How the TikTok Algorithm Works in 2026

Almost every explanation of the TikTok algorithm makes the same mistake: it describes a system that ranks videos. TikTok does not rank videos. It ranks video-viewer pairs — it is constantly predicting whether this specific person, with this specific viewing history, at this specific moment, will watch the next clip to the end.

That distinction sounds academic and it is not. It explains why a video can flop with your followers and then find fifty thousand strangers a week later, why identical content performs differently on two accounts, and why most “beat the algorithm” advice is really just advice about retention wearing a costume.

The test-batch model

Every upload starts by being shown to a small audience — some mix of your followers and people whose behaviour suggests they might like this kind of thing. TikTok watches what that batch does. If the completion rate, replays and engagement clear the bar for that content category, the video is released to a larger and less closely-matched audience, and the test repeats. If it does not clear the bar, distribution quietly stops.

Two things follow from this that matter more than anything else in this article.

First, your video is judged against its own category, not against the platform. A three-minute explainer is not competing on completion rate with a six-second loop. The bar moves. This is why creators who switch format often see a confusing dip — they are being measured against a different benchmark than the one they had learned to beat.

Second, the sample size in the first test is small enough to be noisy. That is the honest, uncomfortable part. A genuinely good video can be shown to a few hundred people who happened to be scrolling at 2am, underperform on a statistical fluke, and never recover. Consistency works not because the algorithm rewards loyalty but because more uploads means more rolls of the dice.

The signals, roughly in order of weight

Watch time and completion — the heavyweight

TikTok is optimising for time spent in the app. Everything else is a proxy for that. A video watched to the end tells the system it successfully held a person, and a video watched more than once tells it something stronger still. Completion rate matters more than raw seconds, which is why a fifteen-second video finished by 80% of viewers can out-travel a two-minute one finished by 12%.

The practical read: your hook is not a stylistic choice, it is the primary ranking input. The first second decides whether the rest of the video gets to be measured at all.

Replays and rewinds

A loop or a rewind is watch time you did not have to shoot. This is why the loop-back edit became ubiquitous and why dense information — a recipe, a list, a text-heavy tip — often outperforms its production value. The viewer scrubs back because they missed something, and the system records it as sustained attention.

Shares

Sharing costs the viewer something. They have to decide the video is worth a specific person's time and then spend social credit sending it. That expense is exactly why the signal is weighted heavily — and why shares tend to correlate with the videos that break out rather than merely perform.

Comments

Comments do two jobs. They are an engagement signal in their own right, and they generate sessions: people return to a thread, read replies, argue. A comment section that is alive keeps the post alive after the initial test window closes.

Saves

The bookmark is the highest-intent tap on the platform because it is a promise to come back. It also predicts replays days later, which is the only reliable way a video gets a second life after its algorithmic moment has passed. Saves cluster around useful content — tutorials, lists, recipes, anything actionable.

Likes

Real, and the weakest of the set. A like is reflexive, costs nothing, and is easily given by someone who was about to scroll away anyway. It is a genuine signal — just not the one that decides distribution, which is worth knowing before you optimise for it.

Relevance and matching

Captions, on-screen text, spoken audio, hashtags and the sound you used all help TikTok work out who this video is for. This is less a ranking factor than a routing factor: it does not make the video travel further, it makes the early tests land in front of people more likely to finish it — which then makes it travel further. Increasingly it also feeds in-app search, which is a growing share of how content gets discovered.

What almost certainly does not work the way you were told

• Follower count does not gate reach. A video from a small account and one from a large account enter roughly the same test. Large accounts do better on average because they have a bigger, warmer initial batch and more practice — not because the number confers distribution.
• Hashtag volume is not a lever. Twenty hashtags do not open twenty doors. They are a weak routing hint; a precise caption usually does more.
• There is no shadowban switch. The overwhelmingly common cause of a “shadowban” is that recent videos failed their test batches. It feels like punishment because the drop is abrupt, and abrupt is exactly what a threshold-based system produces.
• Deleting an underperforming video does not reset anything. It removes a post that could still be picked up later, which is a cost with no offsetting benefit.
• Posting time matters less than the advice industry suggests. It nudges who is in the first batch, which matters at the margin. It does not override retention. If you want to reason about it properly rather than copy a chart, our best time to post tool works from your own audience pattern instead of a generic table.

The cold-start problem, and why it is a real disadvantage

Everything above describes a system that is fair in aggregate and brutal in the particular. A new account has no history, so its first test batches are assembled from weak signals — TikTok genuinely does not know who your content is for yet. That means a wider, worse-matched audience, which means a lower completion rate, which means the video stalls for reasons that have nothing to do with its quality.

The honest fix is volume: post enough that the system builds a model of your audience. Most accounts that “took off suddenly” were feeding it data for weeks beforehand.

Where paid engagement fits is narrower than most sites selling it will admit. It does not change your completion rate and it will not make a weak hook work. What a nudge of TikTok views does is stop a post from dying in silence before anyone has had a chance to react to it — and a profile that reads as established, with a follower base that does not make a first-time visitor bounce, converts more of the attention you do earn. Both are real, both are limited, and neither substitutes for the first two seconds of your video.

What to actually change this week

• Cut everything before the point of interest. If your video has a setup, the setup is the problem.
• Design for completion rather than length. Decide the finish line first, then cut backwards to it.
• Give the loop a reason to close — end on the thing that makes the opening make sense.
• Write the caption for a person, not for a crawler. A question that someone would actually answer beats a wall of tags.
• Post enough to get out of small-sample noise. Judging your strategy on three videos is judging a coin flip.
• Watch retention graphs, not view counts. The graph tells you where you lost people; the count only tells you that you did.

None of this is about outsmarting a ranking system. It is about making something a stranger will finish — which happens to be the only thing the ranking system is measuring.