How do you recommend to a brand-new subscriber who has interacted with nothing yet?

## Why the model cannot help New-user cold start is structurally different from new-item cold start. When an item is new you still hold a description of it, so a content-based or feature-augmented path can place it. When a *user* is new you hold nothing about the person — no history to average into a profile, no interactions to fit a vector from. Item features do not rescue this: the missing side is the user. Everything below is therefore about acquiring signal, not about modelling. ## Lever one: infer from context you already have Before asking a single question, a request carries information: - **Locale and region** — a huge prior on language, and often on topic. - **Referral source and landing page** — someone who arrived on a sports article is not a random visitor. - **Device and time of day** — a phone at 07:40 and a desktop at 22:00 are different sessions. - **Signup channel** — a campaign or partner integration says a lot about the segment. This costs the user nothing and never annoys anyone, so it is always the first lever. It also produces a fallback that is *scoped* rather than global, which is most of the gap between a useful default and a generic one. ## Lever two: elicit with an onboarding prompt Asking a new subscriber to pick five topics before their first session buys an instant profile in the item feature space. It is the cleanest source of signal for a genuinely unknown user, and it is used widely for exactly that reason. The costs are real: - **Friction.** Every screen between arrival and value loses people. A survey that boosts first-session relevance but drops signup completion is usually a net loss, and you find out only if you measure completion *and* downstream retention, not relevance alone. - **Stated versus revealed preference.** People select what they would like to be someone who reads. The picked topics are a useful prior, not ground truth, and treating them as ground truth locks a user into an aspirational identity their clicks contradict. - **Design pressure.** The choices must be few, mutually distinguishable, and expressed in the user's language rather than your taxonomy's. Too many options and people click the first three; too few and the profile is as coarse as no profile. - **Never ask for what you can infer.** Asking for a country you already have from the request is pure friction. A softer variant is interactive: show a small set of items and let the user react to a few. It elicits taste in the same space the recommender works in, and feels like using the product rather than filling in a form. ## Lever three: the non-personalised fallback Some users will refuse onboarding, and some will arrive with no usable context. The fallback must be genuinely good, not a placeholder: - **Trending or most-engaged recently**, scoped to whatever context you inferred, is a strong baseline because popular items are popular for reasons that apply to newcomers too. - **Editorial or curated** selections beat automation when the catalogue is small or the brand voice matters. - **Diversify deliberately.** The first screen is your best chance to learn what this person is about, so a list spanning several areas produces more information than ten near-identical items — and the user's first clicks are worth more than the marginal relevance you gave up. What you should not do is serve a personalised score from an untrained user profile. It is arbitrary, it looks broken, and it is strictly worse than an honest popularity ranking. ## The handover is the real engineering The interesting part is not any single lever but the transition. A blend weighted by evidence works well: ``` score = w(n) * personalised + (1 - w(n)) * fallback ``` where `n` is the number of observed interactions and `w(n)` rises smoothly from 0 toward 1 — for instance `w = n / (n + k)`, with `k` the interaction count at which you trust the two sources equally. This avoids a discontinuity where the feed visibly changes character at some threshold, and it lets you tune a single interpretable constant. The elicited onboarding prior should decay on the same schedule. Onboarding answers are the best evidence you have in session one and among the worst you have after a hundred clicks; a profile that still weights the signup survey a month later is describing a person who no longer exists. ## How to judge it Measure onboarding by **downstream retention**, not by completion rate or first-session relevance. The question is whether users who went through it are still active in week four, compared to users who were routed straight to a good contextual fallback. It is entirely common for a well-scoped fallback to beat a questionnaire, and that result is only visible if the comparison was set up as an experiment.