RN
A recommender does not know what you like. It knows what you did — which thumbnail you clicked, how long you stayed, what you skipped nine seconds in, what you came back to. Taste is the thing it infers from that, and the inference is only ever as good as the record.
Which puts the interesting question somewhere other than where the marketing puts it. The question is not whether a feed can be tuned to you. It obviously can. The question is what has to be stored, and by whom, for that tuning to happen at all.
Two ways to guess, and both are in use
Content-based matching works from the properties of the material. Tags, categories, performers, studio, duration, production style. If the last six things you finished share four tags, the system surfaces more things carrying those tags. It is simple, it explains itself, and it is what most tube-site sidebars are doing.
Its weakness is that it can only work with metadata that exists. Adult tagging is uploader-supplied, wildly inconsistent, and heavily gamed for search traffic — a video tagged with fifteen popular categories it does not belong to poisons every recommendation built on those tags. Garbage metadata produces a confidently wrong feed.
Collaborative filtering ignores the content entirely and works from other people. It finds users whose behaviour resembles yours and recommends what they watched that you have not. This catches things no tag would connect, which is why it feels uncanny when it lands.
Its weakness is that it needs a crowd and it needs you to be legible within it. New users get nothing useful — the cold-start problem — and genuinely unusual taste gets pulled toward whatever the nearest large cluster is watching. If your interests sit at the edge of the distribution, collaborative filtering will quietly try to move you toward the middle.
Most systems of any sophistication run both and blend the outputs, which is why a feed can be simultaneously eerily accurate about one thing and stubbornly wrong about another.
What the system is actually reading
Almost none of it is what you tell it. Explicit signals — likes, favourites, ratings — are sparse, because people rarely bother. The heavy lifting is done by implicit ones:
- Watch time as a share of total length, which is the strongest single indicator most systems have
- Where you stopped, and whether you scrubbed backwards first
- Repeat views
- What you scrolled past without clicking, which is a negative signal and is recorded as one
- Search terms, including the ones that returned nothing
- Time of day and session length, which cluster more strongly than people expect
The last two are worth sitting with. Search queries are the most explicit statement of intent a person makes on these sites, and they are logged as text. Session timing is not obviously personal until you notice that it is enough, by itself, to distinguish one household member from another.
The failure mode nobody warns you about
Recommenders optimise for engagement, and engagement is not the same as satisfaction. The system cannot tell the difference between watching something because it was good and watching it because you were bored and it kept autoplaying.
That produces a well-documented feedback loop. The feed shows more of what you engaged with; engaging with it confirms the inference; the range narrows. After enough cycles you are being served a caricature of your own history, and the sensation people report is not being understood but being stuck — a feed that has decided who you are based on one evening six months ago and will not update.
The escape is unglamorous: use search deliberately, browse by category rather than by feed, and treat anything that offers a history reset as a maintenance tool rather than a privacy one. Systems that let you delete watch history usually let you do it precisely because the recommendation quality depends on the history being accurate.
The privacy part, stated properly
Personalisation requires persistent identity. There is no version of it that does not. To connect this session to your previous ones, something must link them — an account, a cookie, a device fingerprint, or an IP address stable enough to serve as a proxy for one.
This is where the common reassurance breaks down. "Anonymised" and "not linked to your name" are not the same claim, and only the second one is usually true. A profile with no name attached that records every video you watched, every search you ran, and when you were online is not anonymous in any sense that matters — it is pseudonymous, and pseudonymous browsing profiles are routinely re-identifiable when combined with other data.
Three specifics that matter more than the general reassurance:
Adult sites carry third-party code like every other commercial site. Ad networks, analytics, and affiliate trackers sit on the same page as the recommender. Anything the page can see, they can potentially see, and their retention policies are theirs rather than the site's.
Under GDPR, data about sex life and sexual orientation is a special category with a higher bar than ordinary personal data — Article 9 prohibits processing it outright unless a specific condition applies, of which explicit consent is the usual one. A viewing history on an adult site is exactly that kind of data. This is a genuinely strong protection where it applies, and it applies to how a site treats users in scope of the regulation, not to everyone on the internet.
Private browsing does nothing to server-side profiling if you are signed in. It clears local traces. The account-side record is unaffected. If you want the profile not to exist, the lever is the account, not the browser window.
If the trade is not one you want, the honest options are the boring ones: browse signed out and accept a worse feed, keep a separate browser profile, or put a VPN in front of it and understand that this addresses network-level observation rather than anything you hand the site directly.
Where generative systems change the shape
Recommendation picks from things that exist. The newer AI generation tools make something to order instead, which removes the matching problem entirely — and replaces it with a sharper version of the same privacy question, because a prompt is a far more explicit statement of what you want than a click ever was, and it arrives at the provider as legible text.
The consent questions around generated material are a separate and more serious matter than anything in this article, and they do not have a technical answer.
What to take from it
Personalisation is a real improvement over a homepage that shows everyone the same twelve videos. It is also a bargain with an unstated price, and the price is a durable record of your attention held by a company you did not evaluate.
Worth it for plenty of people. But make it a decision rather than a default, and be sceptical of anyone describing the trade as though there were nothing on the other side of it.