"Cold Start Is a Solved Problem" Is the Most Expensive Lie in Retail Tech
September 30, 2026 • Sidd Motwani
TL;DR
- Cold start - serving relevant results to a visitor with no history, no login, and no cookie - is the hardest problem in commerce personalization, and most vendors are not solving it.
- Roughly 85% of traffic on most storefronts arrives cold. Legacy personalization stacks serve that cohort a generic homepage and wait for them to click enough things to become useful. Most visitors bounce before that happens.
- Solving cold start requires three architectural elements: contextual initialization, real-time vector updates, and a single intelligence layer across all surfaces. None of these are features you add to an existing stack.
- In production: Jordan Craig +17% new-visitor RPV in a 50/50 A/B test against the incumbent stack. Fun.com +142.5% recommendation revenue. Brunt Workwear +6.5% RPV and +80% upsell click-through.
I'm going to say something that will annoy at least six vendors and convince at least three retailers to ask a different question on their next RFP. Most personalization vendors hide their cold-start performance because the honest answer is: "We serve everyone a generic homepage until they click enough things for us to be useful." Half of those visitors have already bounced. That cohort - first-time visitors with no login, no cookie, no prior session - is the group losing the most money on your site. It's also the cohort your vendor will not show you a number on. Cold-start activation, real-time personalization from the first pageview, ecommerce personalization without identity: these are the questions that separate the vendors with real architecture from the ones selling features on top of a 2015 stack.
Why Is Cold Start the Hardest Problem in Commerce Personalization?
Cold start is the hardest problem in commerce personalization because, unlike streaming or social media, commerce cannot assume a signed-in, identified user. Roughly 85% of traffic on most storefronts arrives cold - no login, no cookie, no prior session. Legacy personalization stacks were built for identified users and have no real mechanism for the anonymous majority.
In music or video, every visitor eventually shows up with history. They sign in. They open the app on the device they always use. They consent to a recommendation system that gets to know them over time.
Commerce doesn't have that. The single largest cohort of traffic on most online storefronts is a visitor with no history. No login. No cookie (or no third-party cookie). No prior session. Maybe they came from a Google query. Maybe they clicked through an ad. Maybe they got a friend's link.
That cohort is roughly 85% of traffic on most sites. It is also the cohort the legacy personalization stack treats worst - because the stack was built to personalize after identity was known. Cold-start performance is the difference between a retailer's headline number being "we lift conversion on identified users by 5%" and the much more honest version: "We don't actually do anything for the 85% of visitors who arrive cold."
What Does Solving Cold Start Actually Require?
Solving cold start requires three things simultaneously: contextual initialization from first-pageview signals, real-time vector updates within seconds of every in-session event, and a single unified intelligence layer that propagates the cold-start profile to every customer-facing surface. Most vendors don't have any of them. All three are architectural - none can be added as a feature to an existing stack.
One: Contextual initialization. The system has to build a useful initial profile from signals available at first pageview - the search query that brought the visitor in, the device, the geographic location, the time of day, the campaign UTMs, the referring page. If your vendor's cold-start strategy is "show the top-selling item until we know more," they're not solving the problem. They're avoiding it.
Two: Real-time update. The first click matters more than the next thousand. The system has to update the vector within seconds of every event in the new session. Batch retraining overnight is too slow. Hourly is too slow. Vector updates within seconds of the click is the threshold.
Three: A single brain across surfaces. When the visitor clicks something on the search results, the recommendations widget on the next page needs to know. When they dwell on a category page, the product detail page they land on next needs to reflect it. If your vendor's search and recommendations run on separate models with separate state, the cold-start visitor sees three contradicting experiences before they bounce.
All three of these are architectural. None of them are features you add to an existing stack. This is why legacy vendors can't solve cold start by trying harder - the substrate is wrong.
What Does Real Cold-Start Performance Look Like in an A/B Test?
Real cold-start performance shows up in the new visitor cohort specifically - not blended across all traffic, not limited to logged-in users, not measured in engagement proxies. The right number is new visitor RPV lift versus an incumbent stack, at statistical significance, in a production environment. Anything below that threshold is either a different question or a vendor without a real answer.
Jordan Craig is the cleanest cold-start result we have in production.
We ran a holiday 2025 product detail page A/B test against the incumbent personalization stack. Two cohorts, 50/50 split, 95%+ statistical confidence. The headline number across all traffic was strong: +8.29% revenue per visitor, +6.99% lift in PDP recommendation engagement, +1.49% in add-to-cart, +4.65% in downstream payment engagement.
But the cohort that mattered was the new-visitor cohort: +17% revenue per visitor. +5% conversion rate. +5-7% AOV.
That's the cohort the legacy stack was leaving on the floor. First-time shoppers, no history, no cookie, no segment. With an intelligence layer underneath the storefront, they found relevant products faster and bought like they already knew the brand. With the legacy stack, they bounced.
The right question to ask any personalization vendor going forward is: "What was your new-visitor RPV lift in the last pilot you ran?"
- If they pivot to a logged-in-user number, they don't have cold-start activation.
- If they pivot to an engagement metric rather than a revenue metric, they don't have cold-start activation.
- If they have a real number above 10%, they're running an architecture similar to ours - which means there are six to eight companies in the world capable of giving you a real answer, and most retail vendors are not on that list.
What Does Cold-Start Activation Look Like at Production Scale?
Cold-start activation at production scale requires three things to be true simultaneously: a two-headed user vector that initializes from contextual signal and updates from in-session signal, a sub-200ms serving layer that returns a personalized response on the first request, and a unified intelligence layer that propagates the cold-start vector to every downstream surface so the second pageview is already informed by what the first one revealed.
This is not a feature. It is an architecture.
We built this architecture at Spotify for music - documented in two published 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). We have now shipped it for commerce.
It is running in production at:
- Jordan Craig: +17% RPV for the cold-start cohort specifically, +8.29% across all traffic
- Fun.com (live since August 2025): +142.5% rec revenue, +31% RPV, 56% of orders influenced by the intelligence layer
- Brunt Workwear: +6.5% RPV, +80% upsell click-through
None of these brands re-platformed. None of them serve a generic homepage to new visitors anymore.
Cold start is a solved problem if you're willing to replace the architecture underneath the stack. It is not a solved problem if you're still trying to add a feature to a system built in 2015.
The next time a vendor tells you cold start is solved, ask them what their new-visitor RPV lift was. The number - or the absence of one - will tell you everything you need to know.
That's the question most vendors hope you won't ask. And it's exactly the gap Malachyte is closing.
Ready to see a cold-start number worth talking about? See how Malachyte's intelligence layer personalizes the experience from the first pageview - no login, no history, no warm-up period required.

