TikTok has published its own explanation of how the feed works, in a 2020 Newsroom post titled "How TikTok recommends videos #ForYou". This article walks through what that explanation actually says, then maps each part onto the general recommender system ideas a developer would recognise: implicit feedback, feature weighting, cold start and exploration.

The three groups of signals

TikTok says the system ranks videos "based on a combination of factors", in three groups:

In recommender terms, the first group is behavioural feedback, the second is item metadata, and the third is context. None of this is unusual. What is interesting is how TikTok says the groups are weighted against each other.

Strong signals beat weak signals

TikTok's post gives one concrete example. Whether a user finishes watching a longer video is a strong indicator of interest, and it "would receive greater weight than a weak indicator", such as the viewer and the creator being in the same country.

This is the core idea behind implicit feedback. An explicit rating (five stars, thumbs up) is rare and noisy. Watch behaviour is plentiful and honest. Every swipe is a data point: you either stayed or you left, and if you stayed, you either finished or you did not. A short-video app produces dozens of these judgements per minute of use, which is why the feed can adapt so quickly.

For anyone building a recommender, the lesson is that completion and dwell time are often more informative than the actions people consciously take.

What is not a direct factor

The most quoted line from TikTok's post is this one: "neither follower count nor whether the account has had previous high-performing videos are direct factors" in the recommendation system.

That is a deliberate design choice. Many social feeds historically leaned on the social graph, so large accounts got distribution because they were large. TikTok says it ranks the video, not the account. A creator with a big following may still get more views, because followers see their content, but follower count itself is not an input to the For You ranking.

In practice this is what lets a brand-new account reach a large audience with its first video. Each video is evaluated on how people respond to it. For a creator-side look at how these signals play out today, see this guide to the TikTok algorithm in 2026.

Diversity is built in

A recommender that only optimises for predicted engagement tends to collapse into a narrow loop. TikTok's post describes several rules that push against that:

• The For You feed generally will not show two videos in a row made with the same sound or by the same creator.

• The system does not recommend duplicated content, content you have already seen, or content it considers spam.

• It sometimes inserts videos that do not match your existing interests, to help you discover new creators and categories.

The last point is the classic exploration and exploitation trade-off. Exploitation serves what the model already believes you like. Exploration spends a small share of the feed on uncertain items, so the model can learn whether its picture of you is incomplete. Without exploration, a recommender cannot discover that your interests have changed.

The cold start problem

Every recommender has to handle new users with no history. TikTok's answer, according to the post, is to ask new users to pick interest categories when they first open the app. Users who skip that step start with a general feed of popular videos, and the system narrows down from there as the first likes, comments and replays come in.

TikTok describes those first interactions as starting "an early round of recommendations". That matches how most cold start strategies work: show broadly popular items, collect a handful of strong signals, then personalise.

Content that is not eligible

The post also lists content that may not be recommended at all, even if it would perform well. TikTok says some reviewed content, such as graphic medical procedures, may not be shown to general audiences, and that newly uploaded videos, videos under review and spam can be ineligible.

In system design terms, this is a filtering layer that sits before or alongside ranking. A video can be technically eligible to exist on the platform while being excluded from algorithmic distribution.

How users can steer the feed

TikTok lists several controls that act as direct feedback to the model:

• Long-press a video and tap "Not Interested".

• Hide videos from a particular creator or that use a particular sound.

• Report a video.

• Follow accounts and explore hashtags, sounds, effects and trends on the Discover page.

Negative feedback matters as much as positive feedback. Without it, a recommender can only learn what you tolerate, not what you dislike.

What this means in practice

If you strip away the product details, TikTok's description reduces to a few principles that apply to almost any recommendation system:

1. Implicit behaviour is the richest signal. Watch completion tells the model more than a like.

2. Rank the item, not the author. Popularity of the creator is not the same as relevance of the content.

3. Weight signals by strength. Context such as country is useful, but weaker than behaviour.

4. Protect diversity on purpose. Repetition limits and exploration keep the feed from collapsing.

5. Make negative feedback easy. Users should be able to say no without effort.

A note on sources

Everything above about TikTok's own system comes from TikTok's Newsroom post "How TikTok recommends videos #ForYou", published in June 2020. TikTok has not published its model architecture, and the system has certainly changed since then, so treat this as the company's stated design principles rather than a technical specification. The recommender concepts (implicit feedback, cold start, exploration and exploitation) are general terms from the field, not claims about TikTok's internal implementation.

Source: How TikTok recommends videos #ForYou, TikTok Newsroom