Cold start: personalisation for new users and new products
What should you know about cold start recommendation?
Cold start is solved with content-based and popularity fallbacks that switch to collaborative signals as history accumulates. A cold start recommendation strategy decides what a new user or product gets on the first visit, how fast the system moves to behavioural signals, and how each stage is measured.
Cold start recommendation is the problem every recommender faces on day one and then every day after: a new visitor with no history, a new product with no interactions, or a new market where nothing has been sold yet. Collaborative filtering, the technique behind "people who bought this also bought", has nothing to work with in all three cases. The answer is not one clever model but a layered strategy: content-based and popularity fallbacks that serve well immediately, and a graceful hand-over to collaborative signals as history accumulates. This article sets out the layers, how to blend them, and how to know the blend is working.
Why cold start matters more than it looks
For most consumer businesses, the majority of visitors on any given day are new or nearly new. A recommender that only performs for the loyal minority is optimising the wrong slice of traffic. On the catalogue side, new products are exactly the ones merchandising cares about, and a system that cannot surface them until they have sold a few hundred units is fighting the business. Cold start is therefore not an edge case to patch later; it is where a large share of the value of a personalisation engine is won or lost.
The three kinds of cold start
| Kind | What is missing | First-line fallback | What unlocks collaborative signals |
|---|---|---|---|
| New user | No interaction history | Popularity by context (region, device, landing page, referrer) and content similarity to the first item viewed | A handful of clicks or views within the session |
| New product | No one has interacted with it | Content-based embedding from title, attributes, images and description; placement in existing categories | Early exposure through explore slots, then real interaction data |
| New market or tenant | No history for the whole segment | Rules and curated sets, transferred models from similar markets | Enough traffic to train segment-specific models |
Fallback layers for new user recommendations
Layer one: contextual popularity
Popularity is underrated. The best-selling items for a visitor's country, device type, hour of day and landing category are a strong default, far better than a global bestseller list. The context comes free with the first request. Compute these lists on a schedule, keep them small and refresh them often, so the fallback is never stale during a sale or a festival.
Layer two: content-based similarity
The moment a visitor views one product, you know something. A content-based fallback embeds every item using its attributes, text and images, then recommends the nearest neighbours of whatever the user just looked at. This works for a product that has never sold and for a user who has never bought. The embeddings can come from a small open-weight model run on your own infrastructure, which keeps the running cost predictable. Content similarity is also the mechanism that lets new products appear beside established ones from the first hour they are listed.
Layer three: session signals
Within a single visit, a short sequence of views is enough to sharpen the picture considerably. Session-based models take the last few interactions and predict the next one, without needing a long-term profile. This is the bridge between cold start and full personalisation, and it is covered in depth in real-time ranking within a session.
Layer four: onboarding questions
Asking a new user two or three questions on first visit, such as their preferred categories or sizes, is a legitimate cold-start technique when the questions are few and clearly useful. It works well in education, media and fashion, less well where visitors want to browse immediately. Every question adds friction, so measure sign-up completion alongside recommendation quality before adding one.
The hand-over: blending fallbacks with collaborative signals
The mistake is to treat cold start as a switch: rules until the user has ten interactions, then the model. Real systems blend. A common pattern scores each candidate under each available signal, then weights the scores by how much evidence each signal has for this user. A brand-new visitor is served almost entirely by contextual popularity and content similarity; after a few views the session model dominates; after a few purchases the collaborative model takes most of the weight. The weights are tuned against a held-out control, not by intuition. If you have not read why you must run a controlled test, that is the place to start before tuning anything.
Recommender cold start for new products: explore slots
New products need exposure to gather the interactions that will eventually let the collaborative model rank them. Reserve a small share of recommendation slots for exploration: items with high content similarity but little history get shown to a fraction of relevant visitors, and the results feed straight back into the model. Bandit-style allocation does this systematically, increasing exposure for items that convert and cutting it for items that do not. The cost is a small, measurable dip in short-term conversion in exchange for a catalogue that ranks itself within days rather than weeks. Merchandising teams like this because it replaces manual "new arrivals" pinning with something that learns.
