Problem
Some search engines, such as Endeca and Algolia, do not compute a relevance score. They rank by a fixed chain of tie-breaking criteria: matched words, attribute importance, proximity, typos, position, exactness. A team that wants clickstream, session, sales or trend signals to influence the order has no score to adjust. The solution should also work with any engine.
Illustrative example: a shop search on Algolia returns the products in its fixed order, while the team wants items that are added to the cart often to rise in the list.
Solution in brief
Leave the engine's relevance order as it is and re-rank outside it. Turn each result's position in the engine's list into a bounded positional score, combine it with clickstream and product-attribute scores, and sort by the combined value.
Questions this case answers
- How to blend behavioral and catalog signals into an existing result order when the underlying retrieval system exposes only a fixed ordering and no tunable score
- How to add a ranking layer that does not depend on one retrieval engine and survives a migration to another
- How to re-rank search results by clickstream and product signals when the search engine has no scoring model?
- How to keep external ranking logic independent of the search engine so that a later migration is cheap?
Tools and methods
- Tie-breaking relevance model — how the engine orders results: each criterion only splits ties left by the previous one.
- Positional score — an exponential score computed from the position in the engine's list, with parameters and set per query category.
- External signals — clickstream and session scores, dynamic product attributes, seasonal and regional trends.
- Ranking layer — a service after the engine that returns the final order.
Google or a one-time email code. No password.