How does the TikTok algorithm work? We measured 3.2 million videos (2026)

The TikTok algorithm ranks each video for each viewer from what people do with it: what they watch to the end, like, share, comment on or skip. TikTok has said this in writing since 2020. What it has never said is how much the things a creator controls move the result. So we measured it. We took 3.2 million TikTok videos from 49,357 creators and compared every video with the same creator's other videos from the same month. Video length mattered most. Videos of 3 to 10 minutes got 1.54x the views of the same creator's 10 to 20 second videos, and photo posts got 0.81x. Captions, hashtags and sounds came next. Posting time and day barely mattered.

The pages that rank for this question restate TikTok's 2020 post and add tips. Some give signals "in order of weight" or say shares count "3x more than likes", with no sample or method. None of them measures views. This study does, and it holds the creator fixed, so a big account and a small one are judged on the same scale.

What TikTok says about the For You feed

TikTok has explained its recommender in two places: a 2020 newsroom post and a Help Center page it keeps updated. The main points:

• Recommendations come from three groups of signals. User interactions (likes, shares, comments, follows, what you watch in full or skip), content information (sounds, hashtags, captions, the number of views, the country where it was posted) and user information (language, country, device).
• Strong signals count more than weak ones. TikTok's example: whether a viewer finishes a longer video counts more than whether viewer and creator are in the same country.
• Follower count and past hits are "not direct factors", though an account with more followers is likely to get more views.
• The feed avoids two videos in a row with the same sound or by the same creator, and content the viewer has seen.
• The weight of each factor "can change over time".

One more rule shapes what creators make. Since 2024, TikTok's Creator Rewards Program, which runs in eight countries (the US, the UK, Germany, France, Japan, South Korea, Brazil and Mexico), pays only for videos at least one minute long, and not for duets or stitches. TikTok does not say the feed favors those videos. We checked whether the data shows it.

Ten things a creator controls, ranked

We put ten factors into one model: video length, caption length, hashtag count, @mentions, the kind of sound, how many creators used that sound, the hours since the creator's last video, the local hour, the day of week, and whether the caption asks a question. Two rarer markers ran in the same model: replies to a comment and duet or stitch tags. Each factor is measured against the same creator's videos from the same month, with the other factors held fixed.

Video length explains about six times more of the gap between a creator's videos than caption length, the next factor. Hour and day, the most discussed advice, are last.

Longer videos get more views, with no jump at one minute

The longer the video, the more views it gets, and the rise is smooth. A 55 to 60 second video sits 3.4% above the creator's 10 to 20 second videos. A 60 to 65 second video sits 3.7% above. So the one-minute line, which decides Creator Rewards pay, does not show up as a jump in the feed. In the eight Creator Rewards countries we see a step of about 2 points at 60 seconds; in the rest of the world, none.

The rise keeps going. Videos of 90 seconds to 3 minutes got 1.25x, and 3 to 10 minutes 1.54x. This fits what TikTok wrote: a viewer who finishes a longer video sends a stronger signal. It also fits the money. Creators in program countries post more long videos (33% of their videos run over a minute, against 21% elsewhere), but their long videos gain less (1.07x against 1.12x). The premium is not a reward for joining the program.

Two other lengths stand out. Videos under 10 seconds got 1.13x. They loop, and replays can add views. Photo posts, the swipeable carousels, got 0.81x. For accounts with 500,000+ followers a photo post got about 0.70x the views of a video, a 30% drop (chart below).

Not every creator gains from length. Of the 10,622 creators with 15 or more videos on each side of one minute, 65% got more views on their longer videos. The median creator got 1.21x. The other 35% did better short.

Captions and hashtags still help a little

TikTok says captions and hashtags help it understand a video. Our data agrees, with small effects. A video with no caption got 0.88x. Captions of 101 to 200 characters got 1.10x, and 500+ characters 1.18x, against a 26 to 50 character caption. Longer captions give TikTok search more words to match, and they often belong to tutorials, recipes and stories that people watch to the end.

Three to five hashtags got 1.11x against none. Six to ten got 1.14x, but only 0.9% of videos use that many. Our hashtag study found the same on engagement: TikTok videos with 4 or 5 tags got 1.24x the likes, comments and shares of the creator's untagged videos. Generic tags like #fyp did nothing there.

A question mark in the caption came with 1.05x. An @mention came with 0.96x.

Sounds: your own audio beats a song

Videos with the creator's own audio did best. Videos set to a named track got 0.88x. Many guides say trending sounds help. In our data, holding the type fixed, a sound used by 100 or more of our creators got 1.05x against a sound only one creator used. So a popular sound helps a little, but a song in the background costs more than that. One reason: TikTok says the feed avoids two videos in a row with the same sound, so a sound that thousands of videos use has many competitors.

Replies to comments do better, duets and stitches worse

Videos whose caption starts with "Replying to @..." (TikTok adds this when a creator answers a comment with a video) got 1.12x, on 21,675 videos by 4,603 creators. The person who asked, and others who saw the comment, have a reason to watch.

Videos tagged #duet or #stitch got 0.80x, on 3,970 videos. We detect these from the caption only, and creators remove the tag on many, so the sample is small. Duets and stitches also do not earn Creator Rewards.

Posting time, day and gap barely move views

Across the eight 3-hour windows, views ranged from 0.99x to 1.01x of the 9 AM to 12 PM window. Across the days of the week, 1.00x to 1.02x. Our best time to post on TikTok study found the same: 6 to 9 AM wins, by about 2% in views.

Time since the last video made little difference within a month. Posting less than an hour after the previous video came with 0.98x. A gap of 7+ days came with 1.01x. If we compare across months instead (creator fixed, month free), videos after a 7+ day break got 1.15x. That bigger number reflects which months a creator posts rarely in, not the break itself.

