Can AI Solve Choice Overload on Content-Heavy Platforms?

Intro

Recommendation systems cut down the number of choices a person has to make, and the data suggests that's worth more to a platform than most people assume. Whether it comes at the cost of stumbling onto something new is a separate question, not so easy to answer either.

Streaming apps ask a lot of a Tuesday night. Open one, and there are rows of thumbnails to get through before landing on anything worth watching, assuming a person lands on anything at all. A lot of people don't. Banners rotate, the same handful of titles keep resurfacing, and at some point closing the app can feel like the path of least resistance. Casino sites like Vegas Stars casino nz run into a related problem from a different angle: a new account logs in and finds hundreds of unfamiliar game tiles with no obvious place to start. Researchers gave this a name a while back, choice overload, and platforms treat it as some sort of a revenue problem.

A Recommendation Engine's Actual Worth

Swap out Netflix's recommendation engine for a plain popularity ranking, and engagement falls by 12%, according to a November 2025 SSRN working paper from a group of economists affiliated with the company (three of the paper's authors work at Netflix directly). Try a matrix factorization model instead, a step down from full personalization, and the cost is smaller at 4%. Multiplied across a subscriber base in the hundreds of millions, that gap adds up to billions of viewing hours, and money that eventually shows up in renewal numbers one way or another.

Klaus Wertenbroch doesn't fully buy that personalization is an unambiguous good. He teaches marketing at INSEAD and coauthored a paper questioning how much weight people put on algorithmic picks in the first place. "My co-authors and I caution consumers against an over-reliance on AI algorithms when making such picks," Wertenbroch noted in the paper, trying to get people to think twice about biases. He said that he hoped recommendations would be seen just for what they are, not as official endorsement but simply the equivalent of advice.

This is related to popularity bias. When enough accounts already watch a show, it can signal to the model to recommend that show to others (it’s safer to recommend an existing hit than some unknown entity), and as a result, small audiences dwindle in subsequent passes, eventually dropping out of sight for those not actively seeking a title. Economists refer to the rich-get-richer effect to explain this process. It has implications for search engine indexing and social media feeds, anywhere that algorithms implicitly use popularity as their proxy for judging quality.

Netflix and INSEAD ended up arriving at the same conclusion, but they took different paths to get there. Even though they were dealing with totally different data from different industries, they found the underlying reason users stick around is usually the same. Growing libraries that do not improve on the quality of the offerings they already have tend to accumulate junk and become inefficient in drawing new users or retaining existing ones.

Algorithms Are Bad at Happy Accidents

Click history, time on page, and search terms form the foundation of most engines. They rely on those digital footprints to serve up more of what you already spent time watching or playing. It makes sense from a design intent perspective. But then again, there is a chance that you may never get shown this particular kind of suggestion. The reason is somewhat subtle, based on the fact that there has never been any indication of interest because you were never introduced to it. Used to be ‘Word-Of-Mouth’ marketing back in the day, before recommendation systems and AI came around. When your brother or sister would give you a mixtape with just a few tracks on there, and suddenly you’d discover a whole new subgenre without ever looking for it.

Content Type	Standard Engagement	Curated Engagement	Key Observation
Broad appeal hits	Already high	Barely moves	Popularity alone gets these titles found
Mid-tier catalog	Buried, easy to miss	Jumps noticeably	Curation earns its keep here specifically
Niche titles	Almost invisible	Ticks up a little	Even good targeting has limits

Pure data struggles to surface those unexpected, fun picks on its own. A site like Vegas Stars manages a broad catalog packed with card games, high-volatility slots, and live dealers running around the clock. It obviously requires some organization to arrange such a library. The machine learning algorithm takes care of the underlying mathematics while smart UX/product design completes the task. If an individual creates a new account without any past experience, AI and good content categories work their magic. Arranging the platform in such a way makes sure that the grid of games doesn't get cluttered.

Decent Curation Goes A Long Way

First-time visitors to Vegas Stars run into something like 200 separate game tiles, with zero click history behind them as of yet. Plenty of older recommendation engines handle that blank slate in an outdated way, mostly because nobody built them with that scenario in mind. Some end up recycling the same handful of genres until everything on screen starts to blur together. Better systems try something harder, tracking what a player already likes, then slotting in an occasional title from outside that pattern just to see what sticks.

• Clicks and session length say more about a player than any survey does.
• Thousands of accounts get compared against each other through collaborative filtering, hunting for overlap in what got played or skipped.
• Volatility and mechanics are read straight from the game files, so style matches follow data, not guesswork.
• A mood change mid-session gets picked up too, something a static profile would simply miss.

Engineers have a name for that opening scenario, a fresh account with nothing on record yet: the cold-start problem. Some systems handle it by defaulting new accounts into whatever's broadly popular until enough behavioral data builds up to personalize properly. Others run a handful of onboarding questions instead, trading a slightly clunkier signup for a head start on matching. Neither fix is perfect, but both beat showing 200 tiles at random and hoping something lands.

Two techniques do most of the actual sorting when you begin to take a look at everything. First up, collaborative filtering establishes a similarity score between one account and thousands of others by comparing the titles that got clicked, played, or even "skipped in common". Content-based tagging skips that comparison and reads metadata off each title instead, volatility index, theme, paytable structure, then matches those attributes against what one account has favored before (the two techniques often run side by side instead of one replacing the other). Either method saves someone from sitting there and digging through every catalog by hand.

Break it down, and it goes through about five steps, each of them pretty straightforward:

1. Session data, category preferences, and skip patterns feed into the model before anything gets ranked.
2. Individual profiles get matched against titles tied to actual click history, not just whatever is broadly popular.
3. Rank it right. Most of the engagement a homepage grid gets happens inside its first few rows, which is part of why ranking order carries more weight than the size of the list underneath it.
4. New clicks, skips, and completed sessions get folded back in as more training data.
5. New releases and changes in player behavior get folded in within minutes instead of waiting for a scheduled update.

What Better Curation Looks Like at Vegas Stars

Mobile sessions make up most of the traffic on a site like Vegas Stars, and that changes what curation actually has to do for the user. A smaller screen has room for maybe six or eight visible tiles before someone has to scroll, so whatever sits in that first visible row carries an outsized amount of weight, similar to the homepage-grid effect the Netflix research points to. Get that opening row wrong, and a player never sees the rest of the library no matter how good the underlying matching is.

Then there's live dealer table games, which adds a whole new layer of complexity. Say you have a blackjack table with three open seats and a user interested in joining; you need to connect them in milliseconds, not seconds spent digging through an appointment calendar. You can't just hit play on a rerun once the seat is taken. Timing matters here unlike an on-demand video library, because once that live table fills, the opportunity is gone for the player. It's like a multiplayer lobby in any Call of Duty or other team game you find on PlayStation and Xbox.

Payout speed feeds into discovery somewhat indirectly, too. Players moving money through crypto or e-wallets tend to cycle through sessions faster than someone waiting days for a bank transfer, leaving a recommendation engine less time per visit to learn anything before that account drifts to a different platform. Faster money in tends to mean a faster verdict on whether the lobby actually worked for that player.

Pulling this off takes real engineering. It requires fast, responsive backend systems paired with a layout built to match where a player's focus actually goes.

Disclaimer: Gambling is a form of entertainment and carries inherent risk. It should not be viewed as a source of income or a financial strategy. Only bet what you can afford to lose, set strict limits, and seek support if gambling is affecting you. Participation is restricted to adults 18+. Always gamble responsibly.