12 algorithm myths, checked against what the platforms actually published
Everyone has a theory about the feed. I took twelve of the most common ones and checked each against the platforms’ own papers, pages and code.
How to read my tags: confirmedthe source says it · inferredmy deduction from what's confirmed · speculateda guess, and I say so
- Your first audience isn’t random. Instagram says that, for recommendations, each eligible post first goes to a small audience it thinks will enjoy it, and the best performers go wider. confirmed · 1
- TikTok and X rank each viewer’s feed by predicted reactions, as YouTube’s 2016 paper did. confirmed · 2, 6, 17 So a new post mostly reaches first the people it’s predicted to suit, and how they watch decides what comes next. inferred
- Likes are one predicted action among many: TikTok says watch time generally weighs most for many users, and X weights a like at 0.5 against 5 for a reply. confirmed · 4, 17
- Shadowbans exist in a narrower, documented form, and Instagram’s Account Status and X’s Under the Hood pilot let you check your account. confirmed · 12, 17
- No platform publishes the size of that first audience or a golden hour. Anyone quoting one is guessing.
How a new post finds its first viewers
Most myths are about what happens right after you post, so I’ll start there.
“Every video is tested on a small random group first.”
There is a first audience, and on Instagram it’s official. Instagram says that, for recommendations, every eligible post is first shown to a small audience it thinks will enjoy it, whether or not they follow you; as that audience engages, the top-performing reels go to a slightly wider audience, then a wider one again. confirmed · 1 The wrong word is random. Even Instagram’s description is a prediction: an audience it thinks will enjoy the post. confirmed · 1 TikTok, YouTube and X don’t describe a test group, but they describe the same kind of prediction. TikTok ranks each viewer’s feed by how likely that viewer is to watch, finish, share or skip each video. confirmed · 2 In YouTube’s 2016 paper, each viewer is shown only the highest-scoring candidates. confirmed · 6 X’s current code lifts a small account’s new post only in feeds where it was already picked as a candidate and scores highest among eligible new posts. confirmed · 17 TikTok also shows people different videos from time to time, partly to learn what’s popular with wider audiences. confirmed · 3 So the first people to see your video are mostly the ones it’s predicted to suit, and the test is whether they actually watch it. inferred How big that first audience is, how long the test runs and what counts as a pass: no platform publishes any of it. “100 people” is a guess.
so Make the video for a specific audience, so the right people get picked. Then make them watch more and more of it. inferred On Instagram you can also run a version of the test yourself: trial reels go to non-followers first. confirmed · 14
“The first hour decides everything.”
No published paper or official page gives a window in which an ordinary post’s fate is decided. What is confirmed is that early reactions travel fast. Meta says its very large models can be updated within minutes, and that they refine their picture of each person after reactions such as watching a whole video. confirmed · 18 Instagram counts how quickly people engage with a post as a popularity signal, one it says counts much more in Explore than in Feed or Stories. confirmed · 12 X is the one platform with a published clock for its main feed: For You drops posts older than 48 hours, and its new-author lift only applies to posts under 48 hours old with fewer than 1,000 views on the Home timeline. confirmed · 17 A golden hour, as a rule, is folklore with a stopwatch.
so Post when your people are around, and judge a post by who reacted and how, not by its first-hour total. inferred
What the feed actually counts
“There’s one algorithm.”
Instagram says Feed, Stories, Explore, Reels and Search each rank posts their own way, and Meta’s engineers describe more than 1,000 models behind its recommendations. confirmed · 12, 16 YouTube says it built separate recommendation systems for Shorts and long videos. confirmed · 11 X’s published code chains separate services to find posts, score them and filter them. confirmed · 17 “The algorithm” is shorthand for a stack of systems, each tuned for its own part of the app. inferred
so Give each post one job. On Instagram, Stories only reach people who follow you, while the Reels tab is mostly accounts the viewer doesn’t follow. confirmed · 12
“Likes are what the algorithm wants.”
Likes count, but as one predicted action among many. confirmed · 2, 12, 17 TikTok says that for many users, time spent watching is generally weighted more heavily than other factors. confirmed · 4 Instagram’s creator FAQ says what matters most for distribution is engagement per viewer: watch time, sharing, following, liking, commenting and profile visits. confirmed · 13 X is the only one that prints its numbers: a predicted like is worth 0.5, a reply 5 and a share by copied link 20. confirmed · 17
so Build for the action each platform names: watching on TikTok and YouTube, full watches and reshares on Reels, replies and shares on X. inferred
“Longer videos win, because platforms count watch time.”
