You already know: learning to rank and position bias (the e-commerce and ad cases), two-sided marketplaces (the job case), and from the surge pricing case that shared inventory makes experiments interfere. This case adds search over unique, perishable inventory: every listing is different, every night can be sold only once, availability changes with every booking and external calendar, people plan trips over days or weeks, and hosts can decline or cancel. The design questions: how to make availability exact, what to rank for, how to keep the marketplace fair and healthy, and how to test when treatment and control compete for the same listings.
How to read this lesson. No code, on purpose. Numbers marked assumed are assumptions you state out loud.
1. The prompt, and the questions that change the design
A mid-level answer: "rank listings by predicted click-through with a learning-to-rank model". A staff candidate first asks about dates and availability (fixed or flexible, how calendars are synced), how people search (map or list, over how long), how booking works (instant or host approval, cancellations), and fairness and trust (new listings, discrimination). Those answers produce an availability index as a hard filter with a re-check at booking, ranking with session signals, prediction of host acceptance and reliability, and marketplace rules for exposure and fairness.
Assumptions you state: