How TikTok decides what to show

User interactions. What you watch to the end and what you swipe away from, what you like, share and comment on, which accounts you follow, and what you post yourself. This is the group that carries nearly all the weight, because it is the only one that says anything about what a person actually wanted.

Video information. The caption, the sound, the hashtags and the on-screen text — everything the system can read about a video before anyone has watched it. This is how a brand-new upload gets its first small test audience: there is no interaction history yet, so the text is all there is to go on.

Device and account settings. Language, country, device type. TikTok says outright that these are weighted lowest, because they are chosen once for practical reasons and say almost nothing about interest.

TikTok's own illustration of weighting is that a strong indicator of interest, such as whether a viewer finishes a longer video from beginning to end, counts for more than a weak one, such as the viewer and the creator being in the same country.

The reason is what each action costs the viewer. A like is a reflex that takes a fraction of a second and is handed out to things people did not particularly care about. Watching ninety seconds to the end costs ninety seconds, and almost nobody pays that for a video they did not want.

It follows that the two numbers most worth reading in your own analytics are average watch time and the share of viewers who watched the whole thing. Likes are the number most creators look at first and the one that says least.

This is the published line most often disbelieved. TikTok states that neither follower count nor whether an account has had previous high-performing videos are direct factors in recommendation. Each video is tested on its own.

It is why a new account can have one video reach twenty thousand people and the next reach two hundred, and why a large account can post a video that dies. The system is not asking "is this creator popular"; it is asking "did the people we tested this on respond".

It is also why bought followers move nothing. A follower who never watches is not an interaction, and the follow itself does not enter the recommendation decision for the next video. On TikTok, a follower who does not watch is worse than no follower — which is the reason PodSwap built no follow exchange for it.

Every recommender builds a picture of who a video suits out of who responded to it, and uses that picture to choose the next audience. So attention has two effects: the immediate number, and the durable update to what the system believes your audience is.

The second effect is why attention from the wrong people is worse than none. A hundred viewers with no interest in your subject who watch three seconds and leave teach the system, accurately, that the video does not hold attention — and assemble a picture of your audience out of people who were never in it.

TikTok also states that it deliberately intersperses content outside your expressed interests, so a viewer occasionally sees something the model would not have predicted. That is where a video from an unfamiliar creator gets its chance, and it is a small window: a viewer who leaves in two seconds has used it up.

If who responds is what the system learns from, then the useful question about any engagement service is not "how many" but "who". Attention from creators in your follower bracket and your subject teaches it something true about your video. Attention from a farm teaches it something false.

It is also the honest limit. Attention that arrives through a link is not the same thing as attention that arrives through the For You feed, and whether the first moves distribution the way the second does is genuinely unknown — every source in this category assumes it and none demonstrates it. That is why the TikTok side promises matched distribution, early feedback and measurement, and nothing downstream of them.