Why a single watch can reshape a feed
A recommendation system does not know why you watched something. It sees an event: you opened a video, stayed for some portion of it, perhaps searched for a related topic, and then moved on. Those signals become clues about what might hold your attention next. A one-off choice can therefore look, to the system, like the beginning of a new interest.
How the clue spreads
Many recommenders compare patterns among viewers and items. If people who watch a geology documentary often go on to watch videos about fossils, volcanoes, or landscape formation, the system may treat those as plausible next suggestions for someone who watched that documentary too. It does not need to conclude that you are a geologist; it only needs to predict that a related video might earn another click.
That prediction can reinforce itself. If the system places geology videos in prominent spots, you may watch one out of curiosity—or simply because it is there. That new viewing becomes another signal, making the topic seem more relevant than the original, accidental choice really was.
Why the odd recommendation can linger
Platforms balance recent activity with longer-running patterns, and a single signal may remain useful to their models until newer behavior outweighs it. Different services handle that balance differently, but the basic problem is the same: an algorithm can observe what you did, not whether you want that choice to define your future feed.
The result is a feedback loop, not a permanent label. If you repeatedly choose other subjects, those choices can give the system stronger clues. But until they do, one documentary about rocks may keep competing with the romantic comedies you actually prefer.
