The case against the YouTube outlier method
The YouTube outlier method spots videos that beat a channel's average. Research on hits, imitation and AI explains the risks of copying the winners.
On this page
- The same song was a hit and a flop.
- What is the YouTube outlier method?
- Is an outlier video a repeatable formula?
- Why is copying the top performer the wrong lesson?
- Why do outlier tools make YouTube videos look the same?
- Why might a copy of a viral video get fewer views?
- Does model collapse explain the risk of copying outliers?
- Does YouTube punish copied or repetitive content?
- Does YouTube give everyone the same AI suggestions?
- When does outlier research help?
- How do you use outliers without converging?
- Start from your own content.
- What we don’t know yet.
- More Veri Research Reports
- Sources
- Frequently asked questions
The YouTube outlier method ranks videos against their channel’s usual views and uses the winners as references.¹⁶ It can help you find topics to investigate, but a high score does not isolate what caused the views. Music-market and AI-writing experiments suggest reasons to be cautious: popularity can depend on social influence, and shared AI assistance can make independent writers’ stories more alike.¹⁶ We found no direct peer-reviewed test of outlier tools. The case against copying their winners is a reasoned risk assessment, not a measured YouTube effect.
The same song was a hit and a flop.
In a study published in 2006, Matthew Salganik, Peter Dodds and Duncan Watts recruited 14,341 participants to an online music market across two experiments. Everyone could choose from the same 48 songs by unknown bands. In each experiment, participants entered either one of eight social-influence worlds, where they could see their world’s download counts, or an independent condition without those counts.¹
The same songs finished in very different places from one world to the next. Showing people what was popular made success more unequal and harder to predict, and the authors concluded that this unpredictability “is inherent to the process and cannot be eliminated simply by knowing more about the songs or market participants.”¹ Success still partly reflected quality: the best songs rarely did badly and the worst rarely did well.¹
In a paper published in 2008, Salganik and Watts reported an experiment that inverted download counts so the least popular songs appeared to be the most popular. The false rankings affected most songs’ subsequent popularity, although the best songs tended to recover. Projected final rankings tracked appeal less closely in the inverted worlds, with a rank correlation of 0.42 versus 0.82 in the unchanged world. Participants in the inverted worlds also downloaded less music per person.²
The relevance to an outlier dashboard is the uncertainty behind the winning result. A view count records an outcome without identifying how much came from the content or from its distribution.
What is the YouTube outlier method?
The outlier method is a research routine that ranks videos by a multiple of their channel’s recent average or median views instead of by raw views.¹⁶ In a hypothetical average-based example, 200,000 views divided by an 8,000-view baseline gives 25x, while 1 million divided by 900,000 gives about 1.11x.
1of10’s guide calls 3x the working rule for “worth a look” and 5x and up “worth studying,” and notes that those lines “are conventions creators use, not a YouTube metric.”¹⁶ vidIQ offers outlier research and a “Give me ideas” action that starts a new Blueprint from a video’s title.³⁷ TubeBuddy’s Content Outliers tool provides reference videos and suggested titles and descriptions.³⁸
Ranking by ratio helps distinguish a channel’s breakout from its ordinary performance. We would use it to choose what to investigate, rather than let the ranking decide the next video.
Is an outlier video a repeatable formula?
The score alone cannot establish a repeatable formula. It records what happened to one video relative to a baseline without isolating the reason.
Relevant YouTube evidence comes from a 2012 study of clones, videos with the same or nearly identical content and audio. Youmna Borghol and colleagues analyzed 48 sets covering 1,761 videos. “When controlling for video content, we observe a strong linear ‘rich-get-richer’ behavior,” the authors wrote, “with the total number of previous views as the most important factor except for very young videos.” For young videos, they found that uploader characteristics and the number of keywords mattered relatively more.³
The data was collected from February 2010 to April 2011, so the study does not measure today’s recommendation system. It is an observational study, not a test of copying formats or of outlier tools. The authors warn that “inaccurate conclusions can be reached when not controlling for content.”³
| What an outlier tells you | What it doesn’t tell you |
|---|---|
| A video attracted views with one channel | Whether your audience wants the same topic |
| The video beat that channel’s recent average | Which part of the video caused the result |
| The whole video attracted views | Whether its packaging caused that result |
| People watched | Whether viewers were satisfied |
Why is copying the top performer the wrong lesson?
