Before behavior exists, product design supplies the first signal
Topic choices, useful defaults, contextual clues, and deliberate exploration help an app become relevant before it has a history to learn from.
- Onboarding as preference collection
- Defaults that are useful, not empty
- Context without overclaiming certainty
- Exploration that creates learning opportunities
Personalization starts before the system knows you. The first recommendation is a product decision, not an AI miracle.
Open a new app and the blank state has a job. It can show everything, which usually means showing nothing useful. Or it can ask one small question, offer a few good defaults, and create an easy way for the person to say “more like this” or “not for me.”
That is the cold start problem: a recommendation system needs past behavior to personalize, but a new person has no past behavior in the product yet.
In the first article, the loop was:
signals → candidate set → ranked result → feedback
Cold start adapts the same shape. The first signals come from the product experience rather than a long interaction history.
onboarding and context → sensible starting candidates → first useful result → early feedback
Ask for a little, not a biography
Instagram-style topic selection is one familiar pattern: choose a few interests, then see a feed that has a starting direction. A product such as Deepstash can ask what subjects someone wants to learn about. A first visit to Scentum could ask about a few fragrance families or the occasions a person is shopping for.
These are not merely setup screens. They are a compact preference language. The best version does not interrogate people. It asks for the smallest amount of input that materially improves the next result.
Too little: an empty feed with no reason to stay
Too much: a long survey before anyone has received value
Useful middle: a few choices that visibly improve the first screen
The payoff should be immediate. If someone picks “design,” “science,” and “career,” the first set of ideas should reflect that choice. Otherwise the app has asked for effort without proving why it mattered.
Defaults are a point of view
Sometimes a person will not choose topics. That does not mean the product should stop. Popular items, editorially selected collections, recent local activity, or a small set of broadly useful starter paths can make the first session feel alive.
Defaults are never neutral. Choosing “popular this week” favors momentum. Choosing “staff picks” favors editorial judgment. Choosing “near you” favors context. Each can be appropriate, as long as it matches the moment and makes the basis of the recommendation understandable.
The practical rule is simple: show a useful default, then make it easy to move away from it.
Context can help before history exists
The time of day, the device, the entry page, and the task a person just started can provide light context. A person arriving from a link about summer fragrances may benefit from a summer-oriented first result. Someone opening a reading app during a commute may prefer short pieces.
Context is a hint, not a personality verdict. Good products use it to make a starting guess and remain easy to correct. Bad products use it to sound certain about a person they have just met.
Exploration is not a failure mode
Once a product has a first guess, it still needs to learn. Showing only the safest familiar option creates a neat-looking feed but starves the system of new information. A small amount of exploration—one adjacent topic, a new creator, a slightly different scent family—helps the product discover what else may fit.
This is the product version of healthy curiosity. It should be paced so that people get reliable value while the system earns evidence for broader recommendations.
Eventually, topic choices and defaults give way to real behavior: saves, searches, follows, returns, and skips. That is the moment generic relevance becomes personal relevance.
The next piece completes the sequence with the other side of discovery. Feeds push likely discoveries toward us. Scentum lets someone pull a discovery when they can describe the intent—even if the intent is a feeling rather than a keyword.