How the TikTok algorithm works: what TikTok and ByteDance actually published

What I read: TikTok’s explainers, and a ByteDance paper that never names TikTok

The public record comes in two kinds. Plain-language explainers on TikTok’s own sites. And an engineering paper from ByteDance, the company behind TikTok, that many write-ups treat as “the TikTok algorithm”.

• TikTok Newsroom, “How TikTok recommends videos #ForYou”, 18 June 2020. An official statement about the live system, naming the main factors and several feed rules. It’s six years old.
• TikTok Transparency Center, “Introduction to the TikTok recommendation system”. Undated; I read it on 21 September 2026. The most detailed official walk-through, from picking candidate videos to the final rules.
• TikTok’s support site, “How TikTok recommends content”. Undated; I read it on 21 September 2026 as well. Official factor lists for each part of the app, For You included.
• Two shorter official posts, from December 2021 (varying recommendations) and December 2022 (the “Why this video” panel), and the March 2023 version of the For You feed eligibility standards in TikTok’s Community Guidelines.
• TikTok Newsroom, 23 January 2026. The new U.S. joint venture “will retrain, test, and update the content recommendation algorithm on U.S. user data”. confirmed · 9
• Liu and ten co-authors (2022), “Monolith: Real Time Recommendation System With Collisionless Embedding Table”. The authors, all ByteDance staff or interns, presented it at a workshop of the ACM RecSys conference in September 2022. It describes a production system: “Monolith has successfully landed in the BytePlus Recommend product.” confirmed · 4 ByteDance has since put the code on GitHub; its project page doesn’t mention TikTok, and the repository was archived in October 2025. confirmed · 5

None of these sources contains TikTok’s model code or a single weight value. In 2020 TikTok said invited experts would be able to review its source code at its Transparency Center in Los Angeles. confirmed · 1 That was an offer to invited visitors, not a public release.

TikTok says the feed predicts your reactions, ranks by them, then adds variety

1The video is checked and described. Only videos that pass TikTok’s initial content moderation become candidates. confirmed · 2 TikTok enforces its rules with a mix of machine learning models and human review. confirmed · 2 In 2020 it also said videos just uploaded, or still under review, may be ineligible for anyone’s For You feed. confirmed · 1 What TikTok says it uses about the video is descriptive detail: captions, sounds and hashtags in 2020, and today posting time and region, the author’s language setting, the soundtrack, length and hashtags. confirmed · 1, 2 Both lists are examples. Neither says whether models read what’s shown or said in the video to decide who sees it.

2A big pool of candidates is pulled. The system retrieves “a large number of videos” eligible for For You, without saying how many or how they’re chosen. confirmed · 2 The same page also says it uses other people’s behavior: if someone engaged with two of the same videos you did, the system may predict you’ll engage with a third video that person interacted with. confirmed · 2 Your follow relationships are an input too. confirmed · 2

3Each candidate gets a score from predicted reactions. For every video, the system estimates how likely you are to like, share or comment on it; mark it “Not interested”; follow its author or interact with their profile; finish, skip or favorite it; spend a certain amount of time on it; or tap its sound. confirmed · 2 It combines those predictions into one overall score and sorts videos from high to low. confirmed · 2

How the predictions are combined isn’t published. In 2020 TikTok said strong signals count more than weak ones: finishing a longer video from beginning to end counts more than viewer and creator being in the same country. confirmed · 1 Device and account settings count less, it said, because people don’t actively express them as preferences. confirmed · 1 Today it says that for many users, time spent watching a video is generally weighted more heavily than other factors, and that the weighting can change over time. confirmed · 3 Its current For You list also includes a video’s number of views and the country it was published in. confirmed · 3

ByteDance’s Monolith paper describes one unnamed production model built from several “towers”, each predicting a different kind of user behavior. confirmed · 4 That fits TikTok’s description, but the paper doesn’t say this model serves TikTok. inferred

4Final checks add variety and apply rules. If top-ranked videos are too similar, say because they use the same sound, the system swaps some out for others. confirmed · 2 In 2020 TikTok said the feed generally won’t show two videos in a row with the same sound or by the same creator, and doesn’t recommend duplicated content, videos you’ve already seen, or spam. confirmed · 1 Rules then make sure the feed includes creators from your region and leaves out content from accounts aged under 16. confirmed · 2 TikTok also deliberately mixes in videos that don’t match your stated interests or haven’t gathered many likes, and in 2021 it was testing ways to avoid long runs of similar videos on topics such as extreme dieting, sadness or breakups. confirmed · 1, 6

