TikTok Algorithm: For You Page, Monolith and Social Search

The TikTok algorithm decides in milliseconds which video plays next. It does so without regard for who anyone follows: the For You Page pulls candidates from the entire video inventory and sorts them by predicted response. Followers give you a small advantage when a new video starts, because the system can make a more reliable prediction for viewers it already knows. After that, what counts is how strangers react.

That makes TikTok the most consistent implementation of interest-based distribution. How the generic pipeline of candidate generation, light ranking, heavy ranking and re-ranking is built is covered in the article on the recommendation algorithm. This piece is about what sets TikTok apart from other feeds: the evidence on ranking signals, the app's double role as feed and search engine, and the architecture ByteDance has documented itself.

Monolith: what ByteDance has disclosed about its recommendation architecture

ByteDance disclosed the architecture of its recommender system with Monolith in 2022. Three of its building blocks matter in practice.

Collision-free embedding table: Classic systems hash user and item IDs onto a fixed table size. Two different videos then land in the same memory slot and share learning information, which dilutes the prediction for both. Monolith puts a cuckoo hash map underneath that resolves collisions instead of tolerating them. Every video and every user keeps its own representation, even across billions of IDs.

Worker and parameter server architecture: Training and serving run separately, but in sync. The training instance updates parameters, the serving instance receives them at short intervals. That decouples compute load from live operation without letting the serving model go stale.

Online training against concept drift: Interest distributions shift within hours, not weeks. A model trained overnight in a batch is systematically outdated by morning. Monolith keeps training during live operation and treats drift as the normal case, not as an error.

One qualification belongs here: the paper names neither TikTok nor the For You Page, but a different ByteDance product. It describes the company's architectural principles, not verifiably the live system behind the app. It still matches what practitioners see. A video is not scored once and then filed away, it is reclassified again and again. That is why a video can suddenly pick up after days without reach. And that is why reach collapses mid run as soon as a new group of viewers responds differently than predicted.

Which TikTok algorithm ranking signals are documented

In its own explanation of the For You Page, TikTok describes that a longer video watched all the way through is weighted more heavily than a weak interest signal; the interest signals it names there are likes, shares, comments and follows. That text dates from 2020, and the Community Guidelines confirm the signal groups again in 2025, adding search queries. Which of the named signals TikTok classifies as the weak one is stated nowhere. The company has never published concrete weights.

That gives you a clean split between what you can treat as fact and what is industry folklore.

The test pool story is persistent: first 500 views, then the next stage, then viral. The only documented part of it is that new videos are first tested on small audiences. No TikTok source names fixed thresholds. Where such stories come from and which of them hold up empirically is covered in the article on shadowbans, cold start and myths.

Read traffic sources instead of counting views

TikTok Analytics reports per video where the views came from: For You, search, following and the personal profile. That breakdown is the most reliable diagnostic the platform gives you, and it is almost always ignored because the total view count is louder. Four patterns are worth evaluating.

• A high For You share with low watch time: the video was tested broadly and switched off again. The problem sits in the first few seconds or in the length, not in distribution.
• A high search share: the video answers a recurring question and ages slowly. Videos like this deserve successors on the same topic.
• A high share from Following with a minimal For You share: the cold start never got past your own follower base, the model found no audience outside it.
• A conspicuous profile share: viewers enter through your profile, which then becomes a ranking surface in its own right.

These four cases lead to completely different measures. A bare view count leads to none. On top of that, the traffic split of a video does not stay stable. A video can run almost exclusively through the For You Page at first and be found mainly through search weeks later. Judging performance as final on day one confuses the cold start with the result. As a rule of thumb from experience, a second look after a few days and a third after a few weeks pays off, to separate evergreen from short-lived.

Treat TikTok as a search engine

TikTok reported in 2025 that 1 in 4 users starts a search within the first 30 seconds of opening the app and that the platform records billions of search queries per day, up 40% year on year. According to the eMarketer Social Search Report 2025, almost 40% of TikTok users search several times a day and 73% at least once a day; those are US figures, and comparable numbers are not publicly available for the DACH region.

For research, TikTok introduced Creator Search Insights in 2024. The tool shows frequently searched topics, filterable by category, and flags content gap topics: search queries with high demand and few videos to match. You reach it through the search bar and through the creator tools in TikTok Studio.

The rest is search engine logic. The keyword has to be spoken, appear in the on-screen text and show up in the caption, because TikTok transcribes audio and recognises text. Generic hashtag sets are no substitute. Whether a query suits a video at all or rather a website is decided by search intent: comparison questions and how-to questions work in video, transactional queries rarely do. How search differs across TikTok, Instagram and LinkedIn is covered in the article on social SEO.

The originality rule from September 2025

TikTok published the rule on 14 August 2025; since 13 September 2025, unoriginal or reused content without an own contribution has been excluded from the For You feed. What counts as an own contribution is not set out in detail in the guidelines. In our assessment it covers everything that gets uploaded a second time without a new edit, without your own commentary and without new context.

There is no explicit watermark rule in the guidelines. A foreign platform logo in the frame is still the clearest marker of reused material, and as a rule of thumb: anyone repurposing Reels or Shorts on TikTok exports without the logo and re-edits natively. That costs a few minutes per video and removes the riskiest category.

TikTok in the DACH region: reach and limits

According to GWI data from the second quarter of 2025, published via DataReportal, 34.0% of German internet users aged 16 and over use TikTok monthly. That puts the platform behind WhatsApp, Facebook, Instagram and YouTube. Daily usage time reverses the picture: according to Agorapulse and ARD/ZDF data from 2025, TikTok sits at around 75 minutes per day, well ahead of Instagram and Facebook. Among Gen Z in Germany, TikTok reaches around 69% according to the Statista Global Consumer Survey 2025. As of August 2026, no more recent usage rate based on the same methodology is publicly available.

For DACH B2B that means: TikTok is rarely the channel where a purchase decision inside the buying centre is prepared. For recruiting, employer branding and products that need explaining to a broad audience, the combination of long dwell time and working search is hard to replace. The decision should hang on whether your audience searches there, not on whether it scrolls there.

Common mistakes

• Cross-posting without a new edit: unchanged material without an own contribution has been the most direct route out of the For You feed since September 2025; a foreign platform logo also makes it easy to spot.
• Optimising for views instead of retention: views are a result, retention is the lever; without a look at the retention curve, every adjustment stays guesswork.
• Deleting and re-uploading: that destroys the learning data the system has already gathered on the video, and starts the cold start again without a better video.
• An intro that runs too long: if the retention curve already crashes at the start, that is not an algorithm problem.
• Hashtag rituals: a full hashtag set is no substitute for a spoken keyword.

One mistake sits a level higher and costs the most time: trying to reconstruct weights TikTok has never published. Trade articles circulate percentage values for individual signals, factors for shares against likes and supposed thresholds for completion rates. None of it comes from a primary source. What is documented are the signal groups and the statement that a strong interest signal counts for more than a weak one. There are no numbers attached to it. Anyone building content rules on values like these optimises against a guess and never notices, because swings in reach on TikTok can always be explained somehow.

How you measure progress

Three metrics are enough to steer by. The retention curve shows where viewers drop out and is the only metric that points directly at a change in the edit. The share of For You traffic shows whether a video got past your own follower base at all. The search share over time shows whether your content turns evergreen or dies with the trend.

Follower count is not among them. In a system that tests every video afresh against the entire pool, it is a lagging indicator. If you want to understand the TikTok algorithm, measure retention and traffic sources; everything else is interpretation.