---
title: "How YouTube, Instagram and TikTok rank video in 2026: satisfaction signals, AI content understanding, and penalties for reposting"
kind: "Article"
topic: "Attention & feeds"
date: "2026-10-10"
url: "https://canopticon.com/articles/how-youtube-instagram-and-tiktok-rank-video-in-2026"
author: "Chris Telford"
license: "CC BY 4.0"
---
The long-standing complaint about recommendation feeds has been that they count the wrong thing — that a system measuring seconds watched will serve whatever holds a person against their own wishes. The complaint is still fair about the results. It is no longer accurate about the method.

## Watch time stopped being the target

The largest video platform now describes its ranking as a judgement about whether the viewer was glad they watched. It <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Key findings">weights watch time by satisfaction signals like surveys, returns, and session continuation, not just raw minutes</q>, including whether a video <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Key findings">extends the broader viewing session</q>. Instagram's head said the quiet part out loud in a public video series: <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Key findings">The top three signals that matter most for ranking are watch time, likes and sends. So when looking at your insights, pay close attention to average watch time, likes per reach, and sends per reach</q> — sending something to one person now counts for more than a like from many.

There is no single algorithm to describe any more. YouTube <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Key findings">runs separate ranking models for Home, Suggested, Search, Shorts, and Notifications, and in 2025 decoupled the Shorts engine from long-form</q>. Instagram's head added a distinction between audiences: <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Key findings">"likes are slightly more important for connected content, and sends are slightly more important for unconnected content."</q> TikTok's weighting has moved the same way, toward deliberate actions: <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Key findings">comment quality (length/substance) now matters more than comment count</q>. Instagram has also started handing some control to users, launching <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Key findings">"Your Algorithm" (user topic controls)</q> in December 2025.

## The model watches the video

The newer change is less discussed and better evidenced. Recommendation systems used to read a video's surroundings — its title, caption, tags, who watched it. Now a language model watches the thing itself. A peer-reviewed paper by the platform's own engineers describes deploying models for <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Key findings">nuanced content understanding (e.g., discerning content 'vibe') at scale within an industrial short-form video recommendation system</q>. On the judgement that matters most, the model beat the people who used to do the job: <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Key findings">Gemini 2.5 Pro scored an F1 of 81.33% on a nuanced content attribute versus 63.21% for crowd-sourced human raters</q>.

Meta reports the same shift in its own words — <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Key findings">Within six months of developing a state-of-the-art, self-supervised framework for video understanding, we've built and deployed an AI model in Instagram Reels' recommendation system.</q> The practical effect is that the system no longer depends on the creator describing the work accurately: the models <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Key findings">understand semantic meanings of content holistically across different modalities</q>, enough to tell road biking from mountain biking when the caption does not say.

Google is going further than annotation. Its <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Key findings">"Large Recommender Models" (LRM) adapt Gemini for recommendations using "semantic IDs" to tokenize videos</q>, and an engineer's talk said it is <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Key findings">in production for retrieval and being tested for ranking</q>. That last claim rests on a conference talk rather than a paper, and the research flags it as such. Meta put a number on its own change, crediting AI advances with <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Key findings">a 15% increase in Reels watch time</q>, while its largest new model is <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Key findings">currently ads-focused, with multimodal organic ranking described as future work</q>.

Researchers outside the platforms are testing the same idea on a harder problem. A study in Scientific Reports <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Key findings">used LLMs to score the reliability/quality of medical YouTube videos from transcripts and proposed re-ranking search results by LLM quality score</q>. The research draws the implication for anyone making things: <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Details">as models get better at assessing depth, accuracy, and nuance, the gap between "optimized-but-hollow" and "genuinely substantive" content should narrow in the algorithm's eyes</q>

## Where the story outruns the evidence

Not every platform has shown this. The company most often credited with a mind-reading algorithm has published nothing of the kind: <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Key findings">Claims that TikTok "reads" video via multimodal models in ranking appear mainly in marketing blogs, not primary sources.</q> Its documented system is about fast behavioural signals, and its published multimodal work is for moderation. The widely repeated numbers about what it takes to go viral have the same problem — the completion threshold said to have <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Key findings">jumped from 50% to 70%</q> comes from <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Key findings">marketing-blog estimates, not TikTok primary documentation</q>. A good deal of what creators are taught as mechanics is folklore about a system nobody outside has seen.

## Originality became a ranking rule, not a virtue

The second real change is that copying is now penalised by the machinery rather than disapproved of in public. On Instagram, <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Key findings">reposted content is labeled and downranked, with the original creator's version surfaced instead</q>. Across the platforms, <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Key findings">originality and authenticity are now explicit ranking and monetization criteria</q>. An aggregator account that reposts other people's videos is no longer competing on equal terms with the person who made them.

