The Cold Start Problem Is a Belief Problem

TL;DR

• The cold start problem is a question problem: systems wait for behavioral clicks when they could ask users what they believe from interaction one.
• Spotify loses 13.8% recommendation accuracy without onboarding preference signals, and Netflix loses 12% engagement with popularity-based fallback.
• Belief elicitation goes deeper than preference selection by modeling why users make choices, producing a useful self-model in seconds instead of weeks.

The cold start problem is a belief problem because recommendation systems wait for behavioral data when they could be asking users what they believe from the first interaction. Behavioral systems are structurally designed to learn after users have already decided whether to stay or leave. According to AppsFlyer’s retention data [1], the average app retains only about 21% of users after the first 24 hours. This post covers how Spotify, Netflix, Pinterest, and TikTok solve cold start through explicit preference elicitation, the research showing belief-based bootstrapping outperforms passive observation, and how self-models take this approach further.

What Cold Start Actually Looks Like

Consider what happens in practice. A new user signs up for a product. Here is what a behavioral recommendation system knows:

1{← new user, interaction 1
2  "user_id": "usr_new_7291",
3  "behavioral_signals": [],← empty
4  "click_history": [],← empty
5  "preferences": null,← unknown
6  "segments": ["all_users"],← useless
7  "recommendation_strategy": "popularity_fallback",← the default everyone gets
8  "personalization_confidence": 0.0← zero
9}

Personalization confidence: 0.0. Strategy: popularity fallback. This is the same approach that Netflix’s recommendation system [2] and Amazon’s item-to-item collaborative filtering [3] pioneered. Reasonable at scale, but a dead end for new users. The system serves every new user the same thing and calls it a “recommendation.”

How much does that cost? Netflix researchers quantified it [4]: replacing their personalized recommendations with popularity-based ones reduces engagement by 12%. With random recommendations, engagement drops 16%. Their recommendation system as a whole is estimated to save the company over $1 billion per year [5] in retained subscribers. Cold start, the period where every new user effectively gets the popularity fallback, is where that value is most at risk.

The Behavioral Waiting Game

The traditional playbook for cold start has three phases:

Phase 1: Popularity fallback (first sessions). Serve what’s popular. Hope the user clicks on something so you have a signal.

Phase 2: Collaborative filtering bootstrap (early weeks). “Users similar to you liked X.” But “similar” is determined by the handful of clicks collected so far, barely enough to distinguish preferences from noise.

Phase 3: Personalized recommendations (weeks later). Finally, enough data to build a meaningful model. By which point most users have either activated or churned.

The fundamental problem: behavioral systems are designed to learn from usage. But users decide whether to keep using a product before the system has learned enough to be useful. 63% of customers consider the onboarding period [6] when deciding whether to subscribe. The system asks users to have patience with a generic experience while it slowly figures out who they are.

The Platforms That Ask First

The best consumer products already know this. They just don’t call it “belief elicitation.”

Spotify asks new users to select favorite artists immediately at signup. Spotify’s VP of Personalization, Oskar Stal [7], describes how these selections let the system “spin up creators a listener might love based on those they’re already familiar with.” The data backs it up: Spotify Research found [8] that removing onboarding preference signals (selected artists, genres, languages) reduces new-user recommendation accuracy by 13.8%.

Pinterest takes it even further by gating the experience behind preference elicitation. Casey Winters, Pinterest’s former growth lead [9], explains that Pinterest requires users to follow interest topics before seeing their home feed. No interests selected, no feed. The team ran hundreds of A/B experiments to optimize this interest-selection step and found that pairing browser country data with personalized topic suggestions improved activation rates by 5-10%.

TikTok asks new users to select interest categories [10] like pets, travel, and sports to seed the For You feed before any behavioral data exists. Netflix asks users to rate or select titles during onboarding specifically to seed the recommendation engine.

The pattern is universal: every platform that wins the onboarding game asks before it observes.

From Preferences to Beliefs

But here’s what these platforms are still missing: they’re eliciting preferences, not beliefs.

The distinction matters. A preference is “I like jazz.” A belief is “I think music is best when it surprises me.” Preferences are surface-level and unstable; they shift with mood, context, and novelty. Beliefs are structural and predictive. Knowing that a user believes music should surprise them makes it possible to predict their preferences across genres, moods, and formats, without them clicking on a single thing.

