Two-sided marketplaces with dynamic inventory need classifiers for search retrieval, optimization, and recommendations. Because items are dynamically created by the public and because sellers can be untrustworthy, there is no easy way to manually label and review all items. …

The “billable event” for an ad is on what event the advertiser is billed. Only when this event happens is the advertiser billed. Otherwise, the ad delivery is free. For native ads, the best practice is to use cost-per-click or cost-per-impression billing. They are standards, and it’s easier to control…

A “bid” in a modern ad system like on Facebook, Google, or Amazon is a complex term with many meanings. Unlike in a textbook, in implementation, there is no “auction” that advertisers “bid” to win. “The auction” is an abstract concept implemented by a computer system about what to show…

Marketplace search optimization works by learning why users engaged with listings in the past to decide what best to show in the future to increase engagement. Biases are other reasons why users engaged with and purchased an item other than because it was the best match. …

If you’re a marketplace, when you start delivering paid promotions (ads), optimization becomes much more challenging than when you were only optimizing items for sale in your search and feed. In typical search optimization, you only need to maximize one thing: sales. Small changes result in small changes in results. …

Promoted supports many different types of features to make recommendations. Here, we list the major feature domains, provide some examples, and describe how to send these features to Promoted. Feature Domains Feature domains help engineering design feature retrieval and integration systems to estimate the value of (item, user, context) matching. The most…

Good engineering is to build an MVP system, prove fit, and then upgrade components as necessary. However, in search, discovery, and ads, data processing latency is extremely difficult to upgrade from batch to streaming later systems. Grab, Pinterest, Snapchat, etc.: many top companies get stuck for years trying to upgrade…

“Personalization” in search means that each user gets different search results that are optimized for that user. In traditional search, every user gets the same results. Promoted optimizes not only “per user” but also everything we could possibly know at the time of an insertion: relevancy to the query, time…