Why does an AI assistant recommend one brand and not another?

An assistant's answer comes from two places at once. Some of it is what the model absorbed during training, which is a slow-moving impression of your brand built from a large amount of the public web. Some of it is what the assistant retrieved a moment ago, which is fast-moving and much more within reach.

What we can observe, and do

When an assistant retrieves, it usually shows or links what it read. Across hundreds of answers, patterns emerge that are ordinary rather than mysterious: particular review sites, forums, comparison articles and retailer pages recur, and the brands treated favourably on those sources tend to be the brands the assistants put forward. We record every source that appeared, how often, and next to which brand, so a claim about influence can be argued with.

What nobody can honestly claim

There is no published ranking function. No vendor discloses weights. Anyone who tells you they have reverse-engineered the algorithm is describing a correlation and calling it a mechanism. We will not do that, and where the evidence supports only a pattern we say pattern.

What tends to be true anyway

Assistants are answering a buyer's question, so they reward material that answers buyers' questions. In practice that means being described concretely and consistently in the places assistants read: clear specifications, real comparisons against named alternatives, and third-party coverage that addresses the things buyers actually worry about in your category rather than the things your marketing prefers to talk about.

Saved answers and sources can help identify a possible weak point. They do not establish the assistant’s internal reasons. Review proposed diagnoses against confirmed business facts and retain uncertainty.