Search as Marketplace Operating System

This reveals that search is really an operating system for marketplace decisions, not a single ranking model. A buyer search can pass through separate models for query understanding, retrieval, ranking, image quality, ad targeting, and value prediction, all feeding one result page. That is why the real product is the orchestration layer that lets teams swap models in and out without breaking measurement or the buyer and seller experience.

• In practice, each model handles one narrow job. One model can detect blurry listing photos, another can repair them, others decide which items are relevant, which sponsored result is worth showing, and which action is most likely to create a completed order, not just a click.
  1 sacra
• This is why build versus buy is less about owning one magic algorithm and more about owning a large coordination problem. Large platforms like Coupang support separate teams for ads and for search and discovery, plus logging, experimentation, and fulfillment links, because the hard part is making all those systems work together in production.
  2 sacra
• The marketplace payoff is that the same stack can power both organic ranking and monetization. Mirakl turned that logic into Mirakl Ads, giving third party sellers paid placement across retailer marketplaces, which works because the underlying search stack already decides what gets shown, to whom, and with what economic tradeoff.
  3 sacra

The next step in this market is more modular search infrastructure, where marketplaces keep the models that express their unique edge and outsource the rest. As ads, recommendations, and search converge into one decision engine, the winners will be the companies that can update many specialized models quickly while still proving that each change lifts revenue and user outcomes.