Search Lengthens The Tail, Recommendations Shorten It. OnlyFans Has Neither

The most repeated piece of advice in the creator economy assumes a kind of infrastructure that this platform never built.

A large North American retailer installed a recommender system and measured what happened. Sales of niche products went up, exactly as the theory predicted. Sales of popular products went up more. The overall range of what customers bought got narrower.

That result, from a randomized field experiment by Daniel Lee and Kartik Hosanagar, is described in a 2024 survey of the research literature published in User Modeling and User-Adapted Interaction. The survey's authors reviewed 123 papers on the subject and analysed 54 in depth. Their summary of the field is that recommendation algorithms were supposed to surface the long tail and mostly do the opposite.

This matters for a platform that hosts millions of creators and tells all of them to find a niche.

The advice industry treats discovery as one thing. The research treats it as at least two, and they behave differently.

Search, the ability to look for something specific and find it, reduces concentration. When people can express what they want and get it, demand spreads across a wider range of options. This is the mechanism behind the original long tail argument in information systems research, where lower search costs produce a less concentrated distribution of sales.

Recommendation, an algorithm deciding what to show you, increases concentration, at least in aggregate. Individual users see more variety. But different users get pushed toward the same items, so the catalogue as a whole gets used less evenly. The survey traces this to a popularity bias in the underlying algorithms: models trained on skewed interaction data learn that suggesting popular items scores well, and the resulting exposure makes those items more popular still.

A laboratory experiment on consumer search and sales diversity found the split cleanly. With recommendations present, participants looked at a wider range of products and bought a narrower one.

Attention diversified. Purchasing did not.

OnlyFans has no recommendation feed, which by this reading is not a loss. It also has no working search, no category browsing and no filters, which by the same reading is the entire problem.

Its internal search resolves usernames. If you know who you are looking for, it finds them. If you know what you are looking for and not who, it has nothing to offer, and no amount of the creator positioning themselves in a niche will change that, because the niche is not addressable. Nobody can type it.

This is what makes "find your niche" a strange piece of advice in this specific market. In a market with search, the advice is sound: differentiate, and the people who want that thing will locate you. In a market with recommendation, the advice is weaker but not useless, because algorithms do at least surface some tail items. In a market with neither, occupying a niche means becoming harder to find rather than easier, since the only functioning discovery route is a name that somebody already has.

Because the platform declined to build search, the search layer got built outside it.

Independent indexes compile public profile information into something filterable, by category, by niche, by price, by whether a page charges anything at all. Someone who knows the sort of thing they want but not a single username can look them up here and arrive with a name in hand, which is the only input the platform actually accepts.

This is the search-type infrastructure, not the recommendation type. It responds to a query rather than predicting one. On the research above, that is the version associated with a wider distribution of demand rather than a narrower one, although nobody has run the experiment on this market, and that distinction matters more than it might look.

The survey is unusually candid about the limits of its own field, and two of its observations belong in any article that leans on this literature.

First, the evidence is overwhelmingly computational. Studies involving actual users or live field tests are, in the authors' assessment, very rare, most of the work runs offline experiments on historical datasets using abstract metrics. The Lee and Hosanagar study stands out precisely because it was a real deployment. A body of research built mainly on simulations can drift from the situations it claims to describe, and the survey names that risk directly.

Second, the field's default assumption, that surfacing long-tail items is good in itself, is mostly unexamined. There are situations where recommending popular items is the right call, including for new users about whom nothing is known yet, and the survey notes that a deliberate dose of unpersonalised popularity is standard practice in industry, Netflix included.

Neither point rescues "find your niche" as advice. Both are reasons to be careful about the opposite claim, that better discovery infrastructure would automatically redistribute income toward smaller creators. It might redistribute attention. Whether attention converts is a separate question, and the honest answer is that nobody has measured it here.

Whatever discovery infrastructure does to the shape of a market, its absence does something simpler: it leaves the market working on names.

Seventy percent of the research the survey covers was published in the last five years, across 65 different outlets, which is a lot of attention paid to how algorithms distribute visibility. Almost none of it addresses a platform that distributes none at all.

The advice will keep being given. Pick a niche, differentiate, own a corner of the market. It is reasonable advice in a market with a search box that works.