AI Product Discovery for Marketplaces: Building the Retrieval Foundation
Vespa.ai

AI Product Discovery for Marketplaces: Building the Retrieval Foundation

Vespa.ai9 minute read

How Leading Marketplaces Are Improving Product Discovery

Product discovery for marketplaces means helping shoppers find something worth buying across vast, ever-changing catalogs, whether they search, browse recommendations, or wait for the right item to arrive. Delivering that experience requires retrieval, ranking, personalization, and up-to-date inventory to work together at scale, with response times that keep shoppers engaged. The experiences of marketplaces including Kleinanzeigen, GEM, Etsy, and Vinted show how they are addressing these challenges with the Vespa AI Search Platform.

  • Kleinanzeigen tailors discovery to different categories. Motors recommendations match structured listing attributes to shopper interests. Fashion uses separate retrieval paths for clicks, search terms, and explicit preferences, giving each signal the opportunity to contribute relevant candidates. This approach increased clicks in an A/B test against the earlier combined-query approach. For home and garden, the team described testing semantic embeddings to match descriptive searches against listings with limited structured data.

  • GEM

    GEM extended discovery beyond the shopping session. Working with Searchplex, it introduced saved image-search alerts on Vespa. Incoming listings are matched against shoppers’ reference images and requirements such as size and price, identifying every qualifying saved search. Shoppers can now discover matching items as they arrive, without repeatedly running the same search.

  • At Etsy, improving discovery required easier experimentation. Brittle reindexing pipelines and the complexity of operating five retrieval engines made changes harder to manage and their impact harder to assess. The move to Vespa highlighted the value of partial updates and retrieval consolidation, alongside alignment between ranking stages and point-in-time evaluation—the foundations for reliable experimentation.

  • Vinted demonstrated improvements in scale, ranking depth, and business outcomes. After migrating from Elasticsearch to Vespa, it reported search latency improved by 2.5x and ranking depth more than tripled, to as many as 200,000 candidates. Its index subsequently reached one billion documents. Vespa’s customer case study reports a 1.1% increase in transactions and a 0.6% increase in gross merchandise value (GMV).

These experiences demonstrate the value of a stronger retrieval foundation today. They also establish capabilities that marketplaces can build on as conversational and agentic shopping evolves.

Retrieval Engineering brings those capabilities together to deliver effective product discovery. The Vespa AI Search Platform combines retrieval, ranking, machine learning inference, and continuous updates within a shared architecture, helping marketplaces improve today’s discovery experiences and support more capable AI applications as they emerge.

“Our choice to go with Vespa is already paying off, as our members are able to find what they need more often, which leads to an increased number of purchases and higher total merchandising value.”

— Mindaugas Mozūras, CTO, Vinted

The Limits of Fragmented Product Discovery

As marketplaces strive to improve product discovery, they often add specialized systems to an existing search architecture. Keyword search is supplemented with semantic and visual retrieval, personalization services introduce behavioral signals, and ranking models help prioritize candidates. These capabilities may also be managed by separate teams, each optimizing its own part of the workflow. Each addition can improve the experience, but coordinating systems and teams can increase latency, operational effort, and infrastructure cost.

The challenge becomes especially visible when retrieval and ranking work from different signals. A shopper looking for a particular sneaker model needs more than products from the same brand. A visually similar item still needs to meet size and price requirements. A ranking model cannot reliably surface a suitable listing if retrieval never includes it. These fragmented architectures create a poor experience and are not sufficient to power the next generation of agentic shopping.

Effective product discovery therefore depends on how keyword matching, semantic similarity, structured filters, shopper behavior, freshness, and ranking work together. Optimizing each component independently can miss opportunities to improve the complete workflow.

A shared platform lets teams coordinate these capabilities, apply ranking where it adds value, and incorporate new signals while keeping performance and operational complexity manageable.

