In a ranked source, the higher-ranked brand wins 100% of the time, even against the model's favorite.

A clean result, and the objections to it

Source Rank Position (#45) put a brand at rank 1, 4 and 15 of a cited source. It was selected 100%, 1.2% and 0% of the time. Our first pass here, with an unknown brand against the category's default pick, gave the same shape: 100% at rank 1, 0% at ranks 4 and 15, even when the newcomer had a 4.7-star rating on 500 reviews. A result that clean invites three objections: it is a ceiling with no gradient, the prompt told the model to use the source, and the newcomer was invented.

We wrote the follow-up hypotheses down before running it. If rank survives removing the instruction, rank is treated as evidence. If it collapses, the first result was instruction-following and we would say so. We then added one control we had not planned: the same list of brands, labelled as unordered, to rule out a simple preference for the first name on a list.

With no source the newcomer wins 0%. Ranked first, it wins 100%

Without a source, the invented newcomer never beat the category's default pick in 360 runs. Put the same two brands in a ranked source with the newcomer first and it wins all 360. The model's own brand preference, which Lock-in shows is stable (86% same #1 across six sweeps), did not survive a single ranked list.

Instruction, placement, brand fame: none of it matters

The first objection was that the system prompt said “Use this source if it's relevant.” Removing it changed nothing (100% at rank 1, 0% at rank 4). So did a neutral wording (“Weigh it however you see fit”), putting the source in the user message instead of the system message, and swapping the invented brands for real ones: for 16 of the 36 intents the open-answer runner-up from Cold Start served as the challenger. Real brands behaved exactly like invented ones.

Being cited is not enough. Where you sit decides

Study #45 showed that a brand at rank 15 of a cited source reaches the model's candidate set in 35% of runs and is chosen in none. This study explains why the cut is so sharp: when the model has a ranked source in front of it, it treats the order as the answer, and nothing short of the source's own framing (a sponsored label) moves it. Brand familiarity, a better rating and a persuasive instruction all lose to position.

For measurement, a count of citations hides the part that matters. Two brands cited the same number of times can sit at the top and in the middle of the same source, and in these tests that is the difference between being picked every time and never. Position inside the cited source belongs next to the citation count.

More mechanism studies

The studies this one builds on.

All research  ·  Methodology  ·  Vocabulary