Semantic Search for a Multi-Seller Marketplace: Ranking Across Sellers
Relevance is only half the problem
On a single-retailer store, search ranks products. On a marketplace, the same product is often listed by many sellers at different prices and service levels, so marketplace search ranking has to answer two questions at once: which products match the query, and which seller's listing to show for each.
The first job is de-duplication — recognising that twenty listings are the same item and collapsing them into one result with a seller choice behind it. Without that, a search for a popular product returns a page of near-identical rows and buries the variety the shopper actually wants to see.
Seller signals in ranking
Once products are grouped, the winning listing is chosen from signals like price, shipping speed, fulfilment reliability, ratings, and stock. These are structured attributes layered on top of semantic relevance, the same pattern as the reranking stage in our semantic search project — retrieval finds the products, a second pass orders the listings.
Relevance versus business rules
Marketplaces also have commercial goals: promoting fulfilled-by-platform listings, new-seller visibility, category take rate. These can be applied as bounded boosts, but they have to be visible and controlled, because a thumb on the scale that shoppers can feel erodes trust in the whole search.
Fairness and new sellers
Pure performance ranking creates a rich-get-richer loop where established sellers always win and new ones never get a first sale. Some exploration budget — occasionally surfacing a capable new seller — keeps the marketplace healthy without hurting the shopper experience much.
Query understanding still matters
All of the seller logic sits on top of understanding the query. "Cheap wireless earbuds with good bass" still needs semantic retrieval to find the right products before any seller ranking happens. The marketplace layer does not replace search quality, it extends it.
Keeping listings honest
Out-of-stock listings, inactive sellers, and stale prices have to drop out of ranking quickly. A marketplace search that sends a shopper to a listing that cannot be fulfilled fails worse than a single store, because the shopper blames the platform, not the seller.
A worked example
A shopper searches "wireless earbuds with good bass under 80." Retrieval finds the matching products across the catalogue, then grouping collapses the forty listings for one popular model into a single result. For that result, listing selection weighs a fast, reliable seller at 78 against a cheaper seller at 71 with slower shipping and a weaker rating, and leads with whichever the marketplace's tuned balance favours — while still letting the shopper open the full seller list.
Measuring whether the balance is right
Two signals show whether commercial boosting has gone too far: search abandonment on queries where boosted listings dominate, and the conversion gap between the shown listing and the one a shopper picks after opening the seller list. If shoppers routinely override the chosen listing, the ranking is serving the marketplace and not the buyer, and the boost weights need to come down.
Where this stops being right
- A marketplace with one dominant seller per product does not have a ranking-across-sellers problem; standard product ranking is enough.
- Heavy commercial boosting that overrides relevance will show up in conversion and search abandonment before it shows up in take rate.
- Catalogue matching at scale — deciding which listings are truly the same product — is its own hard data problem that has to be solved before ranking can be good.
FAQ
What makes marketplace search different from store search? It has to group the same product across many sellers and then choose which seller's listing to show, using price, shipping, and reliability signals on top of relevance.
How are business goals like promoting certain sellers handled? As bounded, visible boosts on top of relevance ranking. Anything strong enough for shoppers to notice tends to hurt trust and search engagement.
Why give new sellers visibility at all? Pure performance ranking locks new sellers out of a first sale. A small exploration budget keeps the seller base healthy with little cost to the shopper.
ISTRALLEN builds marketplace search that groups listings, weighs seller quality, and keeps commercial boosts controlled — see AI for Retail.