YouTube’s 2016 ranking was built on expected watch time per impression. confirmed · 6 On its own, that leans toward videos that keep people watching longer. inferred Today YouTube says there is no universal ideal length, names average percentage viewed next to view duration, and advises against filler. confirmed · 8, 9 In 2020 TikTok called finishing a longer video a strong sign of interest. confirmed · 3 Instagram doesn’t recommend reels over 3 minutes to people who don’t follow you. confirmed · 13 Length pays only when people stay. inferred
so Make it as long as the idea needs, then cut what people would skip. inferred
“One bad reaction costs more than a like earns.”
On X, chance for chance, yes. It’s the one platform that prints its weights. confirmed · 17 A predicted “not interested” counts −43.2, a mute −58.8, a block −31.2 and a report −234, against +0.5 for a like. confirmed · 17 X’s own code adds two cautions: the weights multiply one viewer’s predicted chances, not counts, so “one report cancels 468 likes” is the wrong reading, and a report is weighted heavily partly because it’s over 1,000 times rarer than a like. confirmed · 17 TikTok, Instagram and YouTube all use “not interested” as a signal but publish no sizes. confirmed · 2, 9, 12 Elsewhere, the exact trade is unknown.
so A hook that wins the tap and loses the viewer is a bad trade everywhere. inferred
Growth, and what nobody outside can see
“Big accounts get pushed, and small ones can’t break through.”
Size helps, mostly through your followers. inferred Instagram even named follower count itself as a Reels signal in 2023. confirmed · 12 In 2020 TikTok said follower count and past hits are not direct factors in For You, while admitting bigger accounts are likely to get more views. confirmed · 3 Instagram said in 2024 that recommended reels had been ranked mainly on how an account’s followers engaged, so the largest accounts often got the most reach, and that it was adding a ranking input to give smaller creators more distribution. confirmed · 1 X’s code can lift one new original post per refresh from an account with 1,000 or fewer followers, if the post has had fewer than 1,000 Home views. confirmed · 17 So the myth was closer to true on Instagram before its 2024 change. inferred
so If a video flops on a small account, it’s more likely about that video than about your size. inferred
“Posting more often gets rewarded.”
No source I read rewards volume. YouTube’s help page for Shorts says there’s no minimum posting cadence for videos to do well. confirmed · 7 X discounts each extra post from the same author in one refresh, never below a quarter of its score. confirmed · 17 TikTok said in 2020 that For You generally won’t show two videos in a row from the same creator. confirmed · 3 And YouTube says one video’s underperformance doesn’t penalize a channel overall, though a channel can suffer over time if particular viewers consistently stop watching its videos when they’re recommended. confirmed · 8
so Post when you have a video your audience will choose. A few they pick likely beat many they skip. inferred
“My reach dropped, so the algorithm changed.”
Maybe. Weights do move: TikTok says the weighting of its factors can change over time confirmed · 4, and X’s repository documents a reply boost it began testing on 10 July 2026, rolled out widely on 13 July and scaled back on 24 July. confirmed · 17 But YouTube names outside reasons too: how many people care about your topic, competition, and seasons such as major holidays. confirmed · 8, 9 From outside, you can’t tell which one hit you. inferred I watch videos all day, and I still can’t see what TikTok or Instagram changed last week. Nobody outside can.
so Compare several posts before you change course, and check the outside reasons first. inferred
Every source here is a snapshot with a date. TikTok’s clearest statements are from 2020, YouTube’s detailed paper is from 2016, and X’s code is from September 2026. When a source changes, a verdict can change with it.
Questions people ask
Does TikTok test videos on a small group first?
Is shadowbanning real?
Do hashtags still help?
Does posting more often help you grow?
Do my followers see my posts first?
Sources
- Instagram for Creators (2024). Helping Creators Find New Audiences. Instagram, 30 Apr 2024. creators.instagram.com · official page · official statementSmall-audience-first ranking of eligible content, originality and aggregator rules; announced as rolling out over the following months.
- TikTok (undated; read 21 September 2026). Introduction to the TikTok recommendation system. TikTok Transparency Center. tiktok.com · official page · official statementCurrent step-by-step description: candidate selection, predicted interactions, combined score, ranking, similarity check, rules, user controls. No weights, candidate counts or update speed.
- TikTok (2020). How TikTok recommends videos #ForYou. TikTok Newsroom, 18 June 2020. newsroom.tiktok.com · official page · official statementNames the three factor groups, the weighting principle (finishing a longer video is a strong signal; device and account settings weigh less), the follower-count statement and feed-variety rules. Six years old; no weights.