When luck and self-reinforcing success strongly influence results, the most extreme performers can be less informative about skill than moderately high performers. Jerker Denrell and Chengwei Liu modeled such conditions in a 2012 PNAS paper. In one simulation, players with 50 successes out of 50 had less skill on average than players with about 40. Under the conditions they studied, “the highest performers may not have the highest expected ability and should not be imitated or praised.”⁴ This is a model of performance, not a finding about which YouTube creators to copy.
Denrell’s 2003 model showed that risky practices “may seem to be positively related to performance” in a sample of survivors even when they are unrelated to performance in the full population.⁵ A winners-only research sample creates the same concern. Studying a hypothetical 25x hit without looking for failed attempts at the same format gives you no comparison group.
Why do outlier tools make YouTube videos look the same?
Shared inputs can make independently produced work more similar. The strongest evidence here comes from creative tasks outside YouTube, so it establishes a possible mechanism rather than the effect of outlier dashboards.
A 2024 experiment in Science Advances analyzed eight-sentence stories from 293 writers assigned to work without AI ideas or with access to one or up to five ideas. Anil Doshi and Oliver Hauser found higher ratings of novelty and usefulness with AI access, especially for writers who scored lower on a creativity test. Stories also became more alike. Similarity to other stories in the same condition rose by 10.7% of the human-only score range with access to one idea and 8.9% with access to five. The authors wrote that “writers are individually better off, but collectively a narrower scope of novel content is produced.”⁶
Other studies examine related forms of similarity. They cover different tasks and do not all measure the same individual-versus-group trade-off:
- Writing style. A 2026 study in Nature Human Behaviour found that LLM polishing and rewriting cut writing-complexity variance by 21% to 50% across tested datasets and models. Separately, it observed falling variance in social media, news and science writing after ChatGPT’s release. That timing alone does not establish causation.⁷
- Brainstorming. A 2025 study in the same journal found ChatGPT improved individual ideas while reducing “the diversity of ideas in a pool of ideas.”⁸
- Art. Eric Zhou and Dokyun Lee analyzed more than 4 million artworks on one art-sharing platform. Their observational study reported 25% higher productivity with AI adoption and a 50% rise in favorites per view, from 2% to 3% by month six. Average content novelty and both average and peak visual novelty declined, while peak content novelty increased.⁹
- Different models. A 2025 preprint found that responses from different AI models were more similar on standardized creativity tests than people’s responses were to one another. It did not test whether creators could regain variety by switching tools.¹⁰
Jon Kleinberg and Manish Raghavan studied a related problem, algorithmic monoculture, in a theoretical model of hiring. Their 2021 PNAS paper showed conditions under which shared algorithmic rankings can reduce overall decision quality “even when the algorithm is more accurate for any one agent in isolation.”¹¹ A 2022 NeurIPS paper found that sharing training data “reliably exacerbates homogenization” on algorithmic fairness benchmarks.¹² Neither paper measured creative content.
The proposed analogy is shared reference material: if creators repeatedly choose from the same winners, they risk making similar choices. We found no peer-reviewed measurement of YouTube title or thumbnail convergence and no evidence that all these generators share a training dataset. The writing and art studies show why individual gains and collective similarity can coexist, but they cannot tell us how often that happens on YouTube.
Why might a copy of a viral video get fewer views?
A copy may reach people who have seen the idea before. Research on music and repeated exposure suggests possible disadvantages, but does not establish a general view penalty for copying a YouTube format.
Noah Askin and Michael Mauskapf studied nearly 27,000 songs from the Billboard Hot 100. They found that “songs sounding too much like previous and contemporaneous productions,” which they describe as highly typical, “are less likely to succeed.” Songs with some degree of differentiation were more likely to rise to the top of the charts.¹³
Repetition can also wear out its welcome. A 2017 meta-analysis found an inverted-U relationship between exposure and liking, with important differences by stimulus type, including visual versus auditory material.¹⁴ In an advertising experiment, ads low in both divergence and relevance wore out with repetition, while ads high in both “show little sign of wearing-out even over repeated exposures.”¹⁵ Neither study tested videos that share a format, so applying them to YouTube is an inference.
