Beyond segmentation: the real-time personalization layer ecommerce needs | Malachyte

The Intelligence Layer for Commerce

By Sidd Motwani, Co-founder & CEO, Malachyte

TL;DR

- Most ecommerce personalization still runs on segment-based, batch-trained models built before 2020- and they're breaking down as anonymous traffic grows.

- A real-time intelligence layer cold-starts a unique visitor profile from page one: no login, no cookies, no history required.

- When search, recommendations, and merchandising share one user vector, they stop contradicting each other, and conversion follows.

- In production: FUN.com +142.5% rec revenue, Brunt Workwear +$9.75M projected annual lift, Jordan Craig +17% new-visitor RPV.

The personalization stack running most online stores today was architected between 2014 and 2019. It was rational for its era. Cookies were rich, traffic was a known quantity, and "personalization" meant "people who bought this also bought that." A separate vendor for search, a separate vendor for recommendations, a separate vendor for merchandising rules — each batch-trained overnight, each maintaining its own model of the visitor, each contradicting the others on the surface.

That era is over.

Cookies are deprecated. Anonymous traffic is the majority of what hits most storefronts. AI agents are about to start arriving on behalf of shoppers, and they don't have history at all. Shoppers expect Spotify-grade relevance from session one, because the only places they get it- TikTok, Spotify, the Instagram feed-have already set the bar.

The category we're building is not personalization. Personalization is the wedge. The category is the intelligence layer for commerce: a foundation model for user behavior, with continuous learning infrastructure underneath, serving every customer-facing surface from one shared brain.

Why the Legacy Personalization Vendor Stack No Longer Works

The legacy personalization vendor stack fails because its three core assumptions- that visitors are identified, that batch retraining is fast enough, and that search, recommendations, and merchandising can run on separate models- have all broken simultaneously. Patching individual vendors doesn't fix an architectural problem.

If you ran a digital retailer in 2017, bolting on vendors was the right move. A search vendor for the search box. A recommendations vendor for the carousel. A merch-rules vendor for the campaign overrides. Each specialized, each batch-trained on historical data, each cookie-aware.

Three architectural assumptions held that pattern together:

- The visitor is identified. A login or cookie was the entry point to a meaningful

experience.

- Batch retraining is fast enough. Overnight model updates were sufficient because

session-level intent didn't shift faster than a day.

- Surfaces can specialize. Search and recs and merch can run on separate models

because intent is mostly stable across them.

Every one of those assumptions has broken. The visitor is rarely identified before they convert. Session-level intent shifts in seconds, not days. And the moment a shopper's behavior on the search surface contradicts what the recommendations widget is showing them, conversion drops. That's not a problem you fix by tuning the vendors harder. It's a problem you fix by replacing the architecture underneath them.

What Real-Time Ecommerce Personalization Actually Looks Like

Real-time ecommerce personalization means generating a single user vector per visitor on the first pageview,from contextual signals like referrer, device, geo, and query, then updating that vector within seconds of every click, scroll, dwell, and add-to-cart. Every surface reads from the same model. No warm-up period. No contradictions.

The intelligence layer is a single brain that produces a single user vector for each visitor, in real time, and exposes that vector to every customer-facing surface. The vector initializes from contextual signals on the first pageview: the query that brought the visitor in, device, geo, time, referrer, campaign. The experience is personalized from page one, with no warm-up period. Then it updates within seconds of every click, scroll, dwell, and add-to-cart.

It's generated by a two-headed transformer model. A slow head learns who the visitor is over weeks. A fast head learns what they want this session. Both operate on the same vector. That's the architecture TikTok and Spotify run in production. Six to eight companies in the world do it at scale. Nobody has done it for commerce. We are.

The intelligence layer routes that vector to every surface that needs to make a decision. Search ranks against it. Recommendation widgets sort against it. Category pages re-order against it. Merch teams set business priorities through a control surface the model respects. Same brain, multiple surfaces, no contradictions.

What Changes When Personalization Runs from a Single User Model

When personalization runs from a single user model, cold start stops being a problem, surfaces stop contradicting each other, and merchandisers retain real-time control, all at once. The results aren't incremental: brands in production with Malachyte have seen rec revenue lifts of 142.5% and RPV gains in the double digits for new visitors.

Cold start stops being a problem. Most legacy stacks serve a generic homepage to 100% of new visitors because they have nothing to personalize against. The intelligence layer starts with a contextually-initialized vector and adapts on the first click. At Jordan Craig, in a holiday A/B test against the incumbent stack, new-visitor RPV moved +17%. First-time shoppers found more relevant products faster and bought like they already knew the brand.

Search, recs, and merch stop contradicting each other. When all three surfaces read from the same user vector, the signal a shopper sends on one surface immediately informs the others. At FUN.com, live since August 2025, rec revenue is up 142.5% year over year. Revenue per visitor is up 31%. 56% of orders are influenced by the intelligence layer. That is what unified intelligence looks like on real money.

The merchandiser stays in the seat. Black-box AI loses every enterprise deal. The intelligence layer ships with a control surface: pin a campaign, override the model, set the personalization intensity per surface, simulate a strategy before it goes live. The AI learns from the override. At Brunt Workwear, the Q4 2025 pilot drove +6.5% RPV and +80% upsell click-through, projecting $9.75M in incremental annual revenue against a $150M baseline. The numbers work because the merchandiser is still allowed to do their job.

Why Agentic Commerce Makes Real-Time Personalization Non-Negotiable

Agentic commerce makes real-time personalization non-negotiable because AI shopping agents arrive with zero history- no cookies, no login, no prior sessions. They require a sub-200ms, in-session user model to query. Batch-trained, segment-based stacks have no vector to return. Brands without an intelligence layer underneath won't get surfaced.

Personalization isn't a feature. It's the substrate.

In the agentic era, McKinsey projects $3–5 trillion in agent-orchestrated retail spend by 2030, and IBM's January 2026 study found 45% of consumers already use AI for part of the buying journey. The brands without an intelligence layer underneath their stack don't get shown to the agents shopping on behalf of their customers.

Agents arrive cold. No history. They need a real-time, in-session model of the visitor that they can query at sub-200ms latency. The 2015-era personalization stack cannot serve them. There's no vector to read, no continuous learning loop to update, no shared brain for the agent to negotiate with.

The entire vendor category has spent ten years selling personalization as a widget — a recs carousel, a search bar, a popup — instead of as infrastructure. The Bessemer Process didn't sell better steel; it changed what was possible with steel. The intelligence layer doesn't sell better personalization; it changes what's possible underneath the storefront.

What This Looks Like in Production

The team building Malachyte led product and architecture on the Spotify version of this. The work is documented in two papers: Generalized User Representations for Transfer Learning (Fazelnia et al., arXiv:2403.00584) and Variational User Modeling with Slow and Fast Features (Fazelnia et al., WSDM '22)- describing the architectural primitives we've carried into commerce.

The results are real:

- FUN.com: +142.5% rec revenue and +31% RPV across a $300M+ business

- Brunt Workwear: +6.5% RPV in pilot and a $9.75M projected annual lift

- Jordan Craig: +17% RPV for new visitors, +8.29% all-traffic

- New York & Company: in pilot

None of these brands re-platformed. None of them serve a generic homepage anymore. None of them are running on segments built in 2017.

The intelligence layer is here. The question is which retailers ship it before the agents start asking for it by name.

If you're a retail or commerce leader thinking about what comes after the 2015 personalization stack — let's talk. Malachyte is the intelligence layer your storefront needs before agentic commerce arrives. → Book a conversation