The preference elicitation literature [11] from the RecSys community has long studied the trade-off between asking and observing. Research on active learning in recommender systems [12] shows that strategically selecting which items to present for rating outperforms random or popularity-based selection for seeding a new user profile. A 2021 survey of cold start elicitation approaches [13] catalogs how auxiliary information, including explicit user input, can bootstrap effective recommendations for new users.

Belief elicitation goes deeper. Instead of asking “what do you like?”, which maps to content, the question becomes “what do you believe?”, which maps to why a user makes the choices they do.

1// Step 1: Elicit beliefs (not just preferences)← seconds, not weeks
2const beliefs = await clarity.elicitBeliefs(userId, {
3  domain: 'developer_tools',
4  questions: [
5    'What matters most in a dev tool: speed, flexibility, or simplicity?',
6    'Do you prefer to read docs or explore by building?',
7    'Is your primary goal shipping faster or reducing bugs?'
8  ]
9});
10
11// Step 2: Build initial self-model← instant
12const selfModel = await clarity.createSelfModel(userId, {
13  beliefs: beliefs,← 3 beliefs, moderate confidence
14  source: 'cold_start_elicitation'
15});
16
17// Step 3: Generate recommendations← personalized from interaction 1
18const recs = await clarity.recommend(selfModel, {
19  type: 'content',
20  count: 5
21});
22
23// Result:← not popularity fallback
24// { confidence: 0.72, strategy: 'belief_based',← 0.72 > 0.0
25//   items: [{ id: 'quickstart-api', reason: 'build-first learner' }] }

Personalization confidence: 0.72. From three questions. In the time it took the behavioral system to serve a popularity fallback.

The Evidence

The data across platforms tells a consistent story:

The pattern extends beyond consumer platforms. A 2025 study in Nature Scientific Reports on active learning for cold start [19] shows that dynamically selecting which items to ask about, rather than waiting for organic interactions, improves recommendation accuracy from the first session. The foundational KDD 2016 work on conversational recommender systems [20] by Christakopoulou et al. demonstrates that even simple interactive questioning can improve personalized recommendations by 25% after asking only two questions.

The key insight from the Netflix research is particularly striking: 41.9% of their recommendation value [21] comes from targeting, meaning matching the right user to the right content, versus only 6.8% from mere exposure. Personalization’s advantage is overwhelmingly about fit, not visibility. Fit is exactly what cold start systems fail to deliver.

Why This Matters for Enterprise AI

Cold start has been a known problem for decades, since at least the early collaborative filtering work by Goldberg et al. (1992) [22] on the Tapestry system. The traditional solutions (popularity fallback, collaborative filtering, content-based bootstrapping) have been good enough for consumer recommendation engines where users interact dozens of times a day.

But enterprise AI products don’t get dozens of interactions per day. They get a handful per week. The patience window is shorter. The cost of a bad first impression is higher. The users are sophisticated enough to know when they’re getting generic content.

Bain & Company research [23] shows that improving retention by just 5% can boost profits by up to 95%. And acquiring new customers costs 5-25x more [24] than retaining existing ones. Every user who bounces during the cold start window, before ever experiencing the product’s actual value, is a compounding loss.

For anyone building an AI product for enterprise, whether a copilot, a recommendation engine, a learning platform, or a support system, cold start is not a minor UX issue. It is the primary reason users bounce before they ever experience the product’s actual value.

The cold start problem is not about lacking data. It is about waiting for the wrong kind. Behavioral data is the exhaust of decisions. Beliefs are the engine. And the engine is available from day one, for those who know how to ask.

Solve cold start in seconds, not weeks. Bootstrap self-models with Clarity.

References

1. AppsFlyer’s retention data
2. Netflix’s recommendation system
3. item-to-item collaborative filtering
4. Netflix researchers quantified it
5. over $1 billion per year
6. 63% of customers consider the onboarding period
7. Spotify’s VP of Personalization, Oskar Stal
8. Spotify Research found
9. Casey Winters, Pinterest’s former growth lead
10. select interest categories
11. preference elicitation literature
12. Research on active learning in recommender systems
13. 2021 survey of cold start elicitation approaches
14. 13.8% accuracy drop without onboarding signals
15. 12% engagement drop with popularity fallback
16. $1B/yr retention value
17. 5-10% activation improvement
18. Seed signal for For You feed
19. 2025 study in Nature Scientific Reports on active learning for cold start
20. KDD 2016 work on conversational recommender systems
21. 41.9% of their recommendation value
22. Goldberg et al. (1992)
23. Bain & Company research
24. acquiring new customers costs 5-25x more