1of10’s own guide warns that viewers “who already clicked the original recognise the near-duplicate and skip it,” and advises: “Copy the structure, never the execution.”¹⁶ That is the vendor’s advice rather than an independently tested result. It leaves a useful question for creators: how much of the reference can you retain before your video feels too familiar?
Film sequels offer a counterexample to any blanket claim that repetition fails. A study of 26 years of movie data found higher total attendance for sequels than non-sequels, but lower total attendance than their parent films.¹⁷ It does not test copying another creator’s format, and it shows why familiarity should not be treated as an automatic disadvantage.
Does model collapse explain the risk of copying outliers?
Only as a limited analogy. In a 2024 Nature paper, Ilia Shumailov and colleagues showed that indiscriminate recursive training on model-generated data can cause “irreversible defects” in which “tails of the original content distribution disappear.” Their language-model demonstration used OPT-125m and WikiText-2, not a creator tool.¹⁸ An ICLR 2024 paper on image generators also found declining quality or diversity in self-consuming training loops without enough fresh real data.¹⁹
A paper presented at an ICML 2024 workshop found that accumulating successive generations of synthetic data alongside the original real data avoided collapse in the settings tested.²⁰ Training-data retention and replacement matter, so these results cannot be applied to a product without knowing its training process.
1of10 says its AI is trained on “62 billion YouTube views across 300,000 videos and 52,000 channels.”²¹ That is a vendor-reported figure, and its page does not establish that generated outputs replace original training data in successive rounds. Training on successful videos and recursively training on synthetic data are different processes. These papers therefore do not demonstrate model collapse in 1of10 or any other outlier tool.
Does YouTube punish copied or repetitive content?
Resembling a successful format is not, by itself, the violation described in YouTube’s monetization policy. The policy allows similar patterns when the substance varies, while restricting repetitive, mass-produced and insufficiently transformed reused content.²²
-
Monetization. On July 15, 2025, YouTube renamed its existing repetitious-content policy “inauthentic content” and clarified that it covers repetitive or mass-produced material. The policy says “channels where content feels interchangeable from video to video are not allowed to monetize” and cautions against template-produced videos without sufficient variation or value.²² YouTube says AI-using channels “remain eligible to monetize” when they follow the policies, including applicable disclosure rules.²³
-
Recommendations. In its 2021 explanation, YouTube described “valued watchtime” using one-to-five-star surveys, with only four- and five-star ratings counting toward that measure and a model predicting responses where surveys were unavailable. It reported an “immediate 20% drop in views” when watch time entered recommendations in 2012.²⁴ Google engineers wrote in 2016 that “ranking by click-through rate often promotes deceptive videos that the user does not complete.”²⁵ In January 2026, CEO Neal Mohan named managing AI slop as a priority and described efforts toward “reducing the spread of low quality, repetitive content.” The letter supplies no enforcement-effectiveness measure.²⁶
-
Packaging tests. YouTube’s A/B testing feature picks title and thumbnail winners by watch time, with its FAQ describing watch time share. The help page warns that third-party tools “often only optimise for click-through rate, which may determine a different ‘winner’ compared to measuring by watch time share.” Tests require advanced-feature access and eligible videos, and may produce no clear winner.²⁷
Other platforms also distinguish original content. TikTok’s 2024 Creator Rewards announcement listed originality as one of four metrics for eligible videos over one minute.²⁸ TechCrunch reported Instagram’s 2024 announcement that, for identical posts in recommendation surfaces, “it will only recommend the original one,” with exceptions for significantly changed material.²⁹ Meta reported that views and watch time on original Facebook Reels roughly doubled in the second half of 2025 versus the same period in 2024, without establishing how much of the increase its originality policies caused.³⁰ Rules about identical reposts do not establish a penalty for sharing a format.
Does YouTube give everyone the same AI suggestions?
