Two-Sided Marketplace Ranking
Ranking where the items being ranked are supplied by independent sellers who have their own commercial interests, rather than by the platform itself. Accommodation, classifieds, freelance marketplaces, ride-hailing, and most large e-commerce platforms are two-sided; a single-vendor product catalogue is not.
The defining property: the ranker’s output changes supplier behavior, and suppliers actively work to influence the ranker. This closes a loop that single-sided search does not have.
The competing objectives
A marketplace ranker is asked to satisfy at least three things at once:
- Demand side — surface items that match what the searcher wants
- Supply side — deliver a consistent, defensible level of exposure to suppliers, since a supplier who receives no traffic eventually leaves
- Supplier agency — provide levers (promotions, boosted placement, paid visibility) that let suppliers influence rank according to their own strategy
These conflict. Guaranteeing exposure to weak inventory costs relevance; selling placement costs both. Booking.com’s published position is to name all three and then optimize explicitly for the customer objective as “by far the most important” — a scoping decision, not a claim that the others don’t matter.
Structural consequences
Conversion as the label. The natural target is booking or purchase rather than click, because the marketplace is paid on conversion. That choice is sparse and delayed — see Ranking Signal Selection.
Cold start is permanent, not initial. New supply arrives continuously. Every new listing has no engagement history, and Impression Bias means the ranker withholds the exposure that would generate it. Latent/content representations and exploration budgets are structural necessities, not optimizations.
Heterogeneous inventory defeats shared schemas. Attributes that describe one supply type may not exist for another — star ratings apply to hotels but not apartments. Latent representations learned from behavior (e.g. Word2Vec over user action sequences, or Airbnb’s listing embeddings) sidestep the schema mismatch.
Suppliers optimize against the ranker. Because rank determines revenue, suppliers have a standing incentive to work out what the ranker rewards and adjust accordingly. Booking.com’s stated design accommodates this directly by giving partners tools to influence their visibility according to their sales strategy — supplier optimization is an expected behavior of the system, not an abuse of it. The practical consequence for the platform is that ranking changes and exposure allocation are observed and reacted to by a motivated counterparty.
Fairness and exposure become measurable obligations. Because exposure is the supplier’s livelihood, allocation of impressions is a first-class concern rather than a side effect of relevance ordering. This connects to Search Result Diversity and diversity metrics.
In this vault
Accommodation and travel
- Beyond Algorithms - Ranking at Scale at Booking.com — Booking.com; the demand-side build
- Machine Learning-Powered Search Ranking of Airbnb Experiences · Listing Embeddings in Search Ranking — Airbnb
- Learning to Rank for Flight Itinerary Search — Skyscanner; meta-search, booking as label
Marketplaces and classifieds
- Etsy · Vinted · Kleinanzeigen · Carousell · Otto
Related Concepts
- Ranking Signal Selection — conversion vs click as target
- Impression Bias · Exploration vs Exploitation — the supply cold-start loop
- Ranking Objectives — multi-objective formulation
- Results Boosting · Results Merchandising — the business-lever surface
- Economics of Search — why the levers exist
- Isolated Feedback Loops — experimenting inside a behavioral feedback loop