Written by: Nicole Ogonowska, IT Growth Manager, Digital Colliers
Whatnot buying Shaped is one of those deals that reads small in the trade press but tells you exactly where the puck is going. Live commerce hosts don't have the luxury of yesterday's batch job. When a seller is holding up a jersey on camera and 4,000 people are watching, the recommendation engine has about 800 milliseconds to figure out who wants the size medium before it's gone. That's not a marketing problem. That's a streaming data problem wearing a video interface.
And the awkward truth for most Shopify-scale brands is that they're still running the batch version of this. Product-affinity scores get rebuilt overnight. Session behavior gets stitched in the morning. By the time your homepage figures out what a shopper actually wants, they've closed the tab.
Batch versus streaming, in plain terms
Most mid-market recommendation stacks look like this. Orders and clicks flow into a warehouse. A dbt job runs at 3am. A recs table gets published. The storefront reads that table. It works. It's cheap. It's also fundamentally blind to what happened in the last ten minutes.
Streaming architecture flips the polarity. Events hit a message bus the moment they happen. Features get updated in memory as the session unfolds. The model scores against the state of the world right now, not the state of the world at breakfast.
The reason this matters more in 2025 than it did in 2021 is pure unit economics. DTC customer acquisition cost is up roughly 40% since 2023, and Meta CPMs have kept climbing through 2024 and 2025. When you're paying that much to land a session, you cannot afford to serve it stale recommendations. UK eCommerce grew about 3% in 2024. Single-digit growth is the new baseline, and the growth is going to whoever converts the traffic they already paid for.
What a session-level feature store actually looks like
Strip the vendor decks away and it's four moving parts:
- An event stream (Kafka, Kinesis, Redpanda, pick your poison) carrying clicks, add-to-carts, dwell time, scroll depth.
- A low-latency store (Redis, DynamoDB, Feast on top of either) holding per-session features that expire when the session ends.
- A feature computation layer that maintains rolling windows: last 5 items viewed, last category, price band drift, cart velocity.
- A model serving layer that reads those features in under 50ms and returns a ranked list.
The model itself is usually the least interesting piece. A well-tuned two-tower retrieval plus a gradient-boosted ranker gets you 80% of the way. What separates the teams that ship from the teams that don't is the plumbing.
And about that. Roughly 95% of enterprise AI projects never reach production. Almost none of them fail because the model was bad. They fail because the features weren't available at inference time, or the latency budget got blown, or nobody could explain why a recommendation changed.
The real blocker is inventory, not the model
Here's the part nobody puts in the pitch deck. Real-time recommendations are only useful if your inventory signal is also real-time. Recommending an out-of-stock SKU is worse than recommending nothing. Recommending a size that will get returned is worse still.
Online return rates sit around 19 to 20% of gross sales, higher in apparel. Roughly 30% of SKUs at a typical multi-channel brand lose money per order once you net out returns and ad spend. Which means your streaming recs engine, if it doesn't know true available-to-promise and doesn't downweight high-return SKUs, will happily optimize you into the losing 30%.
So the questions to ask before you buy any streaming ML product:
- How fresh is our inventory feed to the storefront? Minutes? Hours?
- Do we have per-SKU return rates available as a model feature?
- Can we pull warehouse allocation into the ranking signal, or is it a black box behind an ERP?
If the answer to any of those is uncomfortable, that's the project. The ML is downstream.
Who gets left behind
The operators shipping this well in 2026 tend to have three things in common. They treat the event stream as a first-class product, not a data team side project. They own their inventory signal end to storefront, not just to the warehouse wall. And they measure recommendation lift against margin, not clicks.
Everyone else is going to spend another year rebuilding yesterday's affinity table at 3am and wondering why the paid traffic isn't converting.