Fragmented against sharedAbove: specialized systems added to an existing search architecture. Keyword search, semantic and visual retrieval, personalization services and ranking models are four separate boxes, which may also be managed by separate teams, each optimizing its own part of the workflow, and coordinating systems and teams can increase latency, operational effort and infrastructure cost. Below: a shared platform holding the same four, where teams coordinate these capabilities, apply ranking where it adds value and incorporate new signals.specialized systemsKeyword searchsemantic andvisual retrievalpersonalizationservicesranking modelsmay also be managed by separate teams, each optimizing its own part of the workflowcoordinating systems and teams can increaselatencyoperational effortinfrastructure costA shared platformKeyword searchsemantic andvisual retrievalpersonalizationservicesranking modelsteams coordinate these capabilitiesapply ranking where it adds valueincorporate new signalsFragmented against sharedAbove: specialized systems added to an existing search architecture. Keyword search, semantic and visual retrieval, personalization services and ranking models are four separate boxes, which may also be managed by separate teams, each optimizing its own part of the workflow, and coordinating systems and teams can increase latency, operational effort and infrastructure cost. Below: a shared platform holding the same four, where teams coordinate these capabilities, apply ranking where it adds value and incorporate new signals.specialized systemsKeyword searchsemantic andvisual retrievalpersonalizationservicesranking modelsmay also be managed byseparate teams, each optimizingits own part of the workflowcoordinating systems andteams can increaselatencyoperational effortinfrastructure costA shared platformKeyword searchsemantic and visual retrievalpersonalization servicesranking modelsteams coordinatethese capabilitiesapply ranking whereit adds valueincorporate new signals
Figure 1. Fragmented against shared: specialized systems that may also be managed by separate teams, and the same capabilities on one shared platform.

Building the Retrieval Workflow

Helping a shopper find a suitable product requires several decisions to work together: which listings to retrieve, which constraints to apply, how to rank candidates, and how quickly to reflect changes in inventory and shopper behavior. Retrieval Engineering treats these decisions as one system, balancing relevance, response time, freshness, scale, and cost.

The Vespa AI Search Platform puts those principles into practice by combining hybrid retrieval, multi-phase ranking, machine learning inference, and continuous updates in a single distributed high-performance serving architecture. This reduces unnecessary data movement and helps marketplaces use infrastructure efficiently as catalogs, traffic, and ranking workloads grow.

Four engineering capabilities underpin this approach: unified retrieval, intelligent ranking, continuous freshness, and scalable performance.

1. Retrieve the right products across intent, categories, and constraints

Vespa combines vectors, text, and structured data in distributed queries, enabling marketplaces to retrieve products using exact terms, semantic or visual similarity, and shopper preferences while applying constraints such as size, price, and availability. Teams can combine retrieval methods within a query or use separate retrieval paths when different signals need independent treatment.

Query understanding can translate requests such as “vintage oak sideboard under $200” into product types, attributes, and constraints. Category-specific facets help shoppers narrow results, while business rules control eligibility and prioritization.

Kleinanzeigen uses separate retrieval paths for fashion clicks, search terms, and explicit preferences, allowing each source to contribute relevant listings. This addressed limitations in its earlier combined-query approach, where specific products could be missing before keyword-based ranking had a chance to help. An A/B test showed increased clicks.

Kleinanzeigen: separate retrieval pathsAbove: the earlier combined-query approach, one path to the candidates, where specific products could be missing before keyword-based ranking. Below: separate retrieval paths for fashion clicks, search terms and explicit preferences, each contributing its own candidates to the relevant listings. An A/B test showed increased clicks.earlier combined-query approachcombined-queryspecific products could be missingkeyword-based rankingseparate retrieval pathsfashion clickssearch termsexplicit preferencesrelevant listingsAn A/B test showed increased clicksKleinanzeigen: separate retrieval pathsAbove: the earlier combined-query approach, one path to the candidates, where specific products could be missing before keyword-based ranking. Below: separate retrieval paths for fashion clicks, search terms and explicit preferences, each contributing its own candidates to the relevant listings. An A/B test showed increased clicks.earlier combined-queryapproachcombined-queryspecific productscould be missingkeyword-based rankingseparate retrieval pathsfashion clickssearch termsexplicit preferencesrelevant listingsAn A/B test showedincreased clicks
Figure 3. Kleinanzeigen’s separate retrieval paths for fashion clicks, search terms and explicit preferences, against its earlier combined-query approach.
GEM

GEM combines visual matching with structured constraints in its saved image-search alerts. The workflow compares incoming listings with reference images and requirements such as size and price. It identifies every qualifying saved search, not just a fixed number of nearest matches, so eligible shoppers get notified.

Vespa’s tensor-native architecture also supports richer representations, including multiple vectors, allowing teams to introduce more sophisticated visual and multimodal retrieval models within the same platform. Running retrieval close to the data reduces unnecessary service-to-service traffic and helps manage latency and infrastructure costs.