No. YouTube’s Inspiration documentation says initial ideas use channel data and its title card suggests “options based on similar videos and trends.” Access is not universal, and YouTube began gradually retiring the Inspiration tab in August 2026 in favor of Ask Studio.³¹ In September 2026, YouTube said Ask Studio could suggest and test refreshed thumbnails for older videos in the background. It also reported more than 40 million title and thumbnail experiments since the testing feature’s 2024 launch.³² Variety reported Mohan’s figure of more than 1 million channels using YouTube’s AI creation tools daily on average in December 2025.³³
YouTube also announced Research tools to help creators discover outlier videos.³² The platform’s investment in these features is a reason to avoid a blanket argument against optimization. Personalized suggestions and shared trend references can coexist, and these announcements do not show that every channel receives identical ideas.
When does outlier research help?
The counter-evidence is important. Learning from others can work, and shared AI assistance does not always reduce diversity.
- Information from others. A paper published in 2010 reported a simulated tournament of 100 social-learning strategies. The winner relied nearly exclusively on social learning and weighted information by its age.³⁴ The result supports learning from others in that environment, not a universal freshness rule for videos.
- Convention plus a twist. A study of 17.9 million scientific papers found that the highest-impact work was “grounded in exceptionally conventional combinations of prior work” with “an intrusion of unusual combinations.”³⁵ This is a finding about scientific citations, not a formula for a YouTube hit.
- Lower creativity-test scores. Doshi and Hauser found AI access raised story novelty by up to 10.7% for writers with low creativity-test scores and had little effect for high-scoring writers.⁶ The study did not classify participants by their experience as creators.
- More AI examples. A 2025 study with more than 800 participants found that high exposure to AI-generated examples increased collective idea diversity without changing individual creativity. Low exposure did not produce that diversity effect.³⁶ This is counter-evidence to any claim that shared AI assistance always narrows a group’s work.
Our recommendation is to treat an outlier as a lead for research. Its score alone cannot show whether the topic, packaging, audience relationship or distribution caused the result.
How do you use outliers without converging?
- Start from your own outliers. Rank your own videos against your channel’s recent median, because those results come from your audience. We suggest the median rather than the average: one viral upload pulls an average up and makes the rest of your videos look like underperformers, while the median stays near a typical video. That is our method advice, not a tested rule.
- Let videos mature before you score them. We compare a video only after about 30 days, so a week-old upload is not ranked against videos that have had months to collect views. 1of10’s guide also suggests checking a new video’s 7-day and 30-day views against your recent baseline.¹⁶
- Keep Shorts and long-form apart. Score each format against its own baseline. 1of10’s guide gives the reason: “Shorts typically post very different numbers and mixing them breaks the baseline.”¹⁶
- Use outside outliers to investigate demand. Look for a topic that worked across several channels, then check whether the viewers and context resemble yours.
- Make your version. Keep the topic when the demand is real, and build the angle, opening and thumbnail from your own past choices instead of the original video’s.
- Test eligible videos. Use YouTube’s title and thumbnail A/B testing to compare watch time outcomes rather than relying on a click-through prediction.²⁷
- Keep a record of your choices. Over time those decisions become your reference, and it belongs to you.
Start from your own content.
We recommend keeping your own work and decisions as the starting point for a new video. That is an editorial recommendation, not an intervention these studies tested. Outside references can help you investigate an idea without deciding its whole execution.
We built Veri around your own content and choices. Veri is a creator co-pilot powered by a model you own, and its agents start from your transcripts and published content.³⁹ Validate assesses a concept using the idea and your own content. Ideas develops a topic, source video or question you provide. In Your Model you can review and correct your saved preferences, with sources where available.
Veri does not train a shared model on your content.³⁹ Personal-model training is off by default, by request and currently waitlisted. Agents and saved preferences work without it, and platform API data is never used to train AI models.
If you use an outlier tool, keep it for research. Our 1of10 alternatives guide compares the tools by job, and a vidIQ score cannot tell you who to be looks at the same problem through channel scores.
What we don’t know yet.
As of November 2026 we found no peer-reviewed study that evaluates outlier tools, tests whether creators who replicate outliers outperform those who don’t, or tracks YouTube thumbnail and title convergence over time. The evidence here comes from music markets, near-identical YouTube videos, AI-assisted writing and art, simulations and platform documentation, with each study’s scope noted where it appears. Without direct channel-level evidence, we cannot estimate whether or how quickly outlier copying makes a niche converge.
