← Writing

What TikTok has actually said about its algorithm

26 July 2026 · Joonas Soininen

TikTok publishes far less about its recommendation system than Instagram does. The advice industry filled the gap anyway. Here is the short list of what TikTok has said itself, and what the confident numbers turn out to be.

The one line that contradicts most advice

From TikTok's own explainer, published June 2020 and still their primary statement on the subject:

"Neither follower count nor whether the account has had previous high-performing videos are direct factors in the recommendation system."

Read that against the standard advice to grow your following so the algorithm favours you. TikTok says it does not work that way. Every video is assessed on its own.

This is also the practical case for how you should read competitors. A big account is not getting reach because it is big, so copying big accounts teaches you less than you would think. What you want is the video that beat its own account's typical performance, because that difference is about the video rather than the audience.

The signals TikTok names

Three groups, in their words:

They describe weighting rather than a ranked list: "a strong indicator of interest, such as whether a user finishes watching a longer video from beginning to end, would receive greater weight than a weak indicator, such as whether the video's viewer and creator are both in the same country."

So completion is explicitly named as a strong signal. That much of the standard advice is right. What is not published is a ranked top three, of the kind Instagram's head gave for Reels in January 2025. Anyone presenting a definitive TikTok signal ranking is filling in a blank.

The numbers that do not survive checking

"TikTok evaluates 500+ behavioural signals." TikTok has never published a signal count. There is no primary source for 500, or for any other number. It appears in marketing posts citing other marketing posts.

"TikTok has over 2 billion monthly users." TikTok's most recent official global figure is 1 billion, from September 2021. It has not been updated since. The larger numbers come from third-party analytics firms whose estimates currently range from about 1.92 to 2.21 billion and disagree with each other, which tells you they are models rather than counts.

Where TikTok does publish reliable numbers, it is because the law requires it. Their EU transparency filing, mandated under the Digital Services Act, reports 178 million average monthly active recipients in the EU for July to December 2025. In January 2026 they stated more than 200 million Americans. Those are the citable figures.

"The algorithm tests your video on a small batch of users first, then expands." No primary source. TikTok's own description of the system contains no staged rollout, no test batch, and no numbers like "200 to 500 users in the first hour". It is a plausible-sounding story built to explain why a video underperformed.

What you can actually measure

TikTok Studio gives organic creators a retention rate, defined as the percentage of your video that users watched, plus a per-video drop-off graph. That graph is the most useful thing in there and the least discussed.

Be careful with thresholds quoted at you. TikTok's advertising products count 2-second views and 6-second focused views. Those are ad metrics. If someone gives you a target percentage for a metric with a number of seconds attached, check whether it is a metric you can even see on an organic post.

What the research says, which is less than you would hope

Two studies worth knowing, both using real viewing data donated by users rather than vendor surveys.

A CHI 2024 paper analysed 9.2 million recommendations from 347 real TikTok accounts and found average attention sat at roughly 45%, and did not improve as the algorithm learned more about the user. A CHI 2026 paper trained a model to predict whether someone stays or swipes, using millions of real decisions with the entire video available, and topped out near 70% accuracy. Removing the video's content features barely hurt it.

That second result is the honest ceiling on this whole genre. Most of what decides whether your video gets watched is not in your video. It is who is watching, what they just saw, and what the feed decided to do with you.

How to use any of this

Three things follow, and none of them are hacks.

Completion is named by TikTok as a strong signal, so the middle and end of your video matter at least as much as the opening. Follower count is explicitly not a factor, so study videos that beat their own account's average rather than videos from large accounts. And ignore any threshold that arrives without a source, because on current evidence most of them were invented by someone selling something.

Silma is AI content research for organic social. It ranks the competitors you track by how far each post beat that account's own median, not by raw views, and it reads your own draft frame by frame before you post. Free plan is 25 credits a month, no card. Start free.

Sources

Claims we could not source and therefore did not use: a 500-plus signal count, a global user figure above one billion, and the small-batch testing mechanic.