The first 24 hours

Our crawler reads some videos more than once, so we can see how views arrive. At 24 hours the median video had 64% of the views it had 5 to 10 days later. The ranking hardly changes after day one: the rank correlation between views at 24 hours and the later count was 0.97.

Big videos are the exception. Videos that ended with 1 million or more views had a median of 46% of their views at 24 hours and grew 2.2x after. Videos that ended under 10,000 had 71% and grew 1.4x. This matches TikTok's note that the number of views is part of what it reads about a video. A video that does well keeps being shown to new groups of viewers for days. A video that stalls on day one rarely comes back.

By follower tier and market

Small creators gain most from length and hashtags. For a mega creator, hashtags add 1%. Photo posts cost every tier, and big accounts the most.

The direction is the same in every market. Length matters least in Japan and South Korea, where short clips are the norm, and most in the Philippines and Turkey.

Three creators, three answers on length

@maxgrabel, music quizzes, 3.4M followers. His videos of a minute or longer had a median of 547,300 views, against 115,200 for his shorter ones. "The Best Songs From Every World Cup", 96 seconds of song clips with the question "do you know all of these?", reached 14.1 million views in our crawl. It had 862,500 likes and 10,700 comments when we checked it live. The format needs time: each extra song is one more reason to stay to the end.

@julietsiegel, outfits and cosplay, 3.3M followers. Her longer outfit try-ons ("Your Fairy Outfit v2", 1 minute 41 seconds, own sound) had a median of 48,600 views. Her short clips had 6,100. Length is not the whole story: that fairy outfit video reached 2.3 million views, and so did a 23-second cosplay clip set to a song.

@thesotyfam, family, 3.9M followers. The opposite. Short family clips had a median of 241,500 views and longer ones 54,000. One long run worked: a countdown to a son's return from a mission ended with "Guess who's home!" (68 seconds), which had 1.9 million views in our crawl and 192,500 likes when we checked. The long videos that worked had a story people waited for.

What our data does not see

The ten factors together explain about 1% of the variation between one creator's videos. A typical video lands about 3.3x above or below its creator's average. Almost all of that swing comes from what viewers do with the video: how long they watch, whether they finish, share, or swipe away. TikTok sees those signals. We do not. So read the factors above as small edges on top of the content, not as a way around it.

What to do with this

For creators

• Test longer videos if your format has a reason to run long: a story, a list, a tutorial. Compare your own long and short medians after 15 or more of each, the way we did for the three creators.
• Do not cut a 70-second video to 58 seconds for the feed. There is no step at one minute. In the eight program countries, a minute is still the line for Creator Rewards pay; see how much TikTok pays per 1,000 views.
• Use your own audio where you can. Write a caption with real words, and add 3 to 5 tags that name the subject.
• Answer comments with a video.
• Judge a video at 24 hours. It has about two thirds of its views by then.
• Stop moving your posting time around. It is worth about 2%.

For brands

• Ask for the creator's median views on videos of the length you are buying. A creator whose long videos do worse is a different buy for a 90-second demo.
• Prefer videos over photo carousels for reach. Photo posts got 0.70x the views of a video at large accounts.
• Read campaign videos at 24 hours and again at day 7. Videos that will be big are still growing at 24 hours.
• Sponsored videos are a separate case. In our sponsored post study, sponsored TikTok videos got 0.69x the creator's usual views.
• Virev creator search shows each creator's recent videos with views, and the free TikTok engagement rate calculator reads a public account's last 12 videos. Our other free tools are on the Virev tools page.

For normal engagement rates by tier, see our TikTok engagement rate benchmarks.

How we measured, and what can be wrong

• Creators: public TikTok accounts with 1,000+ followers read by Virev's crawler from June 2026. We removed brands, media and organizations by Virev's profile label, accounts we could not place in a time zone, and accounts that post mostly at night (likely the wrong zone). This is the same set as our best time to post study.
• Videos: posted at least 3 days and at most a year before the read, without ad tags or giveaways, at least 10 videos per creator and 5 per creator-month. The first video of each creator has no gap and is left out. 3,219,734 videos remain.
• Model: log of views, with a separate baseline for each creator and month, so growth over time and account size drop out. Ten factors as categories, one regression, standard errors clustered by creator. A multiplier is the exponent of the coefficient: 1.10x means about 10% more views than the reference video by the same creator in the same month.
• Video length is TikTok's stored duration. A length of 0 marks a photo post; we checked two by hand on TikTok and both were photo carousels.
• Sound: "own original sound" means TikTok's original sound label (in any language) with the creator's name or handle as the author. Creators with a display name that differs from their sound author name are counted under "another account's original sound", which makes that group look a little better than it is. Sound popularity counts creators in our data, not on all of TikTok.
• Replies, duets and stitches come from the caption. We read 20 matches by hand; the reply captions were all comment replies, a few #stitch tags were plain hashtags.
• Early views: Virev's crawler read 4,599 videos at least twice, first within 48 hours of posting and last 5 to 10 days after. Videos we reread are not a random sample; they come from accounts we refresh more often.
• This is an observational study. Creators choose what to make long, which captions to write and which sound to use. A 3-minute recipe and a 10-second joke differ in more ways than length, and the model cannot separate them. It shows which videos get more views, not why.

Download the data

The aggregate tables behind every chart are in one CSV: tiktok-algorithm-study-2026.csv. It has the model multipliers with 95% intervals, the factor ranking, the length curve in 5-second steps, the tier and market tables and the early views curve. Please cite "Virev, October 2026" and link this page.