Measuring cold-start performance separately
A single average lift will hide cold-start failures, because the loyal minority carries the number. Report recommendation quality and conversion by user tenure: first visit, first week, first month, established. Report product performance by days since listing. When the first-visit segment shows no lift over a plain bestsellers list, the fallback layers are not doing their job, whatever the headline says. This segmentation is cheap to build if the event pipeline carries user tenure and item age as attributes.
Privacy and the cold-start temptation
Cold start tempts teams to buy third-party data or fingerprint devices to fill the gap. Under India's DPDP Act and the GDPR, that path carries consent and purpose-limitation obligations that are hard to meet for a marginal improvement. Contextual signals such as region, device and referrer are available without any of that, and content-based similarity needs no personal data at all. Build the honest layers first; they usually close most of the gap.
A worked example
A D2C brand selling across a storefront and WhatsApp had a recommender that performed well for repeat buyers and did nothing for anyone else, which mattered because most daily visitors arrived from ads and had never been seen before. We rebuilt the serving path in layers. First-visit traffic received popularity lists keyed on the landing category and the ad campaign that brought them, refreshed hourly. As soon as a product was viewed, content embeddings drove a "similar to this" strip, which also gave newly listed items exposure from day one. Session signals took over after a few views. A holdback on the old system stayed in place throughout. Lift for first-visit users moved from indistinguishable to clearly positive, and the merchandising team retired their manual new-arrivals pinning. The wider engagement is described in the D2C personalisation and WhatsApp case study.
Team and timeline
A cold-start layer is a natural first phase of a personalisation build because it needs no long history: a data engineer for the popularity and attribute pipelines, an ML engineer for embeddings and blending, and a product owner from merchandising. The fallback layers ship in roughly three to four weeks; the blending and explore slots follow once collaborative data exists. This is delivered within the personalisation engines service, from $21,000 / ₹13.6L, or scoped first in a Sprint Zero discovery at $3,250 / ₹2,00,000, credited to the build. Running costs are dominated by embedding refreshes and the feature store, and the pricing page lists the Care Plans that keep them monitored.
Before you start: a checklist
- Measure what share of daily visitors are new or near-new, so you know how much cold start matters
- Confirm product attributes, descriptions and images are complete enough to embed
- Decide which context signals (region, device, referrer, landing page) are available at first request
- Agree with merchandising how much exposure to reserve for new products
- Check that user tenure and item age flow through the event pipeline for segmented reporting
- Set up the control group before the first fallback layer goes live
- Decide whether onboarding questions are acceptable for your audience
Questions clients ask
- How many interactions before collaborative filtering works? There is no fixed number; the blend shifts weight gradually as evidence accumulates, and the control test tells you where the crossover sits for your catalogue.
- Can we use an LLM to solve cold start? LLM embeddings of product text are a good content-based signal, but an LLM is not a recommender; it feeds one layer of the blend.
- Do we need onboarding questions? Only if visitors will tolerate them; for most storefronts context and content similarity are enough.
- What about a brand-new tenant in our SaaS? Start with curated sets and rules, transfer a model from similar tenants where the data allows, and switch as their own history grows.
- How do we stop new products being buried? Explore slots allocated by a bandit, with merchandising able to set a floor for launches.
Related reading
See recommendation engines explained for e-commerce leaders for the basics and personalisation for onboarding for the SaaS case. Google's Recommendations AI documentation discusses cold-start handling in a managed context. Engagement models and prices are on the pricing page.
Cold start is not a gap to apologise for; it is where a well-layered recommender earns its keep on the visitors that matter most.
Frequently asked questions
What is the cold start problem in recommendation?
▾
It is the lack of interaction history for a new user, a new product or a new market, which leaves collaborative filtering with nothing to learn from. It is solved with contextual popularity, content-based similarity and session signals, blended as evidence grows.
What is a content-based fallback?
▾
A method that recommends items similar in attributes, text or images to what the user just viewed. It works with no interaction history, so it serves both first-time visitors and newly listed products from the first hour.
How do you measure cold start recommendation quality?
▾
Report lift by user tenure and by item age against a held-out control, not as one average. If first-visit users show no gain over a bestsellers list, the fallback layers of your personalisation engine need work.