Primary Bias
A model's prior association with a brand, held before it retrieves anything. What it already thinks colours what it does with the evidence. Attributed to Dan Petrovic and DEJAN, 2025 to 2026.
What it actually means
Retrieval does not start from nothing. A model carries associations from training, and those priors influence which retrieved material it trusts, how it frames the answer, and whether it names you.
A brand with a strong favourable prior gets the benefit of the doubt on thin evidence. A brand with a weak or wrong prior has to overcome it with retrieval, every time.
Why it is hard to fix
Priors come from training data, which is historical and not editable. You cannot correct what a model already thinks; you can only supply retrievable evidence that competes with it in the moment.
The long game is being in the record accurately enough and for long enough that the next generation of training data carries a better prior. That is slow, and it is the strongest argument for consistency over campaigns.