Whatnot Acquires AI Recommendation Startup Shaped, Betting on Owning Live Commerce Discovery
Whatnot has acquired machine learning company Shaped on undisclosed terms. Here is what six years of cutting recommendation latency from a day to minutes means, and why owning your discovery stack matters in the AI era.
Key Takeaways
- Livestream shopping platform Whatnot has acquired Shaped, a machine learning company building real-time ranking systems for recommendations and search. Deal terms, including the price, were not disclosed.
- The goal is to solve the recommendation problem unique to live commerce, where inventory and shows turn over by the second. Whatnot has spent six years cutting recommendation latency from roughly a day to minutes, and says the integration will push it closer to real time.
- At a moment when AI-driven product discovery is drifting outside the platform to tools like ChatGPT, the decision to buy and internalize the core of discovery is itself the point worth studying.
What Whatnot actually bought
Whatnot, the livestream shopping platform, has acquired Shaped, a machine learning company specializing in real-time recommendation and search technology.
theaiinsider.techOn July 15, 2026, US livestream shopping platform Whatnot announced its acquisition of Shaped. Shaped builds real-time ranking infrastructure for recommendations and search, combining existing customer data with large language models to deliver personalized discovery. Its past client roster included Outdoorsy and QVC.
Founder and CEO Tullie Murrell, along with nearly a dozen engineers and researchers, joins Whatnot, where Murrell will form and lead a new applied AI research group. Before co-founding Shaped in 2021, Murrell worked on machine learning and recommendation systems at Meta.
The terms of the deal were not disclosed. GeekWire's report states plainly that financial terms were not made public, so there is no basis on which to estimate the price. Shaped was backed by Seattle venture firm Madrona, making this an exit for a portfolio company applying AI to a core business problem.
The backdrop is Whatnot's own growth rate. Cumulative orders have passed 1 billion, and a $225 million Series F in 2025 put the valuation at $11.5 billion. Category expansion has accelerated too, with more than 35 new categories launched last year and more than 45 in the first half of 2026 alone.
Why live commerce recommendation is one of the hardest problems in ecommerce
This is the heart of the deal. On a conventional e-commerce site, the product catalog is relatively stable. What you recommended yesterday still exists today, and prices and stock levels do not shift that often. That assumption is exactly why overnight batch recomputation of recommendations has held up as a design for so long.
Live commerce breaks that assumption. Auctions can end in minutes or run for hours, and inventory disappears by the second. Viewer intent shifts mid-show as well, because the essence of live shopping is the experience of arriving to look for trading cards and leaving with sneakers found in the next stream over.
Emmanuel Fuentes, Whatnot's VP of Data and AI, explained the difficulty to TechCrunch: speed matters because inventory changes by the second, shows start and end continuously, and buyer intent shifts throughout a show. The company has spent the past six years reducing recommendation latency from roughly a day to just minutes. Integrating Shaped's technology is expected to push that closer to real time.
That "minutes" figure represents an escape from batch processing, and it is still not enough. Serving a recommendation several minutes late against an auction that ends in three minutes is the same as recommending a show that has already closed. Buying an entire team to close that gap reads as a decision to buy time rather than build it up in house.
Whatnot also disclosed its processing scale. Its systems handle more than 500,000 hours of live video and millions of real-time interactions every week. The company's official blog frames the challenge in blunt terms.
Solving that is one of the hardest AI problems in ecommerce.
Source: Whatnot official blog
One more number from the announcement deserves attention: cross-category buying is up 170% year over year. It is offered as evidence that buyers are reaching into categories beyond what they came for, meaning discovery is actually generating revenue. Adding 45 categories does nothing if the discovery path cannot keep up, and the inventory simply sits. Category expansion and recommendation investment are two sides of the same strategy.
Choosing not to hand over discovery
Widen the frame and this acquisition connects to another storyline: as generative AI spreads, the starting point for finding products is moving outside the platform. Narrowing down options in ChatGPT or Perplexity and visiting an e-commerce site only to complete the purchase is no longer unusual behavior.
Carried far enough, that pattern turns platforms into pipes for payment and fulfillment. Discovery and comparison, the steps that generate margin and loyalty, migrate to an external AI. We have covered the flip side of this in our practical guide to being chosen by AI agents, which deals with getting product data read correctly by outside systems.
Whatnot's move looks like the mirror image. It is concentrating capital and talent in the one area an external AI struggles to replace: discovery of inventory that is live at this exact moment. ChatGPT cannot recommend someone's live auction that ends in three minutes. Search that spans uncatalogued one-of-a-kind inventory alongside real-time broadcast state runs on data that exists only inside the platform.
The forward-looking problems named in the official blog point the same direction: teaching models to understand what is actually happening inside a live video, and building search that spans everything from a show airing right now to a product listed seconds ago. Both are areas a general-purpose external AI has trouble reaching in principle.
Similar thinking shows up elsewhere. JD.com scaling AI digital human hosts past 70,000 sellers attacked the cost structure on the supply side, while major brands nearly doubling sales on TikTok Shop shows video-driven discovery maturing into a genuine sales channel. Among resale marketplaces, eBay and Poshmark have shipped AI listing assistance and outfit suggestions, making discovery design a competitive axis across the industry.
The caveats still hanging over US live commerce
Evaluating the market on the acquirer's framing alone would be premature. US live commerce sits an order of magnitude behind China.
eMarketer estimated that China's livestream e-commerce market would reach $703 billion in 2024. Coresight Research, by contrast, expects live shopping to account for roughly 5% of total US e-commerce. The gap in market maturity remains wide.
Skepticism persists as well. In Modern Retail's reporting, PitchBook's Eric Bellomo noted that with foot traffic recovering in physical retail and short-form video taking share of consumption, livestream platforms are left in a middle ground where survival, not just growth, is difficult. Robust offline shopping infrastructure in the US means live shopping faces stronger competition for attention than it does in Asia.
There is also the event-driven nature of the format. It performs during the broadcast and its replay, then the effect fades. That is precisely why recommendation quality matters, but no data has yet been published showing AI investment converting into revenue. On what a winning US playbook looks like, the Twitch executive's argument that trusted voices are the precondition is worth reading alongside this.
What ecommerce operators should take away
Few merchants can acquire an ML team outright. Three points still transfer.
First, ask whether your recommendation refresh rate actually matches your business model. If you sell fast-turning inventory, flash sales, or pre-orders, overnight batch recommendations are structurally late. That Whatnot needed six years to go from a day to minutes also says this improvement is not an overnight project.
Second, treat traffic arriving via external AI and discovery inside your own site as separate KPIs. Growing generative AI referrals and deepening on-site browsing require different investments and different tactics. Leaning entirely on the former hands control of the comparison stage to someone else.
Third, audit whether discovery data is accumulating on your side. What shoppers viewed, what they passed over, where they dropped off. That behavioral record is an asset external AI never receives. Whatnot emphasized 500,000 weekly video hours and millions of interactions precisely because competitors cannot copy them.
Conclusion
The price stays undisclosed, but what Whatnot bet on is clear. As AI rewrites the entry point to product discovery, the company chose to concentrate capital on the discovery experience only it can build.
Two things are worth watching next. Whether Murrell's applied AI research group ships features that genuinely parse live video content. And how far faster recommendations show up in metrics like cross-category buying and repeat visits. Whether live commerce can break past its 5% ceiling in the US depends on those answers.