- TikTok (undated; read 21 September 2026). How TikTok recommends content. TikTok support site. support.tiktok.com · official page · official statementFactor lists for each surface (For You, Following, Friends, LIVE, Search, account suggestions); says time spent watching is generally weighted more heavily for many users and that weighting can change. No numbers.
- TikTok (2023). For You feed Eligibility Standards. TikTok Community Guidelines, last updated March 2023. tiktok.com · official page · official statementContent allowed on TikTok but kept out of For You, still findable by search or following. This is the March 2023 version; the Guidelines have been revised since and the newest version could not be fetched.
- Covington, P., Adams, J., Sargin, E. (2016). Deep Neural Networks for YouTube Recommendations. RecSys '16 (ACM Conference on Recommender Systems), Boston. Google. doi.org · paper · productionDescribes YouTube’s production recommender as of 2016 (candidate generation, ranking by expected watch time, example age); ten years old, and YouTube has changed since.
- YouTube Help (undated, read September 2026). Search & discovery tips (Shorts version). support.google.com · official page · official statementShorts ranking signals (share who chose to view, view duration, percentage viewed, likes, surveys), Shorts surfaces, posting cadence, Shorts and long-form; no weights.
- YouTube Help (undated, read September 2026). Good to know about recommendations for YouTube's recommendation system. support.google.com · official page · official statementYouTube’s answers to common creator questions (length, formats, channel penalties, titles and thumbnails, subscribers, monetization); YouTube’s own account, not independently testable.
- YouTube Help (undated, read September 2026). Search & discovery tips (Video version). support.google.com · official page · official statementHow long-form videos are personalized and judged once recommended (watch or ignore, view duration, percentage viewed, likes, surveys) and outside factors; no weights.
- Goodrow, C., VP of Engineering, YouTube (2021, September 15). On YouTube's recommendation system. YouTube Official Blog. blog.youtube · official page · official statementSignals (clicks, watch time, surveys, shares, likes, dislikes), history of objectives, borderline-content demotion; no model details or weights.
- The YouTube Team (2025, May 6). Shorts truths: Debunking common myths about YouTube Shorts. YouTube Official Blog. blog.youtube · official page · official statementStates that Shorts and long-form have separate recommendation systems and that YouTube has seen no evidence of Shorts hurting long-form; no technical detail.
- Mosseri, A. (2023). Instagram Ranking Explained. Instagram (about.instagram.com), 31 May 2023. about.instagram.com · official page · official statementSignals and predictions for Feed, Stories, Explore and Reels, plus recommendation rules; describes the live app in plain words, gives no weights, and is now three years old.
- Instagram for Creators. Frequently Asked Questions. Instagram, undated (read 21 Sep 2026). creators.instagram.com · official page · official statementEngagement per viewer, the 3-minute rule, hashtags and what limits distribution; undated, so check it for changes.
- Instagram for Creators (2024). Trial reels: Try content with non-followers first to see what performs best. Instagram, 10 Dec 2024. creators.instagram.com · official page · official statementHow trial reels are shown and measured; a product feature, not a ranking explanation.
- Instagram for Creators (2026). Rewarding original creators on Instagram. Instagram, 30 Apr 2026. creators.instagram.com · official page · official statementExtends the 2024 originality rules to photos and carousels and defines original content; does not change what followers see.
- Levis, L., Ma, S. S., Nava, E. (2025). Journey to 1000 models: Scaling Instagram's recommendation system. Engineering at Meta, 21 May 2025. engineering.fb.com · official page · productionInfrastructure post: more than 1,000 models, a ranking funnel per surface, weights adjusted in experiments; little on what is predicted.
- xAI / X (2026). x-algorithm: X For You Feed Algorithm. GitHub, xai-org/x-algorithm (README; home-mixer/params/param.rs, synced 18 September 2026; home-mixer/scorers; docs/BIDIRECTIONAL_BOOST_CHANGE.md; phoenix/README.md; user-cred-v2; visibility-filtering). github.com · open-source code · productionThe current For You code and default settings as of 18 September 2026, including the weights, the 48-hour limit and the new-author lift (home-mixer/scorers/author_cold_start.rs); settings can change or differ inside experiments.
- Meta AI (2023, June 29). The AI behind unconnected content recommendations on Facebook and Instagram. Meta AI Blog. Meta. ai.meta.com · official page · productionContent understanding, the shortlist funnel and minute-level model updates for recommended posts; high level, no numbers on weights.
This is the plain-language version. The full research — every stage written out formally — is Jean-Paul Azzi's, and it's becoming a book. Every claim here points at the platform's own paper, page or code; nothing comes from marketing blogs.