked the vendor listed first ~85% of the time. This is entirely precedented based on other work I've read.
Happy to answer any questions or run other variations/scenarios people are curious about.