OfferUp has replaced part of its recommendation system with what Swiss AI firm Albatross calls “perception models,” and the marketplace says the swap lifted homepage feed engagement 84% in an initial pilot. The two companies announced a long-term partnership this week to roll the technology out across OfferUp, following Albatross’s earlier deployment with European marketplace Wallapop.
The technology targets a specific weakness in marketplace recommendations: on a platform where every listing is one-of-a-kind and thousands turn over hourly, ranking by popularity or recency rewards sellers who are already visible and buries a first-time seller’s item no matter how well it matches what a buyer wants. Albatross’s models instead interpret in-session user behavior in real time and reshape the homepage feed accordingly. Over the pilot, listing views rose 19% platform-wide, purchase intentions climbed 15%, and OfferUp buyers on the new system discovered 3 million additional unique listings from 131,000 additional sellers who would have stayed buried under the old ranking.
“Marketplaces are fundamentally a discovery problem: understanding what someone wants right now and matching them with the right item among millions of unique listings,” said Dr. Kevin Kahn, cofounder and CEO of Albatross. “LLMs understand words; perception models understand behavior.”
That distinction is the real signal for marketing leaders running personalization anywhere near a large or fast-turnover catalog: the current generation of consumer-facing recommendation upgrades is not defaulting to a generative AI layer, it is defaulting to purpose-built behavioral models, with generative AI reserved for content and copy. Marketers evaluating vendors pitching “AI personalization” should ask specifically which category a tool falls into, since the two solve different problems and get measured differently. Walmart’s Scintilla update, announced with similar aims of surfacing more relevant marketplace inventory, and this publication’s broader look at automation reshaping how campaigns run both point at the same underlying shift: catalogs are increasingly ranked by real-time behavioral inference rather than static rules.
Source: Albatross via GlobeNewswire