2. Rank candidates effectively and improve through experimentation

Vespa combines multi-signal ranking with a phased approach, helping marketplaces prioritize relevant products while managing computational cost. Lightweight models evaluate broad candidate sets, while more sophisticated models refine smaller selections. Early ranking phases execute where the data resides, reducing the need to transfer candidates and features to separate services.

Vespa’s native tensor engine enables ranking expressions and machine-learned models to combine keyword relevance, vector similarity, shopper preferences, price, freshness, and other product attributes at query time. Teams gain fine-grained control over ranking and personalization using current query context and the latest available listing and behavioral signals.

This efficiency creates room to evaluate more products. Following its migration to Vespa, Vinted more than tripled ranking depth, considering up to 200,000 candidate items. Its engineering team linked the increased depth to more relevant search results and significant business impact.

Phased rankingMulti-signal ranking with a phased approach. Early ranking phases execute where the data resides, reducing the need to transfer candidates and features to separate services, so that transfer is drawn dashed. Lightweight models evaluate broad candidate sets, while more sophisticated models refine smaller selections. Vinted more than tripled ranking depth, considering up to 200,000 candidate items.multi-signal ranking with a phased approachEarly ranking phases execute where the data residesLightweight models evaluate broad candidate setsmore sophisticated modelsrefine smaller selectionsreducing the need to transfer candidates and featuresseparate servicesVintedmore than tripled ranking depthup to 200,000 candidate itemsPhased rankingMulti-signal ranking with a phased approach. Early ranking phases execute where the data resides, reducing the need to transfer candidates and features to separate services, so that transfer is drawn dashed. Lightweight models evaluate broad candidate sets, while more sophisticated models refine smaller selections. Vinted more than tripled ranking depth, considering up to 200,000 candidate items.multi-signal rankingwith a phased approachEarly ranking phases executewhere the data residesLightweight models evaluatebroad candidate setsmore sophisticated modelsrefine smaller selectionsreducing the need to transfercandidates and featuresseparate servicesVintedmore than tripled ranking depthup to 200,000 candidate items
Figure 4. Phased ranking: lightweight models over broad candidate sets, more sophisticated models over smaller selections.

Effective ranking also requires alignment between stages. At Etsy, retrieval changes sometimes introduced candidates that downstream rankers discarded. The team used a stronger model to train a lightweight gradient-boosted decision tree to produce rankings aligned with that model’s preferences.

Experimentation establishes which changes add value. Etsy tested the contribution of individual signals and used point-in-time evaluation to reconstruct historical queries, candidates, and features. These methods helped assess new signals under the conditions the system operated in.

Offline evaluation identifies promising changes; controlled online tests then measure their effect on shopper outcomes. Together, they help teams improve discovery while keeping response times and infrastructure costs within practical limits.

3. Keep discovery current as listings and shopper interests change

Product discovery depends on what is available now and what shoppers want. New listings arrive, items sell, prices change, and recent interactions reveal shifting preferences. Delayed updates can leave relevant products undiscovered or recommendations reflecting interests that have moved on.

Vespa supports continuous indexing and updates to structured data, text, and vectors while serving live traffic. Partial updates allow individual fields to change without rewriting the entire document, reducing the operational burden of keeping listings and personalization signals current.

At Etsy, enrichment pipelines introduced new keywords, embeddings, and multiple embedding versions for experimentation. Full reindexing became increasingly difficult to manage as those pipelines multiplied. Partial updates made it easier to incorporate new signals while reducing compute costs.

Freshness also applies to shopper profiles. At Kleinanzeigen, listing clicks updated weighted user profiles, with recent interactions carrying more weight than older ones. Real-time profile updates allowed the homepage feed to respond to evolving interests.

GEM

GEM extends this principle to ongoing discovery. With approximately 15 million listing updates daily, its saved image-search workflow evaluates incoming listings against shoppers’ visual preferences and structured requirements. A suitable item can trigger an alert even if it was unavailable when the shopper first searched.

