Ranked lists have a self-reinforcing problem: items near the top get more attention, clicks, and purchases, so they look even more relevant than they are. Eugene Yan's guide to position bias lays out how this distortion affects both the measurement of system performance and the data used to train ranking models.
The proposed remedies fall into two broad families. The first is to introduce randomness deliberately, such as by varying the order of items or randomizing placements during evaluation. The second is to learn from the inherent randomness already present in user behavior, using those natural variations as a source of signal to mitigate the bias.
Since the article is the only source consulted, there is no differing viewpoint to reconcile. Its value lies in the practical framing of a familiar problem: position bias is not just a nuisance to be removed, but something that can be measured and addressed by working with randomness rather than against it.