How Topsort Designs Auction Infrastructure for One-of-a-Kind Marketplaces

Most commerce media auction engines were built for retailers selling standardized products.

If a shopper searches for "wireless headphones," there are often hundreds of similar products that can participate in the auction. Category hierarchies are well defined, products are largely interchangeable, and retrieving candidates for an auction is relatively straightforward. But not every marketplace looks like that.

Luxury, vintage, collectibles, and other curated marketplaces are built around millions of unique products. Every listing has its own characteristics, and relevance depends on much more than category alone. Style, creator, material, period, price, and buyer intent all influence which sponsored products make sense for a shopper.

That changes the engineering problem.

The Auction Starts Before Bidding

Most discussions about commerce media focus on the auction itself—how bids are ranked, how winners are selected, or how machine learning improves optimization.

In reality, one of the hardest problems happens before the first bid is ever evaluated. Before an auction can begin, the platform has to determine which products should be eligible to compete.

For marketplaces with standardized catalogs, that process is relatively simple. For marketplaces with millions of one-of-a-kind products like 1stDibs, it becomes one of the most computationally expensive parts of the entire auction pipeline.

As catalogs grow, evaluating every possible product is no longer practical. The system must reduce millions of potential candidates to a small, highly relevant set, all within milliseconds.

Designing an Auction Around the Catalog

This challenge became especially clear during 1stDibs' migration to Topsort.

1stDibs operates a marketplace with more than 2.6 million products, the majority of them one-of-a-kind, and runs sponsored listings across search, browse, category, creator, and product detail pages. As its commerce media program expanded, the company needed greater flexibility in how products were retrieved, ranked, and monetized across those different discovery experiences.

The migration itself involved several thousands of active campaigns and more than 18 months of historical reporting data. Campaigns were mirrored while the existing system remained live, allowing catalog integrations, auction configurations, and performance to be validated before 1stDibs moved directly to 100% of production traffic on Topsort.

The scale and structure of the catalog created a specific auction challenge. Certain categories, particularly Art, generated candidate pools large enough that a conventional retrieval approach would have struggled to maintain real-time performance.

Rather than asking the auction engine to evaluate every potentially eligible product, Topsort redesigned the retrieval stage of the auction pipeline. The platform applies targeted pre-filtering before bidding begins, narrowing the candidate pool based on relevance signals while preserving the products most likely to matter to the shopper. This allows the auction engine to focus its computation where it creates the most value instead of spending resources evaluating products unlikely to compete effectively.

Balancing Speed with Relevance

Filtering products aggressively can improve latency, but it also risks removing products that may actually be the best sponsored recommendation.

For one-of-a-kind marketplaces, traditional category matching is far from enough. Two products may belong to the same category while being completely different in style, craftsmanship, designer, or buyer appeal.

To address this, Topsort combines efficient candidate retrieval with semantic product matching. Instead of relying solely on category relationships, semantic matching evaluates how closely products relate to the item a shopper is viewing. Sponsored listings are selected based on product similarity rather than simple taxonomy, allowing the auction to surface recommendations that better reflect shopper intent.

This approach is particularly valuable on product detail pages, where shoppers are exploring highly specific products and expect equally relevant recommendations.

Working together, targeted pre-filtering and semantic matching allow the platform to optimize for both performance and relevance. The result is reliable sponsored listings coverage while maintaining sub-50ms auction latency, even across the marketplace's busiest discovery experiences.

Commerce Infrastructure Should Adapt to the Marketplace

Every marketplace has its own catalog structure, discovery journey, and definition of relevance.

A grocery retailer, a luxury marketplace, and an automotive marketplace may all operate commerce media programs, but the assumptions behind their auctions are fundamentally different.

The 1stDibs migration illustrates why modernizing commerce media infrastructure is not simply a matter of replacing one auction engine with another. Moving thousands of campaigns and years of historical data without interrupting advertisers was one part of the challenge. Making the underlying auction architecture work for millions of largely unique products was another.

Rather than forcing every marketplace into the same architecture, Topsort designs its auction engine to adapt to the characteristics of the catalog it serves.

As commerce media expands into increasingly specialized marketplaces, the next generation of auction engines won't be defined only by better bidding algorithms. They'll be defined by their ability to understand unique inventory, retrieve the right candidates efficiently, and deliver relevant sponsored experiences.

The best auction is not necessarily the one that can evaluate the most products. It is the one that can identify the right products to evaluate in the first place.