Spotify’s recommendation playbook is heading to e-commerce, and that’s a bigger deal than the funding round

Three former Spotify employees have raised $10 million to do something that sounds obvious after the fact: use behavioral signals to shape shopping in real time, not after the session is over.

Their startup, Malachyte, builds on the kind of infrastructure the team worked on at Spotify, where its Vector AI system reportedly drives about 90% of recommendations for 800 million users. The pitch is simple. Most stores still depend on blunt personalization signals like purchase history, demographic buckets, logged-in profiles, and overnight segmentation jobs. Malachyte wants to react to what a shopper is doing right now.

That sounds like a small product tweak. It isn’t.

Why this matters

E-commerce personalization has been stuck in a fairly crude mode for years. A first-time visitor often gets the generic homepage. A returning customer usually sees recommendations shaped by what they bought last month, not what they’re trying to buy this minute.

That’s fine for repeat purchases. It breaks down when intent shifts quickly. Someone who usually buys office clothes might be shopping for hiking gear today. A parent looking for school supplies doesn’t need a recommendation engine trained on last season’s cart.

Malachyte’s approach is to treat each session as a live signal stream. Search terms, hover events, click sequences, scroll depth, add-to-cart behavior, device type, referral source, and time of day all feed the model. The system updates as the session unfolds, then carries that signal into the next visit.

Retailers already have most of that data sitting in logs. The problem is that a lot of them aren’t doing much with it while it still matters.

Session intent and long-term taste

The company calls its model a “two-headed Vector AI.” Marketing language aside, the idea is familiar and sensible.

One head predicts near-term intent. What is the shopper trying to do right now? Looking for a gift. Replacing work boots. Browsing without buying. The other tracks broader taste, the slower preference signal that says this person tends to buy premium brands, prefers minimalist design, or avoids certain categories.

That split matters because a single embedding or user profile often mixes two different things: durable preference and temporary context. You don’t want someone who clicked on steel-toed boots to be stuck in workwear forever. You also don’t want to ignore the fact that they clicked on steel-toed boots at all.

Motwani says the system starts forming before the first click, using whatever context is available when the page loads. That can include referral source, device, time, and sometimes geography or campaign metadata if the merchant has it. The model then refines itself as the session goes on.

Personalization doesn’t have to wait for identity resolution. Anonymous traffic still leaves a useful trail.

Why this is harder than it sounds

Moving from recommendation engines into commerce search and merchandising is not trivial. Spotify can optimize for clicks, listens, skips, and long-term engagement in a relatively controlled environment. Retail is messier.

The catalog is larger. Inventory changes constantly. Product attributes are inconsistent. Merchandising teams want to push margin, promotions, and seasonal stock, not just whatever looks most relevant. Then there’s the cold-start problem for products as well as users. A new SKU can’t depend on interaction history it doesn’t have.

Real-time systems also carry real infrastructure costs. If Malachyte is reading every hover, scroll, and search refinement continuously, it’s running an event-driven pipeline with tight latency expectations. Tracking, feature generation, model inference, and ranking all have to stay fast enough not to slow the page down. Retailers are not forgiving when conversion slips.

There’s another risk too: overfitting to momentary behavior. A shopper can click around in ways that look like intent and still be nowhere near a purchase. If the model reacts too fast, the storefront gets strange. A few bad updates and you’re showing work pants to someone who only wanted to compare boot styles.

More signal can mean better relevance. It can also mean more noise.

Why the Spotify connection matters

This story is less about generic “AI for commerce” and more about recommendation expertise moving from media to retail.

Spotify’s recommendation stack had to handle intent, context, and sequence at huge scale. Music consumption has a rhythm that looks a lot like browsing behavior: short sessions, repeat visits, shifting moods, and lots of weak signals. That makes the jump to shopping feel more credible than a standard AI startup pitch.

Commerce has one big difference, though. The system can directly affect revenue in ways that are easy to measure and hard to hide. If the recommendation layer pushes the wrong products, conversion drops. If it nudges shoppers toward the wrong price tier or category mix, margin can suffer. If it gets too aggressive, customers leave.

Retailers won’t just ask whether the model is smart. They’ll ask whether it’s controllable.

Can merchandisers override it? Can brands protect certain placements? Can teams set guardrails around inventory, exclusions, and promotion rules? Can the system explain why an item is being recommended, at least enough for internal debugging? Those questions matter more in commerce than they do in entertainment.

What the rollout says about the product

Malachyte says it started testing in 2024, worked with more than 20 enterprise customers across travel, grocery, and retail, then narrowed its focus to e-commerce. The platform went live with Fun.com in fall 2025. Since June 2026, it’s been generally available to Shopify merchants through a native integration, with API access for larger retailers.

That setup says a lot about where the company is headed.

The Shopify route points to faster adoption with merchants who want better recommendations without building their own ML stack. The API route is for larger retailers with existing commerce systems, data teams, and more complicated integration needs.

The native integration is easier to sell, but it comes with limits. Shopify merchants will want something that drops into the storefront with minimal engineering work. That usually means less customization and fewer ways to tune the model to a specific business. The API gives more control, but the integration gets heavier. The retailer has to manage event collection, identity stitching, ranking hooks, and probably experimentation infrastructure too.

That’s where these products often stall. The model can be good. The operational work can still be miserable.

What technical teams should ask

For engineering teams looking at something like this, a few questions matter more than the pitch.

First, what’s the event schema? If the system is learning from clickstream behavior, the quality of the ingestion layer decides a lot. Bad event names, missing timestamps, duplicate events, and inconsistent product metadata will wreck the model long before any embedding layer helps.

Second, how is identity handled? Anonymous browsing, logged-in users, cross-device sessions, and privacy consent all create edge cases. A system that claims to personalize “before the first click” still has to respect consent rules and cookie restrictions, especially in stricter privacy markets.

Third, what’s the latency budget? Real-time personalization only matters if ranking comes back fast enough to shape the experience. If inference adds visible delay, users won’t care that the model is clever.

Fourth, how does the system avoid turning into a black box for merchandising teams? Retailers usually need a mix of automatic ranking and business rules. Pure model output is rarely enough.

Finally, how does it handle catalog churn? E-commerce inventories change constantly. Out-of-stock products, price changes, new collections, and seasonal shifts can all invalidate yesterday’s relevance signals.

The money is real, but the category still has to prove itself

A $10 million seed round from Bessemer Venture Partners, Gradient, and Harpoon Ventures is a solid signal that investors think this team has something defensible. The Spotify pedigree helps. So does the fact that the product is already live with customers.

But commerce personalization has a habit of sounding smarter than it is. Plenty of vendors can talk about embeddings, vectors, and contextual signals. Fewer can ship a system that reliably improves conversion across messy catalogs, uneven traffic quality, and real merchant constraints.

Malachyte looks more credible than most because it comes from people who already built high-scale recommendation infrastructure, and because the product premise is narrow enough to test. If a retailer can show better session-to-cart behavior, stronger product discovery, or higher AOV without hurting site performance, that’s a real win.

If not, it’s another recommendation layer with a better pitch.

Either way, shopping is starting to look a lot more like streaming behavior analysis. That’s probably where it was headed anyway.