5The feed learns from what happens next. TikTok says every new interaction helps the system learn your interests, and calls its signals “dynamic and constantly updated”, without saying how fast. confirmed · 1, 2 Monolith is where ByteDance documents speed: user actions such as clicks and likes stream back into training while the live model keeps serving. confirmed · 4 What the model stores about each individual user, item or other ID was refreshed at minute-level intervals; the rest of the model, daily. confirmed · 4 In a seven-day live test on a ByteDance ads model, this real-time training beat periodic batch retraining on all seven days, by the paper’s accuracy measure. confirmed · 4

Monolith’s other main idea is memory: instead of letting different users or items share a slot, which blurs them together, it gives each one its own. confirmed · 4 It also skips IDs seen only a handful of times, and expires inactive ones, such as “a short-video that is out-of-date”. confirmed · 4

The paper never discusses the For You feed’s goals, signals or weights; its tests used public movie-rating and ad-click data, plus unnamed ByteDance models. confirmed · 4 Because ByteDance built Monolith for fast-feedback products and names short-video ranking as an example, TikTok’s systems likely learn quickly too — but no source confirms it. inferred

My read: watching to the end leads, and each video stands alone

None of these is a published TikTok rule. Each is my own deduction, and I name its basis.

• Make people want to stay to the end. TikTok says time spent watching generally weighs most for many users, and calls finishing a longer video a strong sign of interest — which suggests watching is your main scoreboard, and extra length helps only if people stay. inferred
• Don’t win the first second and lose the viewer. Skips and “Not interested” are among the reactions it predicts, and TikTok says it adjusts for what people mark as not interesting — so a hook the video doesn’t pay off is likely to work against you. inferred
• Judge each video on its own. Because TikTok said follower count and past hits aren’t direct factors, a small account isn’t shut out by its size, a past hit doesn’t carry the next video, and a flop is more likely about that video than a mark against your account. inferred
• Label the video for the people you want. TikTok lists captions, sounds and hashtags as video information and lets viewers filter out words and hashtags, so accurate labels probably help the right people find it. inferred No published paper or official page says piling on popular hashtags adds reach. A trending sound isn’t a free slot either: the final check swaps out top-ranked videos that are too similar, like ones sharing a sound. inferred
• Stay eligible. TikTok’s 2023 eligibility standards list reposts without new or creative edits, very short clips, GIF-only videos and “like-for-like” bait as ineligible for For You. confirmed · 8 Those videos can still be found through search or by following. confirmed · 8 If a video only gets views from followers and search, check those standards before you assume a penalty. inferred
• Read results across several posts. New uploads may be held back while under review, TikTok deliberately mixes in videos outside people’s known interests, and the weighting can change over time. confirmed · 1, 3 So one video’s first hour is a noisy reading; compare several posts before you change course. inferred

Nobody outside TikTok knows the weights, me included

I watch videos all day, and I still can’t see TikTok’s weights. The published record leaves out most of what you’d want to know:

• The weights. TikTok says the weighting can change, but gives no numbers. confirmed · 1, 3
• How candidates are found. The Transparency Center says only “a large number of videos”. confirmed · 2
• Whether a video’s pictures and speech are analyzed to decide who sees it. No published paper or official page confirms this for the For You feed.
• What the U.S. retraining changes. TikTok said in January 2026 that the algorithm would be retrained on U.S. user data, but not how ranking will differ. confirmed · 9

Four common beliefs, checked against the sources:

• “Big accounts get pushed.” Contradicted as a direct factor: in 2020 TikTok said follower count and past hits are not direct factors, while admitting bigger accounts are likely to get more views. confirmed · 1 Popularity still plays a part: the current For You list includes a video’s number of views, and the “Why this video” panel can cite content popular in your region. confirmed · 3, 7 Today’s pages neither repeat nor withdraw the 2020 follower statement.
• “Every video is tested on a small batch first.” Half true. TikTok ranks each viewer’s feed by how likely that viewer is to watch, finish, share or skip each video. confirmed · 2 So a new video mostly reaches first the viewers it’s predicted to suit, not a random batch, and how they react shapes who sees it next. inferred TikTok also deliberately shows people videos outside their known interests, partly to learn what’s popular among a wider range of audiences. confirmed · 1 What no published paper or official page gives is a set size, a fixed window or a pass mark (see the myths page).
• “Rewatches are the strongest signal.” No evidence. TikTok’s pages talk about finishing, skipping and time spent watching, not rewatching, and their clearest ranking is that time spent watching generally weighs most for many users. confirmed · 1, 2, 3 A replay may simply add to time spent, but no source says how replays are treated. speculated
• “The TikTok algorithm is Monolith, and it updates every minute.” Not confirmed. The paper documents minute-level updates for unnamed ByteDance production models, and names only BytePlus Recommend as a product. confirmed · 4 That TikTok uses the same or similar technology is plausible, but unconfirmed. inferred

Dates matter. The most specific statements are from 2020 and 2021. The current explainers are undated. Monolith is from 2022. TikTok announced a U.S. retraining in January 2026. confirmed · 9 Any of this may have changed.