Engineering Notes 1 - How Topsort Builds Marketplace-Specific Ranking and Prediction Models

By Ricardo Zilleruelo, Principal Software Engineer, Topsort

A grocery marketplace, a fashion marketplace, and a marketplace for one-of-a-kind design do not generate the same shopping behavior. Their catalogs are different. Purchase cycles are different. The signals that indicate relevance and intent are different.

But at the same time, they are not completely unrelated. Across marketplaces, there are broader patterns in how shoppers engage with products, how products behave, and how context shapes a purchase decision.

Topsort’s ranking and prediction architecture is designed to learn from both.

Large shared components can learn general shopper and product behavior across marketplaces, while smaller marketplace-specific components specialize those representations around the data and behavior of each individual marketplace.

That combination gives Topsort a model that can start with broader commerce intelligence, then adapt to how a particular marketplace actually works.

Clicks and Purchases Are Different Signals

One of the most important distinctions in commerce ranking is also one of the simplest: the product most likely to get a click is not always the product most likely to generate a purchase.

Topsort models those outcomes separately.

The system predicts a quality score, an estimate of how likely a product is to earn meaningful engagement in the current context, and a separate conversion score estimating how likely it is to sell if clicked. Those predictions, together with the advertiser’s bid, contribute to the final ranking decision.

This matters because commerce media needs to balance advertiser value, shopper relevance, and the likelihood of a real commerce outcome rather than optimize for attention alone.

Learning From the Full Commerce Context

A good prediction depends on more than past performance.

On the context side, the model can incorporate signals such as the shopper’s search query, the products currently being viewed, recent shopper behavior, placement, and time.

On the product side, it can reason over structured attributes such as price, brand, and category, alongside richer catalog information such as product text and images.

Behavior adds another layer: what has been happening around this product recently, and what has this shopper interacted with before?

These signals are converted into learned representations and combined so the model can evaluate a product in the context of the current auction rather than in isolation.

The competitive set matters too. A product that is a strong recommendation in one auction may be weaker when the shopper’s intent or the products competing against it change.

Learning Globally, Specializing Locally

Marketplace specialization is only half of the architecture.

Some of the system’s largest components are trained across data from many marketplaces, allowing them to learn more general patterns in shopper and product behavior than any individual marketplace could provide on its own.

Those shared components can scale to billions of parameters.

The marketplace-specific prediction heads remain much smaller, roughly thousands to a million parameters—and can specialize that broader intelligence around the marketplace’s own behavior.

The result is a pretrain-then-specialized approach: learn broad commerce patterns at scale, then adapt them to the marketplace where the decision is actually being made.

That distinction matters especially when onboarding a new marketplace.

Instead of starting with an empty model and waiting for months of marketplace-specific history to accumulate, the system can begin from representations already informed by broader commerce behavior and then become increasingly specialized as local data arrives.

Ranking Should Reflect How Commerce Actually Works

The goal is not to build one generic model for every marketplace or one entirely separate system for each one. It is to combine the strengths of both approaches.

Large shared models can learn patterns at a scale that individual marketplaces cannot. Marketplace-specific components can then capture the differences in catalog structure, shopper behavior, conversion dynamics, and commercial context that make each marketplace distinct.

That is what allows Topsort’s ranking system to answer a more useful question:

Given what we know about commerce broadly, and what we are learning about this marketplace specifically, which product should win this auction right now?

FAQ

How does AI improve product ranking in marketplaces?

AI can improve marketplace ranking by combining more than historical performance. Topsort’s models can consider shopper intent, search context, product attributes, behavioral history, conversion likelihood, and the competitive set in the auction to make more relevant ranking decisions.

Why do marketplaces need different ranking models?

Marketplaces can differ significantly in catalog structure, purchase frequency, product similarity, shopper behavior, and conversion patterns. Topsort combines learning across multiple marketplaces with marketplace-specific specialization so the model can benefit from broader commerce intelligence while adapting to each marketplace’s own behavior.

How does marketplace ranking balance bids, relevance, and conversion?

The advertiser’s bid is one input into the ranking decision. Topsort also predicts how likely a product is to earn meaningful engagement and how likely it is to convert if clicked, helping balance advertiser value with relevance and commerce outcomes.

How can AI rank new products with little or no performance history?

Topsort can use product information such as category, brand, price, text, and images alongside representations learned from broader commerce data. This helps the model form an initial understanding of a new product before it has accumulated significant click or purchase history.

What signals are used in marketplace ranking and prediction?

Depending on the use case, ranking can incorporate search queries, shopper behavior, placement, time, product attributes, catalog text and images, recent product activity, predicted engagement, conversion likelihood, and auction context.

Can a marketplace ranking model work with limited historical data?

Yes. Topsort’s architecture can start from shared components trained across broader commerce data and then specialize as marketplace-specific impressions, clicks, purchases, and behavioral signals accumulate. This gives newly onboarded marketplaces a stronger starting point than training every component from scratch.