What TikTok itself says about recommendations
In its Newsroom, TikTok names three groups of signals for the For You feed: viewer actions, information about the video, and device and account settings. The platform explicitly gives the last group less weight.
Two conclusions follow from the same post. Subscriber count is not a direct ranking factor, and a past hit gives no head start to the next video: every publication is judged anew. TikTok calls completion of a long video a strong signal of interest.
- Viewer actions: completions, rewatches, shares, comments, likes, follows after watching.
- Information about the video: caption, sound, hashtags. The system uses them to determine the topic and the first audience.
- Settings: language, country, device type.
Level 1. Account and device
If reach dropped for all publications at once and that coincided with a change in how you work, start with the account, not with the videos.
A quick test: publish one file from an old, live account and from a new one. The difference in the first 24 hours shows where the problem is, in the file or in the account.
- Publications go out from one device and one network: mass publishing from a single IP links accounts into a group.
- Accounts are fresh, with no activity beyond posting.
- Descriptions and hashtags repeat across different publications. A text duplicate like that is visible without analysing any video.
- The device, the network and the interface language changed all at once.
Level 2. First audience and first second
The feed first shows the video to a small group of viewers. For that group to turn out right, the platform needs hints: caption, sound, hashtags, the first frame. If they contradict each other, the video goes to the wrong people and gets a weak reaction.
- Is it clear from the first second, without sound or captions, what the video is about.
- Do the topic, the caption and the hashtags match.
- Is the target audience interested in the topic: what competitors in the niche are pulling in.
- Does the video avoid opening on a logo, an intro card or an empty frame, the most common reason viewers leave in the first seconds.
Level 3. Retention and length
Platforms weigh retention above everything else. A typical mistake is a 45-second video where the value runs out at eight: viewers leave, and the algorithm reads that as not interesting.
- Cut the video to where the value ends, not to a 30 seconds is the format rule.
- Drop the wind-up: the first frame has to carry meaning.
- Build the ending so that it loops back to the start: a rewatch is a strong signal.
- Test different lengths on one topic, 8, 15 and 30 seconds, and compare completion.
Level 4. Duplicates and reposts
TikTok's community rules as amended on 24 September 2026 classify non-original content and unauthorised reposts as unacceptable, and set separate requirements for the For You feed.
In March 2026 Meta said it reduces the reach of non-original posts in the feed and Facebook Reels and favours creators of original content.
In practice this means: a video with another platform's watermark will not make recommendations; the same file across ten accounts reads as one group; renaming a file and editing metadata changes nothing. What really shifts a copy's fingerprint is covered with measurements in the piece on what platforms detect in copies.
Level 5. Moderation and restrictions
TikTok does not use the term shadow ban. What people call that looks like cut-off distribution with no notifications, and most often it is an ordinary dip rather than a sanction.
What to do: check notifications and account status, remove questionable material, stop publishing in batches for a couple of weeks and watch the trend.
- Reach dropped for all new publications at once, harmless ones included, which looks like a restriction.
- Reach declines gradually and older videos keep working, which means the issue is the content.
- There is a violation notice or a feature restriction. That is a fact, not a guess.
A checklist for an hour
Seven steps that separate a problem with the video from a problem with the account.
- Open notifications and account status: are there any violations.
- Compare the first three seconds of the best and the worst video of the month.
- Check the caption, sound and hashtags against the topic.
- Check whether texts repeat across different accounts.
- Look for watermarks and traces of other platforms on the copies.
- Match impressions against completion in analytics.
- Post a test video on the same topic from a live account.
When it is not about the videos
One video proves nothing: impressions are distributed probabilistically, and the difference between works and does not work only shows on a batch of 20-30 videos of the same format. Mass posting is built on exactly that volume: reach adds up from many tests.
It also happens that videos come out steadily but do not sell: the topic is entertainment, the product in frame is random. Then the problem is the brief, not the reach, and other formats are needed, which the article on 9:16 vertical video ads covers. If every scene matters, teams choose handcrafted videos, for example in the style of 1990s cartoons: from $190 per video up to 90 seconds.
When one topic goes to different geos and different platforms, reach is assembled through different publishing channels rather than edits to one account; copies for platforms are prepared by large-scale video versioning.