How the TikTok Algorithm Works

Published: 20 August 2026 · Reading time: about 8 minutes

The For You Page is a recommendation system, and recommendation systems are not mysterious. They measure what people do and predict what they will do next. TikTok has published a reasonable amount about its own, and the rest can be inferred from the analytics it shows creators. Here is what it actually measures, in rough order of how much it matters.

The one-sentence version

TikTok shows your video to a small group of people, watches how they respond, and either expands the audience or stops. Everything else is detail.

The test batch

When you post, your video goes to a modest initial audience — some of your followers, plus users whose interests match the video's apparent subject. Creators often describe this as a fixed number like 300 views. It is not fixed, and it varies with your account history and the topic, but the shape is right: the first audience is small.

That group's behaviour is measured. If the signals are good, the video goes to a larger group. If those respond well too, it goes wider again. Each round is a fresh test, which is why a video can sit at 400 views for two days and then suddenly accelerate — it cleared a threshold late.

This is also why the follower count matters less than people expect. Followers improve your starting position slightly. They do not carry a video that the test audience ignores.

The signals, ranked by weight

1. Watch time and completion rate

How much of your video people watch is the dominant signal, and completion rate — the share of viewers who reach the end — is the version of it you can most directly influence.

The logic is straightforward. TikTok's business depends on keeping people in the app. A video that holds attention does that job; a video that gets skipped does not. Everything else is secondary to this.

Two practical consequences. First, shorter is usually safer: it is much easier to get 80% completion on twenty seconds than on ninety. Second, rewatches count for more than their share of the runtime suggests — a seven-second video watched three times registers as 300% watched, which is why short looping videos punch above their weight.

2. Retention shape, not just the average

The system looks at where people leave. A video that loses half its audience in the first two seconds is judged differently from one that holds everyone until a slow ending, even if the average watch time is identical.

This is why the opening second gets so much attention from experienced creators. It is not superstition — it is the steepest part of the curve. Your analytics show this graph directly; see TikTok analytics explained.

3. Shares

Shares are the strongest engagement signal because they are the most expensive action. Liking costs a tap. Sending a video to a friend costs social capital — you are vouching for it.

Content that gets shared tends to be useful, surprising, or precisely targeted at a relationship ("this is so you"). If you want one engagement metric to design for, it is this one.

4. Comments

Comments indicate the video provoked a reaction strong enough to type about. Replying to comments helps twice: it extends the conversation, and your replies can be posted as videos, which turns one piece of content into several.

5. Likes and saves

Likes are the weakest of the engagement signals — cheap to give, and given reflexively. Saves are more interesting: a save means "I want this again later", which is a strong quality indicator for tutorials, recipes and reference content.

6. Follows from the video

Someone watching a video and following the account as a result is a powerful signal, because it says the viewer wants more of specifically this.

7. Negative signals

These work in reverse and are worth taking seriously: skipping quickly, tapping "not interested", hiding the video, reporting it, or muting the account. A video can be actively suppressed rather than merely ignored.

8. Content and context signals

Lower weight than behaviour, but still used: captions and on-screen text, hashtags, the sound, the subjects detected in the video itself, device language and region. These mainly help the system decide who the first test audience should be. They do not rescue a video that audience does not watch.

Myths worth dropping

• "Posting at exactly 7:03pm matters." Timing has a modest effect on your test batch and no effect on whether a video is good. It is the most over-discussed variable on the platform. Our timing guide puts it in proportion.
• "Deleting a flop hurts your account." There is no evidence for a lasting penalty. A video that underperformed simply stops being distributed.
• "You get shadowbanned for using certain words." Community-guideline violations do restrict distribution, and that is a documented, visible process — you are notified. The folk version, where invisible word filters silently kill accounts, is mostly creators explaining an ordinary slow week.
• "Thirty hashtags will get you seen." Hashtag stuffing dilutes the topical signal and looks like spam. Three to five specific tags is the whole technique.
• "Follower count determines reach." It determines a small part of your starting audience. Nothing more.
• "Engagement pods work." Coordinated engagement from accounts with no topical relationship to your content teaches the system to show your videos to the wrong people.

Niches and how the system finds your audience

TikTok is effectively building two profiles: one of each viewer's interests, and one of what each video is about. Distribution is a matching problem between them.

This is the real argument for picking a niche. An account that posts consistently about one subject becomes easy to classify, and the system can find the right audience reliably. An account that posts a recipe, then a gym video, then a comedy sketch is hard to place, and each video has to find its audience from scratch.

It also explains a common frustration: one video goes viral with a general audience, thousands of people follow, and subsequent videos underperform. The new followers were attracted by an outlier, not by the account's normal subject. A viral video outside your niche can actively hurt you.

What to actually do

1. Design for completion. Before posting, ask what happens at second two, and whether anything in the middle can be cut. The answer is usually yes.
2. Open with the payoff. Show the finished thing, state the claim, ask the question. Save the build-up for people who have already decided to stay.
3. Post consistently in one subject. Consistency of topic matters more than frequency.
4. Read retention, not views. Views tell you what happened; the retention graph tells you why.
5. Treat every video as an experiment. Change one thing at a time — the hook, the length, the format — and keep what works.

The uncomfortable summary is that there is no algorithm trick. The system is a fairly honest measurement of whether people want to watch what you made. Most "algorithm hacks" are attempts to avoid that conclusion.