Connect a channel at joinveri.co and start your next idea from something you made.
More Veri Research Reports
- Who owns your AI clone? We read the terms of 7 clone tools.
- What is your content worth to AI companies? Follow the money
- Why AI writes like everyone else
- The AI credits trap: what a thumbnail or clip really costs
- Paying twice for creator AI: what 14 tools' terms say about your data
Every report, with its data, is on Veri Research.
Every source in this report is listed as a downloadable table on the data page.
Sources
- Salganik, M. J., Dodds, P. S., & Watts, D. J. (2006). Experimental study of inequality and unpredictability in an artificial cultural market. Science, 311(5762). PDF
- Salganik, M. J., & Watts, D. J. (2008). Leading the herd astray: An experimental study of self-fulfilling prophecies in an artificial cultural market. Social Psychology Quarterly, 71(4), 338–355. PMC
- Borghol, Y., Ardon, S., Carlsson, N., Eager, D., & Mahanti, A. (2012). The untold story of the clones: Content-agnostic factors that impact YouTube video popularity. ACM KDD. arXiv
- Denrell, J., & Liu, C. (2012). Top performers are not the most impressive when extreme performance indicates unreliability. PNAS, 109(24). PMC
- Denrell, J. (2003). Vicarious learning, undersampling of failure, and the myths of management. Organization Science, 14(2). INFORMS
- Doshi, A. R., & Hauser, O. P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 10(28). Science
- Sourati, Z., Karimi-Malekabadi, F., et al. (2026). The shrinking landscape of linguistic diversity in the age of large language models. Nature Human Behaviour, 10, 1925–1942. Nature
- Meincke, L., Nave, G., & Terwiesch, C. (2025). ChatGPT decreases idea diversity in brainstorming. Nature Human Behaviour, 9. Nature
- Zhou, E., & Lee, D. (2024). Generative artificial intelligence, human creativity, and art. PNAS Nexus, 3(3). OUP
- Wenger, E., & Kenett, Y. N. (2025). We’re different, we’re the same: Creative homogeneity across LLMs. Preprint. arXiv
- Kleinberg, J., & Raghavan, M. (2021). Algorithmic monoculture and social welfare. PNAS, 118(22). PNAS
- Bommasani, R., Creel, K. A., et al. (2022). Picking on the same person: Does algorithmic monoculture lead to outcome homogenization? NeurIPS. arXiv
- Askin, N., & Mauskapf, M. (2017). What makes popular culture popular? Product features and optimal differentiation in music. American Sociological Review, 82, 910–944. Columbia Business School
- Montoya, R. M., Horton, R. S., et al. (2017). A re-examination of the mere exposure effect: The influence of repeated exposure on recognition, familiarity, and liking. Psychological Bulletin, 143(5), 459–498. PubMed
- Chen, J., Yang, X., & Smith, R. E. (2016). The effects of creativity on advertising wear-in and wear-out. Journal of the Academy of Marketing Science. Springer
- du Randt, K. (2026). How to Find Viral Video Ideas With the Outlier Method. 1of10. Checked November 5, 2026. 1of10
- Dhar, T., Sun, G., & Weinberg, C. B. (2012). The long-term box office performance of sequel movies. Marketing Letters, 23. Springer
- Shumailov, I., Shumaylov, Z., et al. (2024). AI models collapse when trained on recursively generated data. Nature, 631. Nature
- Alemohammad, S., Casco-Rodriguez, J., et al. (2024). Self-Consuming Generative Models Go MAD. ICLR 2024. ICLR proceedings
- Gerstgrasser, M., Schaeffer, R., et al. (2024). Is Model Collapse Inevitable? Breaking the Curse of Recursion by Accumulating Real and Synthetic Data. ICML 2024 FM-Wild Workshop poster. OpenReview
- 1of10. About 1of10: YouTube Growth Platform for Creators. Checked November 5, 2026. 1of10
- YouTube Help. YouTube channel monetization policies. Checked November 5, 2026. Google
- Sarah (TeamYouTube). (2025). Response to creator questions about YPP policies (July 2025). Checked November 5, 2026. Google
- Goodrow, C. (2021). On YouTube’s recommendation system. YouTube Official Blog. Checked November 5, 2026. YouTube
- Covington, P., Adams, J., & Sargin, E. (2016). Deep neural networks for YouTube recommendations. ACM RecSys. Google Research
- Mohan, N. (2026). From the CEO: What’s coming to YouTube in 2026. YouTube Official Blog. Checked November 5, 2026. YouTube
