Popularity is partly manufactured by the feed
When people say a song, a video or an opinion is “everywhere”, they usually mean it keeps appearing in the feeds they use. That is a statement about exposure, and exposure on most large platforms is decided by software rather than by a simple count of what people want. Understanding how that software works, at least in outline, is one of the most useful habits for reading any trend story built on online activity.
The details of any specific platform’s system are generally not public, and they change often. What follows describes the general design ideas that most ranking systems share, with the caution that real systems vary.
What a recommendation system is trying to do
A feed has far more possible items than a person could ever look at. Something has to decide which few appear first. In outline, many systems work in stages:
- Candidate selection. From a very large pool, the system picks a smaller set that might be relevant, often based on who you follow, what similar users engaged with, or what is new.
- Scoring. Each candidate gets a predicted score, typically an estimate of how likely you are to do something the platform values, such as watching to the end, replying or sharing.
- Ranking and adjustment. Items are ordered by score, then often adjusted for variety, freshness, safety rules or business goals.
The key point is that the target is a prediction about behaviour, not a judgement about quality or importance. If the system is tuned to reward watch time, content that holds attention will tend to rise. If it is tuned to reward replies, content that provokes responses may rise. The goal chosen by the platform quietly shapes what the public sees.
From prediction to apparent popularity
Popularity numbers such as views, likes and shares are produced inside this process. An item that the system decides to show to more people will usually collect more of each. That creates a loop:
- The system shows an item to a small test audience.
- If the early response looks good, it shows the item to a larger audience.
- More exposure produces more visible engagement, which can lead to still more exposure.
This kind of feedback can make small early differences grow. It resembles the dynamics discussed under network effects and tipping points, where a modest initial advantage can compound. It also means that two equally good items can end up with very different counts, partly because of timing, the first audience they happened to reach or random variation.
Because the loop runs on predicted engagement, what the platform shows is also what generates its own evidence. The numbers then look like confirmation that the item was popular, when part of the popularity may be the result of being shown.
Why this matters for reading data
Several common reasoning problems follow from this design.
Counts reflect exposure as well as appeal. A video with many views was shown many times. Without knowing how often it was shown to people who scrolled past, views alone say little about how strongly anyone liked it. The distinction between kinds of signals is explored in the discussion of engagement metrics.
Different people see different platforms. Personalised feeds mean there is no single shared view of what is popular. Your impression of what “everyone” is watching is built from your own feed, which is itself a selection. This is a form of the selection effects problem: the sample you see is not chosen to be representative.
Design changes can look like behaviour changes. If a platform changes its ranking goals or its interface, the volume of some content can move sharply even if people’s tastes have not. This is treated separately under platform rule changes, but the underlying point belongs here: a chart of platform activity partly records the platform.
Cause is hard to separate from amplification. It is tempting to say a post went viral because of its qualities. Often the more cautious reading is that it had some qualities, met a favourable early audience and was boosted by the system. Telling these apart is a problem of causal claims, and outside researchers usually lack the access needed to test it directly.
What can and cannot be known from outside
Researchers and journalists often work with what is visible: public counts, scraped samples, or data supplied by platforms. Each has limits, which is the broader theme of platform data and its blind spots. Typically:
- Public counts show outcomes, not how many times an item was displayed.
- Samples gathered by scraping may be skewed toward what is easy to collect.
- Platform-supplied data may be limited in scope or chosen by the platform.
- Experiments run by platforms themselves are rarely described in full detail.
So a careful reader treats claims about what “algorithms do” with some humility. Studies that have tried to measure effects of ranking often find that results depend on the platform, the time period and the group of users, which is one reason the evidence is described as mixed on questions such as whether feeds narrow people’s information diets.
Some other effects worth keeping in mind
Ranking systems are not the only force. People choose whom to follow, what to search for and what to share, and these choices interact with the software. Journalists then report on what appears to be rising, and that coverage can itself push more attention toward it, a pattern described as media amplification. Creators also respond to what appears to be rewarded, changing what they make, which feeds back into what the system has available to rank. The creator economy is shaped in part by guesses about what the system favours.
Short-form video is a useful example. Feeds there often put heavy weight on how users respond within seconds, and the sequence is chosen largely by the system rather than by the viewer’s explicit choices. That can make the link between what people say they want and what they are shown looser than on a feed built mainly from accounts they follow.
A practical way to read a popularity claim
When you meet a claim that something is popular online, a few questions help:
- Is the measure exposure (views, impressions) or action (shares, purchases)?
- Who was it shown to, and could the platform have chosen that audience?
- Does the claim cover the whole population, or only the users of one service?
- Did anything change in the platform recently that could explain a jump?
- Is there evidence from outside the platform, such as surveys or sales records?
None of these questions proves a claim wrong. They simply separate what the feed displayed from what people independently chose, which is the gap this whole cluster of topics keeps returning to.