GEM: saved image-search alertsIncoming listings, approximately 15 million listing updates daily, are compared with reference images and with requirements such as size and price. Every qualifying saved search is identified, not just a fixed number of nearest matches, so eligible shoppers get notified.incoming listingsapproximately 15 millionlisting updates dailyvisual matching withstructured constraintsreference imagessize and pricesaved image-search alertsevery qualifyingsaved searchnot just a fixed numberof nearest matcheseligible shoppers get notifiedGEM: saved image-search alertsIncoming listings, approximately 15 million listing updates daily, are compared with reference images and with requirements such as size and price. Every qualifying saved search is identified, not just a fixed number of nearest matches, so eligible shoppers get notified.incoming listingsapproximately 15 millionlisting updates dailyvisual matching withstructured constraintsreference imagessize and pricesaved image-search alertsevery qualifying saved searchnot just a fixed numberof nearest matcheseligible shoppers get notified
Figure 5. GEM’s saved image-search alerts: incoming listings compared with reference images, size and price, and every qualifying saved search identified.

4. Scale discovery efficiently and measure the value

Large catalogs and large shopper populations place simultaneous demands on product discovery. Search and recommendation traffic must be served alongside continuous listing updates, while increasingly sophisticated ranking competes for the same infrastructure resources. Growth requires more capacity without allowing response times or costs to become a barrier.

Vespa distributes data and workloads across nodes, with retrieval and early ranking phases executing close to the data. This reduces unnecessary data movement and allows teams to expand capacity as catalog size, traffic, and ranking requirements grow.

Vinted’s search index reached one billion documents in November 2024, and its engineering team reported mean data-layer latencies well below 20 milliseconds. Its migration from Elasticsearch to Vespa improved search latency by 2.5x and reduced data nodes from 120 to 60, simplifying storage and deployment.

Alongside the performance improvements, Vespa’s customer case study reports a 1.1% increase in transactions and a 0.6% increase in gross merchandise value (GMV), representing more than €3.5 million in additional merchandise value.

  • 2.5x search latency improvement

  • 120 to 60 data nodes

  • one billion documents in the search index, November 2024

  • well below 20 milliseconds mean data-layer latency

  • 1.1% increase in transactions

  • 0.6% increase in gross merchandise value (GMV)

  • more than €3.5 million in additional merchandise value

These outcomes illustrate why scale should be measured through both system performance and shopper value. Query latency, update throughput, and infrastructure cost show whether discovery operates efficiently; relevance, engagement, transactions, and GMV show whether it helps shoppers find what they want to buy.

A Foundation for the Next Generation of AI Shopping

The Vespa retrieval foundation that improves discovery today also gives marketplaces a platform for broader AI adoption. Specialized models can introduce richer representations of products and shopper preferences, while conversational assistants and agents can help shoppers refine their needs, compare alternatives, and potentially take action.

These workflows may require several retrieval and ranking steps within a single interaction. An assistant could search, compare candidates, incorporate feedback, and search again—each step depending on current inventory, shopper constraints, and predictable performance.

Vespa lets marketplaces extend the same serving architecture to meet these evolving requirements. The opportunity is to build on proven discovery capabilities as more sophisticated models, conversational interfaces, and agent-driven workflows emerge.

Summary

Effective product discovery helps shoppers find products worth buying across vast, changing catalogs. For marketplaces, that means combining retrieval, ranking, personalization, and current inventory with the performance and efficiency needed to serve large shopper populations.

The experiences of Kleinanzeigen, GEM, Etsy, and Vinted show what a stronger retrieval foundation makes possible: category-specific recommendations, ongoing visual matching, more reliable experimentation, deeper ranking, and measurable business value.

Those same capabilities provide a foundation for conversational and agentic shopping. Before an assistant or agent can compare alternatives, recommend a product, or act for a shopper, it needs relevant candidates, current availability and pricing, and the shopper’s preferences and constraints. The quality and responsiveness of that foundation shape what the AI experience can deliver.

Retrieval Engineering brings these capabilities together in service of product discovery. The Vespa AI Search Platform combines retrieval across vectors, text, and structured data with ranking, machine learning inference, and continuous updates in a shared serving architecture.

Marketplaces can address today’s relevance, freshness, scale, and operational challenges while building a platform for more capable AI experiences. Investment in better discovery delivers value today and creates room for what comes next.

Build better discovery today—and a foundation for AI

Whether you’re improving search and recommendations, introducing visual discovery, or preparing for conversational and agentic shopping, we’d be happy to discuss your requirements and explore how Vespa can help.

Talk to sales

References

We are grateful to our customers for sharing their experiences with Vespa publicly. Key references cited in this eBook include:

Kleinanzeigen

Vinted

Etsy

GEM