For twenty years, the dominant model in adult entertainment has been the catalogue. A studio or an amateur shoots a scene, the scene gets uploaded, the scene gets tagged, and the user goes looking for it. Everything downstream of that — the tube sites, the aggregators, the affiliate networks, the tag taxonomies, the recommendation feeds — is infrastructure built to solve one problem: helping a person find, among millions of pre-existing files, the one that most closely resembles what they had in mind.
Nobody ever solved that problem. They just got better at approximating it.
The search problem nobody solved
Think about what actually happens in a typical session on a tube site. A user arrives with something specific in their head — a look, a scenario, a dynamic, a mood. What they get is a grid of thumbnails filtered by a dozen or so blunt categorical tags. They click, scrub, close, click again, scrub again. The industry has a metric for this and it is not flattering: the median session involves opening several videos and finishing none of them. Every one of those abandoned clicks is a gap between what the user wanted and what the catalogue could offer.
This is not a UX failure. It is a structural limit. A catalogue can only ever return what already exists, and what already exists was produced for someone else. The long tail of human preference is effectively infinite; the library is finite and expensive to extend. Tagging, semantic search, and recommendation engines all attack the same bottleneck from the same side — better retrieval over a fixed corpus — and all of them hit the same ceiling.
From retrieval to generation
Generative models remove the corpus constraint. When the content is produced at request time, “search” stops being the interaction model. The user describes what they want and receives exactly that, rather than the nearest available approximation someone else happened to film in 2019.
Platforms such as Porn.ai are built around that inversion. There is no library to browse because there is nothing to browse until you ask for it. The unit of consumption shifts from the scene that exists to the scene you specified, and the difference in satisfaction between those two things is much larger than most people in this industry have internalised. Anyone who has watched a first-time user go from typing a prompt to seeing their own idea rendered back at them recognises the reaction — it is the same one people had the first time a search engine actually understood what they meant.
The second-order effect matters more than the first. Once generation is instant, iteration becomes free. The user does not get one result and settle; they adjust, regenerate, refine. The session becomes a loop rather than a funnel, and loops are, for obvious reasons, extremely good for retention.
The economics are not close
The traditional pipeline carries fixed costs that scale poorly: talent, crew, location, licensing, compliance, editing, storage, and the ongoing legal overhead of every one of those. Extending the catalogue by one scene costs real money. Extending it by ten thousand scenes costs ten thousand times as much.
For a generative platform, the marginal cost of an additional piece of content is compute, and compute has been falling in price for the entire history of the industry. Inventory becomes effectively unbounded. Catalogue depth stops being a moat, because catalogue depth stops being a scarce resource — which is precisely why the incumbents whose entire competitive position rests on library size should be paying close attention.
The objections worth taking seriously
Some of the pushback here is fair, and the sites pretending otherwise are not helping the category.
Quality is genuinely uneven. Video generation in particular is in the awkward phase where the good outputs are very good and the bad outputs are unusable, and users notice. That gap is closing on a timescale measured in months, not years, but it is real today.
Provenance and consent are the harder problem, and the one that will define which platforms survive the next regulatory cycle. Likeness misuse is not a hypothetical risk; it is an active enforcement priority in several jurisdictions. The platforms that will still be operating in three years are the ones investing in identity filtering, provenance signals, and age-verification infrastructure now, while it is still a differentiator rather than a legal requirement.
And there is a genuine open question about whether an infinitely customisable experience is one people actually want long-term, or whether the friction of the catalogue was doing something for us that we did not notice. That one will not be settled by argument.
Where this lands
Tube sites are not going to disappear. Radio survived television. But the centre of gravity moves, and it moves toward whichever model resolves the user’s intent most directly. For two decades that was “find me a file.” Increasingly it is “make me the thing I am imagining.”
The interesting question for anyone operating in adult right now is not whether generation replaces retrieval. It is how quickly, and whether your traffic model assumes a user who is still willing to scroll.
