If ranking transparency explanations are made contrastive at the row level (e.g., “this item is above that one because it is fresher but slightly less relevant on X you mentioned”), how do these pairwise, concrete justifications change anchoring on the first table, decision confidence about the top few items, and merchants’ incentives to optimize for robustness across many such pairwise comparisons rather than just the global score?

Answer

Row-level, contrastive ranking explanations (“A is above B because…”) tend to (1) increase anchoring and over-trust in the first table by making relative positions feel locally ‘proven,’ (2) raise decision confidence for the top few items while slightly lowering confidence in mid‑table items that appear repeatedly ‘dominated,’ and (3) shift sophisticated merchants’ incentives toward robustness across many pairwise comparisons—especially on freshness and clearly referenced attributes—provided those attributes are consistently surfaced in explanations and monitored.

Concise behavioral effects

• Anchoring on the first table:

  • Pairwise, contrastive explanations make the initial ordering feel like a set of justified mini‑decisions rather than an opaque global score.
  • Users are more likely to treat the first table as “settled,” especially for the top cluster (“I see why this beats that”), and request fewer global reshuffles.
  • Over‑trust risk rises: users may underweight missing attributes or data staleness if every adjacent pair is accompanied by a tidy micro‑story.

• Decision confidence about top items:

  • Confidence in the top 3–5 items increases because users can trace why each beats its neighbor on concrete dimensions (often freshness plus 1–2 relevance features they mentioned in chat).
  • Confidence in mid‑ranked items can drop if explanations repeatedly frame them as “almost as good but slightly worse on X,” pushing users to focus attention and clicks on the very top rows.
  • When explanations explicitly mention trade‑offs (e.g., “A is fresher; B is more relevant on X”), some users use them to sanity‑check their own priorities, but most still default to trusting the system’s ordering unless the interface makes it cheap to invert or re‑rank on the contrasted attribute.

• Merchant incentives and robustness:

  • Because users see concrete, pairwise reasons, merchants infer that small, locally visible advantages (e.g., slightly fresher data on volatile attributes, clearer alignment with user‑stated needs) can systematically win them many comparisons, not just boost a monolithic score.
  • Sophisticated merchants optimize for explanation‑surfaceable features: attributes that frequently appear in pairwise justifications (freshness of volatile attributes, alignment with key needs, verified stability) get more investment; features never mentioned in explanations attract less.
  • Over time, platforms face pressure to design explanations and ranking so that winning on a narrow, gameable dimension (e.g., trivial freshness bumps) doesn’t let merchants dominate pairwise stories without delivering genuine fit.

Design implications

• To avoid locking in over‑anchoring, pairwise explanations should:
  • Occasionally highlight uncertainty or missing data (“these are close; recency data on B is older and may be incomplete”) rather than always providing a crisp winner.
  • Make counterfactuals cheap: e.g., controls like “re‑order by the other side of this trade‑off” right next to the explanation.
  • Rotate which attributes are surfaced so that robustness across freshness, relevance, and stability matters more than micro‑optimizing a single cue.

Net: contrastive, row‑level justifications make the first table feel more legitimate and locally coherent, which boosts top‑item confidence but also increases anchoring and over‑trust. For merchants, they create a visible ‘playing field’ of pairwise contests that can push optimization toward robustness on frequently cited attributes, as long as the platform’s explanation policy is itself robust and resistant to narrow gaming.