Cold Start Recommendations: What to Show a Visitor You Know Nothing About

The cold start problem is recommending well to someone with no history. The working answer layers signals: start with popularity adjusted for context such as landing page, device, and location, then learn aggressively from the first clicks, searches, and dwell in the session, shifting weight from the crowd to the individual within a handful of interactions.

Popularity is a prior, not a strategy

With zero information, the best single guess is what similar sessions wanted, which usually means current bestsellers filtered by the page the visitor arrived on and the category they are browsing. That prior gets the first screen roughly right and buys the system time to observe. Treating it as the whole strategy is where most cold start implementations quietly fail.

For more on this, see this related guide and our earlier explainer on the same theme.

Pure popularity loops. The same items get shown because they sell, and they sell because they are shown, so the catalogue narrows and new or niche products never gather the evidence to surface. Reserve a slice of early slots for exploration: credible alternatives near the popular items, so the system learns preferences instead of merely confirming its own display choices.

Session signals are the fast teacher

A new visitor teaches quickly if the system listens. The first search terms, the filters applied, the products lingered on and the ones skipped in a list all carry intent, and they are available within the first minute. Session based models that re rank after every interaction routinely beat static bestseller lists by the third or fourth click.

Weight recency over everything at this stage, and forgive fast. An accidental click or a shared device pollutes a fresh profile easily, so early signals should decay quickly unless repeated. The profile of a new visitor is a draft in pencil. Systems that write it in ink on click one spend the rest of the session apologising.

Context you already have, used honestly

Landing page, campaign, referrer, geography at a coarse level, time of day, and device class are all known before the first click, and each shifts the odds. A visitor arriving from a gift guide in December on a phone is a different recommendation problem from a returning category search on desktop, even with identical lack of history. Use that context for the first screen and let behaviour take over from there.

Readers also found this companion piece from our archive useful.

Honesty matters in how context is gathered. Coarse, purpose limited signals declared in your privacy approach are both the ethical and the durable choice, because they survive regulation and browser change. Cold start built on fragile third party data is a system with an expiry date already printed on it.

Measure cold start separately

Blended recommendation metrics hide cold start performance, because known users dominate volume and convert anyway. Segment by session age and profile depth: click through and conversion on first session, second session, and the speed at which recommendations diverge from the popularity baseline for each visitor. That divergence speed is your learning rate, and it is the number to optimise.

Also measure catalogue coverage for new visitors specifically. If new sessions only ever see a small fraction of the range, discovery and margin both suffer, and suppliers notice. A healthy cold start system broadens what newcomers see while narrowing what each individual sees, and those two motions should both show up in your reporting.

Frequently asked questions

How many interactions before recommendations get personal?

Session based systems start improving after two or three meaningful signals such as a search, a filter, or sustained viewing of a product. Full personalisation deepens over repeat visits, but the first session should already feel noticeably less generic by its second half.

Are bestseller lists a bad cold start default?

They are a reasonable first guess and a poor final answer. Bestsellers adjusted for the landing context perform acceptably, but without exploration slots and session learning they trap the catalogue in a feedback loop and teach the system nothing about the individual.

Can I ask new visitors questions instead?

Yes, a short optional preference prompt, common in gifting and style categories, can shortcut learning significantly. Keep it skippable and act on the answers immediately and visibly, or visitors learn that answering honestly changes nothing and stop doing it.

What data should I avoid using for cold start?

Anything you cannot explain or keep lawful: purchased third party profiles, inferred sensitive traits, and identifiers gathered without consent. Coarse context and in session behaviour are effective, durable, and defensible, which is why mature systems lean on them.