AI-designed search modules

The ranker behind the main search results does a very good job, but it has many jobs to balance at once. A module offers a different lens on the same search, focused on one clear value: a complete kit, a well-kept used item at a lower price. That gives buyers a faster way to a good decision.

When the project started, only about 25% of searches had a module to show. Each one took a PM and a data scientist to define by hand, so the long tail of searches got nothing.

Those uncovered pages also carried fewer ads than our standard. That left real headroom to add useful modules without raising ad load beyond the norm on any page, a constraint I held throughout. The goal was never 100% coverage, but there was a lot of room between 25% and that.

I used the LLM as a designer, offline, and kept it off the live path. It reads a large sample of searches and their top ad candidates and proposes modules a buyer would find useful, each with a title, a pitch and rules for which items belong.

1. 1The LLM's choices become training labels.
2. 2A transformer tags the whole catalog in a batch job.
3. 3A lightweight XGBoost model picks items at search time: fast and cheap, with no LLM in the request.

Complete Value Bundles shows listings that include everything for a task, like a camera with lens and bag. Quality Pre-Owned shows well-kept used items at good prices, and stays off searches where the buyer is already looking for used. Both only fill slots that would otherwise be empty, so every impression is new revenue.

The launched modules prove the method: LLM-designed modules can meet the bar for revenue and for buyers. The next step removes the last manual work. Instead of designing modules in advance, the system will generate each module’s title, pitch and item rules at search time, for that query and its candidates.

That is how coverage reaches the long tail: searches too specific, too niche or too new for any pre-built module. It isn’t live yet; the launched modules set the quality bar it will be measured against, with guardrails for speed, accuracy and unsupported claims.