Consumer discovery is broken

If you look at recent activity on Spotify, odds are you listened to the same twenty or thirty songs on repeat. You went on YouTube or TikTok and you ended up in the same end zone. You probably tried discovery but nothing genuinely new got through. Nothing surprised you. It felt bland. The feed calls this personalization but it feels like a holding pattern.

There is a great piece in the Financial Times written by Jemima Kelly that made the cultural version of the case — that fashion in 2026 looks like 2006, cinema runs on sequels, and that the bestselling artist of 2025 debuted twenty years ago. She refers to it as mimetic regression.

I don’t believe we lost our taste, but it certainly got optimized away by the consumer internet platforms. The cultural stagnation has lasted over a decade. Algorithmic feeds have flattened taste into more generic sameness. Interfaces are rewarding bite-size catchiness over depth. It feels hard to find genuine inspiration anymore.

The machine is mimetic by design

Look under the hood and the sameness is structural, not incidental.

• Modern recommendation systems are mostly oriented around collaborative signals (people like you also liked…), plus content embeddings (this item is intrinsically similar), plus sequential models (what are you likely to want next?). They are still optimized mostly for short-term engagement — clicks, completion, dwell time, session length. That objective function dominates.
• Heavy weighting of recent behavioral signals, combined with engagement maximization, produces powerful feedback loops. Systems can incorporate exploration, diversity constraints, or longer-term signals. But short-term engagement metrics usually win the A/B tests.
• The research literature has long distinguished “beyond-accuracy” objectives — novelty, diversity, serendipity — precisely because pure accuracy and engagement optimization produces homogenization: repetitive, popularity-based outputs.
• The data is endogenous. A recommender trains on reactions to content that an earlier version of the recommender chose to expose. What’s underlying taste versus revealed behavior?
• Flywheel effect: the demand-side loop rewires the supply side. Creatives are pushed toward instant catchiness over development. Songs are increasingly engineered for the first eight seconds. They’re shorter. The structures or formats are no longer novel.
• Previously much of the highest-scale short-form and feed recommendation talent sat concentrated inside ByteDance, Meta, and YouTube. That concentration, plus the data and compute moats of the incumbents, slows the spread of alternative approaches into new consumer products.

What’s changed

A fundamental shift is underway in the inputs to the discovery machine. Models can now take high-quality multimodal embeddings — audio, visual, text — combined with natural language understanding, and convert stated intent directly, instead of only guessing from previous behavior.

• Multimodal understanding: the machine can understand the item.
• Language: it can understand your stated intent.
• Conversation: it can help steer the exploration process.

Discovery can move beyond pure retrieval-and-ranking of items toward controlled exploration. Platforms already experiment with novelty and exploration in limited ways. The difference is that those efforts remain secondary to the core engagement objective.

The opportunity

Consumer discovery and recommendations are broken. Startups have an opportunity today to define the discovery engine for the end user with novelty and serendipity as primary objectives, not secondary constraints.

The incumbents face real structural headwinds with their objective functions, measurement systems, existing product surfaces, and the fact that their creator economies are all optimized around the engagement loop. Changing that is hard even when the underlying models improve. You need to change the denominator completely. The atomic unit changes.

The old system learnt taste from your behavior. The new system can help you articulate taste you haven’t yet developed. It should feel like an extension of yourself.

That might not just make discovery better. It might make culture feel surprising again — and the status engines that drive cultural change, scarcity, subcultures, gatekeepers, should start to hum once discovery stops defaulting to the familiar.