S
Surya Kalipattapu
Guest
Why exploration is essential for users, new content, and creator mobility
TL;DR: Serendipity is not random content sprinkled into an otherwise deterministic feed; it is relevant surprise produced by disciplined exploration. Platforms need an explicit exploration budget because without one, they cannot learn new user interests, evaluate new content, or give emerging creators a credible path to an audience.
A perfectly exploitative recommendation system has a strange failure mode: it becomes increasingly certain about an increasingly narrow version of you.
If you watch tennis, it shows more tennis. If you prefer one creator, it finds adjacent creators. If you complete short videos, it becomes better at finding short videos you will complete. Each decision can be locally correct while the experience slowly loses discovery.
That is why exploration is not a decorative diversity feature. It is how a recommendation system learns anything the previous system did not already know.
Recommendation is also experimentation
Every impression has two possible functions:
- Serve: deliver the item currently expected to create the most value.
- Learn: reduce uncertainty about the user, item, creator, or context so future decisions improve.
Pure exploitation maximizes the first using current knowledge. Pure exploration sacrifices the current experience. The product challenge is to learn at the lowest reasonable user cost.
This is the exploration-exploitation problem behind contextual bandits and reinforcement learning. It is also a product portfolio decision: how much attention should be reserved for uncertain possibilities, for which users, on which surfaces, and under which safeguards?
Serendipity is relevant surprise
Teams often confuse diversity, novelty, and serendipity.
- Diversity means the slate contains meaningfully different items.
- Novelty means an item is unfamiliar to the user.
- Serendipity means an unfamiliar or unexpected item becomes valuable.
Randomness can create diversity and novelty. It does not reliably create serendipity.
Relevant surprise sits just beyond the boundary of known taste. It might be jazz for a tennis fan because a cohort with similar tennis interests also loves jazz, as YouTube has used in a public explanation. It might be an audiobook at a commute time when a listener usually chooses podcasts. It might be a new creator whose content is semantically close to a user's interest but lacks behavioral history.
The system needs both a reason for the bet and uncertainty about the outcome.
New items have no labels until someone takes a risk
Cold start is frequently framed as a modeling problem. It is also an allocation problem.
A new creator's post cannot accumulate positive engagement without impressions. But a system trained on engagement may withhold impressions because the post lacks evidence. This circularity creates a “rich get richer” dynamic even when the model contains no explicit preference for established creators.
Semantic and multimodal representations can estimate relevance from the content itself. Cohort transfer can borrow evidence from similar items. Creator priors can help. Yet none eliminates the need for controlled exposure.
A credible exploration policy asks:
- What minimum evidence does a new item need before broader distribution?
- Which users are most likely to evaluate it fairly?
- How quickly should evidence update the item's opportunity?
- How do we distinguish low quality from unlucky initial exposure?
- How do we prevent exploration pools from becoming a low-quality experience?
The aim is not equal distribution. It is a fair opportunity to generate evidence.
Explore at the point of indifference
One of the safest places to pursue ecosystem goals is where predicted user value is nearly tied.
If the top candidate has a much higher expected value than the alternatives, replacing it may impose a visible cost. If several candidates are within a small relevance band, the system can prefer novelty, creator diversity, freshness, or learning value with little expected downside.
Spotify's Mostra framework uses a related idea: candidates within an epsilon of the top satisfaction score can be reranked for creator and platform objectives.[^16]
This suggests a general policy:
Exploration can also be personalized. A user in an active discovery mode may welcome a wider radius. A user trying to continue a show, find a specific tutorial, or catch up with close friends may not.
Explore slates, not isolated items
Many surfaces present collections: a feed viewport, a row, a playlist, an Up Next panel, or a Stories tray. The value of one candidate depends on its neighbors.
Five individually relevant political clips, breakup songs, or videos from the same creator can form a poor slate. Conversely, one exploratory item can work because the surrounding slate provides familiarity.
This is why item-level relevance is not enough. The system must reason about:
- topic and creator repetition
- format and duration balance
- emotional intensity
- familiarity versus novelty
- sequence effects
- the learning value of the collection
Google's SlateQ research addressed the difficulty of optimizing long-term value when recommendations are presented as interacting slates rather than isolated actions, and validated the approach in live YouTube experiments.[^17] Spotify's contextual calibration work similarly constructs a slate against a predicted content-type distribution instead of choosing every item independently.[^18]
The product insight is simple: the experience is the sequence, not the sum of item scores.
Measure discovery after the click
Teams often count a novel impression or click as successful discovery. That is too early.
A stronger discovery funnel includes:
- Exposure: Was the user given a meaningful chance to notice the item?
- Sampling: Did they start or inspect it?
- Quality: Did consumption and feedback suggest value?
- Adoption: Did they follow, save, add, or return?
- Expansion: Did the item open a durable new interest or creator relationship?
The later steps distinguish curiosity from value. They also help the system learn which exploratory bets compound.
Spotify's long-term podcast work is useful here because its reward is not merely an initial stream; it estimates how many days the listener engages with a newly discovered show over a longer window.[^19]
Exploration needs a budget and a governor
Leaving exploration implicit is risky. The system may under-explore because short-term metrics punish uncertainty, or over-explore because a diversity initiative lacks relevance constraints.
A product team should define:
- Budget: What share of attention or candidate capacity can be exploratory?
- Radius: How far from known interest can the system travel?
- Eligibility: What quality and safety floor must exploratory content meet?
- Targeting: Which users and contexts are appropriate evaluators?
- Learning rule: How quickly does evidence change allocation?
- Stop rule: When should the system conclude that an item or direction is not promising?
- Success: What downstream adoption or satisfaction outcome justifies the exploration?
These policies should vary by surface. An endless discovery feed has room for more experimentation than a notification, where every interruption is expensive. A homepage row can label its exploration thesis; an autoplay queue may need tighter constraints.
The risk of no exploration is invisible
The cost of a bad exploratory recommendation appears immediately as a skip. The cost of failing to explore is counterfactual: the user never discovers a creator, the platform never learns a new taste, and the creator never receives the evidence needed to compete.
Because one cost is visible and the other is not, organizations naturally underinvest in exploration.
But a platform that only harvests known relevance eventually depletes it. Users change. Culture changes. New formats emerge. Creators enter and leave. The system needs a renewable process for discovering value.
Serendipity feels like magic to the user. Underneath, it is disciplined uncertainty management.
Sources and further reading
[^16]: Spotify Research, “Mostra: Balancing multiple objectives for music recommendation”, 2022.
[^17]: Ie et al.,“Reinforcement Learning for Slate-based Recommender Systems”, 2019.
[^18]: Spotify Research,“Calibrated Recommendations with Contextual Bandits on Spotify Homepage”, 2025.
[^19]: Spotify Research,“Optimizing for the Long-Term Without Delay”, 2023.