How recommendation feeds decide what you see

YouTube and TikTok say their recommendation feeds rank videos using individual behavior signals, watch time, clicks, completions, survey responses and content information like captions and sounds, filtered through account settings such as language and country. Those descriptions come from the platforms' own published blogs and help pages, and they are the only authoritative account of how the feeds work, version-dated and subject to change.

What does YouTube say influences its recommendations?

In its own explanation of how the search and discovery system works, published on YouTube's official blog and help documentation, the company lists the signals it uses to rank suggested videos: the viewer's watch history, how long people watch a given video, clicks, likes, shares, survey responses and the frequency with which the viewer has already seen a channel or topic. YouTube explicitly says no single creator or employee decides what trends, and that each viewer's home page and Up Next row are personalized, not universal.

Two points in YouTube's own account run against popular belief. First, the company says it down-weights videos viewers click on and then quickly abandon, because satisfying watch time, not raw clickbait, predicts retention. Second, YouTube publishes an annual reminder that its recommendations are evaluated against user surveys about satisfaction, which is why a video watched heavily but rated poorly can lose suggested-video traffic. Both statements appear in company explanations current through 2025.

What does TikTok say about the For You feed?

TikTok's blog post on recommending videos, first published in 2019 and still the company's canonical description years later, breaks its signals into three groups. User interactions include which videos a person likes, shares, comments on, favorites, creates and watches to the end. Video information covers captions, sounds and hashtags. Device and account settings cover language preference, country setting and device type, which TikTok describes as one-factor signals used to optimize performance rather than strong personalization drivers.

TikTok also states that the feed weights signals by indicated interest: a video watched to the end counts more than a fast scroll past, and skipping or pressing not interested removes similar content. The company says follower count is not a direct ranking factor, which matches the observable pattern of accounts with small followings reaching large audiences through the For You page.

Why do the platforms publish these explanations at all?

Transparency here is largely regulatory and reputational work. YouTube has published recommendation-quality updates since at least 2019, when it said it would reduce recommending borderline content that violates no rule but misinforms, and TikTok's For You post sits inside a broader transparency push that included European Commission pledges under the EU code of practice on disinformation. Both companies describe their systems as continuously tuned, and both frame the published explanations as simplifications of ranking systems that are far more complex than any blog post can convey.

That caveat deserves emphasis. What the companies publish is the category of signals, not their weights, not the models, and not the experiment schedule. An engineer's tweak can shift behavior more than any documented principle, and none of that is visible from outside. The honest reading is that the official descriptions are necessary but not sufficient: they rule out some folklore, and they say nothing precise about magnitude.

What can creators legitimately infer?

From the companies' own words, a few practices follow directly. Since both platforms reward completion and satisfaction, holding viewers to the end matters more than inducing a click they regret. Since captions, sounds and hashtags are content signals on TikTok, accurate labeling helps the system place a video. Since YouTube tracks how often a viewer has already seen a channel, fatigue effects are real and documented by the platform itself.

What does not follow is the folklore: posting-time rituals, hashtag stuffing, engagement pods. None of these appear in either company's description, and several run against the stated satisfaction goal. Creators who test such tactics and report results are reporting noise, correlation with the actual signals, or effects of changes the platforms never announced.

How often do the documented rules change?

The published descriptions are stable in shape but not in detail. YouTube's blog has documented successive changes, the 2019 borderline-content reduction, periodic updates on harmful misinformation handling, and annual statements on recommendation satisfaction surveys. TikTok's account of the For You feed has stayed essentially the same since 2019 in its three-group structure, while the product around it, video length defaults, commerce integration, search features, has changed repeatedly. As of 2025, the descriptions summarized here remained the companies' official account, subject to change without notice, which is the caveat both companies attach themselves.

What about Instagram's feed?

Meta's own explanations of the Instagram feed follow the same pattern of named signal categories without weights. The company's published overview ranks its signals by importance, with watched-time and activity history near the top, followed by likes, relationship signals and post information, and it describes ranking in stages: candidate sourcing from followed and suggested accounts, then prediction of likely actions, then ordering. Meta has also published transparency-style cards for its ranking systems across Facebook and Instagram as part of its response to regulatory pressure, mirroring the EU commitments noted above. As with the other platforms, the company frames these descriptions as simplifications of systems under continuous experimentation, and as of 2025 none of the three companies had published model weights or training details. The correct reading across all three: the documented signals set the boundaries of what matters, while the undocumented tuning decides outcomes, which is why creator folklore fills the gap and why official explanations deserve citation when it is cited at all.

Everything here reflects the companies' published accounts as of 2025, subject to change without notice.