YouTube wrote the rule into its monetisation policy. <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Details">YouTube's July 15, 2025 policy renamed "repetitious content" to "inauthentic content," explicitly demonetizing mass-produced/templated videos at the channel level — a direct response to "AI slop."</q> The reported effect on Instagram is large, though the figures come from creator guides rather than the platform: <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Details">"aggregator accounts saw 60–80% reach drops after this change, while original creators saw 40–60% reach increases,"</q> and the research treats specific weightings like these <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Caveats">as estimates</q>.

## What it means for anything long or detailed

The research is blunt about where careful work lives. Detail-heavy material <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Key findings">underperforms in pure algorithmic swipe feeds but overperforms in search, subscriptions, and self-selected long-form</q>. The subscription route now has numbers behind it: <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Key findings">In the past three months, the Substack app has driven nearly half a million paid subscriptions and more than 32 million free subscriptions to writers and creators. Most of these connections come from the feed.</q> Even the most formula-driven creator on the platform has moved — <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Key findings">MrBeast in 2025–2026 publicly pivoted toward emotional storytelling and narrative, saying audiences have become "numb" to spectacle</q>, and his former retention director called retention-maximising algorithms <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Key findings">poisonous</q>.

Even the word at the centre of the change is the platforms' own. The research cautions that <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Caveats">"Satisfaction," "session contribution," and similar terms are platform-defined and not fully transparent; creators cannot directly measure them.</q>

> **Objection: Satisfaction is just engagement with better manners**
>
> Partly. The measured gains these systems report are still platform gains — the content-understanding deployment was judged on <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Key findings">+0.49% user participation in content creation, +0.21% satisfied consumption</q>, which is to say more posting and more watching. But the instrument has genuinely changed: a system that asks viewers afterwards, that counts a private send over a public like, and that downranks a repost, can be wrong in ways the old one could not even express. Judge it on what it ranks, not on the word.

The numbers for long, careful work are good where people choose it. Documentary-style true crime <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Key findings">earns 30–50% higher CPMs than casual commentary</q>. Podcasts hold listeners too: <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Key findings">True-crime podcasts hit 80–85% completion rates.</q> Publishers have started charging for depth directly, with subscription investigative products from USA TODAY and the Daily Mail; the strategy, in the research's words, is to <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Key findings">paywall niche depth rather than chase algorithmic reach</q>.

## The platforms built on subscriptions rank differently

Where revenue comes from subscribers rather than advertisers, the feed is built to produce subscribers. Substack <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Key findings">explicitly rebuilt its Notes recommendation system to optimize for subscriptions and paid conversions rather than time-spent</q>. Its head of machine learning describes the goal as <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Details">"what's best for you… and eventually subscribe and pay for,"</q> and the system rewards generosity between writers. That has a catch the research records: <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Details">some creators note the amplified Notes drive subscriptions to others, making reciprocal partnerships the workaround.</q>

Other tools skip the ranking altogether. Flipboard's Surf gathers Bluesky, Mastodon, RSS, podcasts and YouTube into <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Key findings">human-curated feeds rather than algorithmic timelines, explicitly positioned for users "exhausted by algorithmic timelines."</q> The pitch from the membership and newsletter platforms is about who owns the audience: growth platforms <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Key findings">"rent" attention while owned platforms (email, memberships) survive algorithm changes</q>.

## The formula trap

The most influential playbook for gaming the old system came from MrBeast's company. Its leaked production memo is <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Key findings">the canonical formula playbook: obsessive second-by-second retention analysis, CTR/thumbnail/title focus, "crazy progression," and escalating-stakes formats</q>. Even it <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Key findings">warns against reusing formats back-to-back to prevent boredom</q>.

The strategists who sell this knowledge now say something closer to the platforms' own line: <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Key findings">"the algorithm will reward what people want to watch"</q>. The research's summary of the trade-off: <q doc="the-state-of-algorithmic-quality-ai-content-assessment-and-discovery-a" section="Key findings">formulas raise the floor but commoditize; durable advantage is audience loyalty, community, and "wow factor" no one else can replicate.</q>

## What this explains

It explains why advice about feeds ages so badly: most of it describes the behavioural scoring of a decade ago, while the live systems are reading the work and forming a view of it. It explains why reposting stopped paying and why originality advice suddenly has teeth — the penalty is in the ranking, not in the etiquette. And it explains the split that creators keep running into: the same careful, detailed piece that sinks in a swipe feed does well in search, in a subscription, in a long video someone chose. The feed is no longer indifferent to what a thing is. It is just still counting its own gains when it decides.