Illustrative example: making 2M marketplace offers findable with hybrid search

Illustrative example. This is a composite scenario written to show how AACSearch can be used. It does not describe a named customer, and the figures are illustrative, not measured results.

The problem

Hundreds of sellers, each with their own feed format and vocabulary: the same brake pad appears as «колодки передние», «тормозные колодки FT» and a part number. Keyword search missed meaning; pure vector search missed part numbers. Sellers paid for visibility, but sponsored items drowned out organic results, and nobody could measure either layer separately.

The setup

The marketplace consolidated all seller feeds into one AACSearch collection through the YML connector, with seller, brand, OE-number and fitment attributes as filterable fields. Queries run in hybrid mode — keyword retrieval for exact part numbers, vector retrieval for natural-language queries — merged with reciprocal rank fusion and reranked at the top.

Sponsored placement runs as scheduled merchandising campaigns with reserved slots and separate reporting. Organic ranking is untouched by paid placement, and both layers report impressions and clicks independently.

What could change (illustrative figures)

• Search CTR: +27% (illustrative), driven mostly by natural-language queries («подходит на Camry 2018») that hybrid retrieval answers and keyword search couldn't.
• Query reformulations: −38% (illustrative) — shoppers find what they needed on the first query instead of iterating.
• A full refresh of 2M offers in about six hours (illustrative) through the scheduled feed pull, with incremental updates landing within minutes of a seller's feed change.

Why it matters

The two-sided fairness problem turned out to be a reporting problem: once sponsored and organic metrics were separate, disputes about placement moved from opinion to dashboards, and the ranking formula itself became something experiments iterate on rather than a political argument.