- YouTube Help. A/B test titles and thumbnails. Checked November 5, 2026. Google
- TikTok Newsroom (2024). Introducing the new Creator Rewards Program. Checked November 5, 2026. TikTok
- Malik, A. (2024). Instagram is updating its algorithm to surface more content from smaller, original creators. TechCrunch. TechCrunch
- Meta (2026). Rewarding original creators on Facebook. Checked November 5, 2026. Meta
- YouTube Help. Explore Inspiration tab on YouTube. Checked November 5, 2026. Google
- Pappu, A. (2026). New tools to power your creation journey from start to finish. YouTube Official Blog. Checked November 5, 2026. YouTube
- Spangler, T. (2026). More Than 1 Million YouTube Channels Are Using AI Tools Daily (headline excerpt). Variety. Variety
- Rendell, L., Boyd, R., et al. (2010). Why copy others? Insights from the social learning strategies tournament. Science, 328(5975). Science
- Uzzi, B., Mukherjee, S., Stringer, M., & Jones, B. (2013). Atypical combinations and scientific impact. Science, 342(6157). Science
- Ashkinaze, J., Mendelsohn, J., et al. (2025). How AI Ideas Affect the Creativity, Diversity, and Evolution of Human Ideas: Evidence From a Large, Dynamic Experiment. ACM Collective Intelligence. arXiv
- Adrian. (2026). Outliers. vidIQ Help Center. Checked November 5, 2026. vidIQ
- Davidson, A. Outlier YouTube videos and how to find them. TubeBuddy. Checked November 5, 2026. TubeBuddy
- Veri. Veri, the creator co-pilot, powered by a model you own. Checked November 5, 2026. Veri
Frequently asked questions
What is an outlier video on YouTube?
An outlier video is one that got far more views than its own channel usually gets, measured as a multiple of the channel's recent average or median. Creators and tool makers often treat 3x as worth a look and 5x or more as worth studying. Those thresholds are conventions, not a YouTube metric.
Does the YouTube outlier method work?
It can show you what an audience responded to, but the score alone can't tell you why or whether it will happen again. Music-market experiments found that popularity partly depended on social influence. A study of near-identical YouTube videos found that prior views were an important predictor even when content was similar. Neither study tested outlier tools directly.
Why do outlier tools make videos look the same?
They may encourage similarity when creators use the same examples, but we found no direct study establishing that effect on YouTube. Experiments with AI-assisted writing show that individual work can improve while the group's work becomes more alike. That is evidence for a risk, not a measured effect of outlier tools.
Does YouTube penalize content copied from another creator's outlier?
Resembling a successful format is not, by itself, the violation described in YouTube's monetization policy. YouTube renamed its existing repetitious-content policy inauthentic content on July 15, 2025, clarifying its coverage of repetitive or mass-produced content. Its separate reused-content policy can apply when someone repurposes another creator's material without substantial original contribution.
Why do all YouTube thumbnails look the same now?
Shared references and generators could encourage similar choices, but we found no peer-reviewed study measuring YouTube thumbnail convergence. A study of more than 4 million artworks found declining visual novelty among AI adopters alongside higher productivity and favorites per view. That art-platform result does not establish what causes thumbnail similarity on YouTube.
How do I use outliers without copying them?
Start from your own idea or your own channel's best videos, use outside outliers to investigate demand, and develop your own angle and packaging. Compare eligible titles and thumbnails with YouTube's A/B tests, and consider watch time and satisfaction rather than click-through rate alone.
Has anyone studied whether outlier tools work?
Not that we could find. As of November 2026 we found no peer-reviewed study that evaluates outlier tools directly or measures whether creators who replicate outliers outperform those who don't. The research in this article comes from music markets, near-identical YouTube videos, AI-assisted writing and art, simulations and platform documentation.