I Analyzed 2,500 YouTube Videos to Understand Why RedLetterMedia Are Internet Darlings

There are millions of channels that chase the YouTube algorithm.

The legendary Hack Frauds from Milwaukee are not among them.

And yet, somehow, every video they post racks up nearly as many views as they have subscribers — they have 1.59 M subscribers as of writing — in fact, as I found, they get 4.5x more views per video than MrBeast, 8x more than MKBHD, and 16x more than Linus Tech Tips. This is incredible. As someone who works with data all day, I simply had to know why.

So I pulled 2,500 videos from RLM and YouTube’s biggest channels, treating my favorite YouTube channel like a dataset, and deep dove it all.

What I found was fairly obvious if you’ve ever watched them before: people can smell when you’re optimizing for the platform instead of them. RLM’s secret weapon isn’t SEO or outrage bait — it’s that their audience trusts they’ll always be genuine, that they’ll never chase trends.

If you’d like to read at your own pace — here’s the Table of Contents. Enjoy!

1. Methodology
2. What YouTube Metadata Looks Like
3. Audience Efficiency
4. The Value of Authenticity (The Neil Breen Effect)
5. The ‘Video Essay’ Format
6. The Anti-Clickbait Manifesto
7. Going Against The Attention Economy
8. Self-Sustaining Content
9. Other YouTube Metrics
10. Conclusion

Methodology

I started by pulling every RedLetterMedia video I could use — they have 884 of them, but only 611 have subtitles, too, something I needed for my analysis — across their major playlists: Half in the Bag, Best of the Worst, re:View, and a few oddballs that didn’t fit anywhere.

But to understand whether RLM’s approach was just different or actually effective, I needed a baseline. So I also pulled roughly 1,500 videos from YouTube’s biggest success stories: Linus Tech Tips (and their network of channels), Marques Brownlee (MKBHD), Cinemassacre (similar niche and format to RLM + another 2010s internet phenomenon), and MrBeast. These channels represent the platonic ideal of YouTube success — algorithmic darlings that follow best practices, optimize aggressively, and scale like startups. If RLM was breaking the rules and still winning, I wanted to see how much they were breaking them, and what they were doing instead.

Altogether, about ~2500 videos. For this, I used yt-dlp, the open-source command-line tool that lets you fetch both metadata and subtitles from YouTube. It’s a fork of the old youtube-dl project, but faster, better maintained, and less allergic to modern web changes.

GitHub - yt-dlp/yt-dlp: A feature-rich command-line audio/video downloader

The plan was simple:

• Download JSON metadata for every video (title, views, likes, upload date, tags, etc.)
• Grab the subtitles in SRV2 format (Needed these for further analysis — I’ll explain more in a bit)
• Merge and clean everything into one dataset
• Analyze the heck out of it.

To extract all this video data, I did something fairly hacky to actually automate yt-dlp at scale.

import subprocess
import json
import time
from pathlib import Path
from datetime import datetime

# Configuration - RLM Playlists
PLAYLISTS = {
    "BOTW": "https://www.youtube.com/playlist?list=PLJ_TJFLc25JR3VZ7Xe-cmt4k3bMKBZ5Tm",
    "HITB": "https://www.youtube.com/playlist?list=PL34C1F26D03F5F9B8",
    "RE-VIEW": "https://www.youtube.com/playlist?list=PLJ_TJFLc25JSmtBkyIYqgD5KabbU57yzY",
    "MOVIE TALK": "https://www.youtube.com/playlist?list=PLJ_TJFLc25JSpxBDbSYnlyjHlg0YIp2ZX",
    "BLACK VOID": "https://www.youtube.com/playlist?list=PLJ_TJFLc25JRdChhr9sp00ajl7FAeKDx-",
}

# Comparison channels - fetches last 200 videos from each
# These channels are here as representatives of "YouTube success" 
# by the numbers.
COMPARISON_CHANNELS = {
    # 1. Cinemassacre - James Rolfe (same niche, same era)
    "Cinemassacre": {
        "url": "https://www.youtube.com/playlist?list=PLbQ-gSLYQEc7Ie64tC64qIYYPCVL6Ftn9",
        "sample_size": 200,
        "type": "playlist",
        "creator": "Cinemassacre"
    },
    
    # 2. MKBHD - Marques Brownlee (main + secondaries)
    # channel details here

    # 3. Linus Tech Tips - Linus Media Group (main + active secondaries)
    # channel details here
    
    # 4. MrBeast - Jimmy Donaldson (main + active secondaries)
    # channel details here
}

OUTPUT_DIR = "youtube_data"

def download_specific_videos(video_ids, output_dir="youtube_data"):
    """Download metadata and subtitles for specific video IDs"""
    output_path = Path(output_dir)
    output_path.mkdir(exist_ok=True)
    
    print(f"\nDownloading {len(video_ids)} videos...")
    print(f" Output directory: {output_dir}")
    
    for i, video_id in enumerate(video_ids, 1):
        print(f"\n[{i}/{len(video_ids)}] Downloading: {video_id}")
        
        cmd = [
            "yt-dlp",
            f"https://www.youtube.com/watch?v={video_id}",
            "--skip-download",              # Don't download video, just metadata
            "--write-info-json",            # Save metadata as JSON
            "--write-auto-subs",            # Download auto-generated subtitles
            "--sub-format", "srv2",         # SRV2 format (includes timing data)
            "--sub-lang", "en",             # English only
            "--output", f"{output_dir}/%(channel)s/%(upload_date)s-%(id)s-%(title).100s",
            "--ignore-errors",              # Continue on errors
            "--no-overwrites",              # Skip existing files
        ]
        
        try:
            subprocess.run(cmd, check=True, capture_output=True)
            print(f" Success")
            time.sleep(1)  # Rate limiting
        except subprocess.CalledProcessError:
            print(f" Error/already exists")
            continue

def get_videos_from_channel(channel_url, channel_name, sample_size=50):
    """Get most recent videos from a YouTube channel"""
    print(f" Fetching {sample_size} most recent videos from: {channel_name}...")
    
    cmd = [
        "yt-dlp",
        f"{channel_url}/videos",
        "--flat-playlist",
        "--playlist-end", str(sample_size),
        "--print", "%(id)s|%(title)s|%(view_count)s|%(upload_date)s",
        "--quiet"
    ]
    
    result = subprocess.run(cmd, capture_output=True, text=True, check=True)
    videos = []
    
    for line in result.stdout.strip().split('\n'):
        if line:
            parts = line.split('|')
            if len(parts) >= 3:
                videos.append({
                    'id': parts[0],
                    'title': parts[1],
                    'views': int(parts[2]) if parts[2] != 'NA' else 0,
                    'upload_date': parts[3] if len(parts) > 3 else 'unknown',
                    'channel': channel_name
                })
    
    print(f"Found {len(videos)} videos")
    return videos

# Example usage:
# python extract.py --playlists ALL              # Download all RLM content
# python extract.py --comparison ALL             # Download comparison channels

The idea is to first extract a list of all videos in the playlist/channel, and then download them — but only the metadata and subtitles.

The HTTP 429 Problem

In practice, this was… not simple. YouTube really doesn’t like being scraped at scale, especially when you’re pulling metadata for 2,000+ videos in a short window. I quickly started hitting HTTP 429 errors (“Too Many Requests”), and adding sleep timers between requests only slowed things down, and didn’t really fix the issue.

TL;DR I needed proxies. After some research, I just used these. This routes requests through a rotating pool of residential IPs — real addresses that look like organic traffic that YouTube can’t easily block.

Web Unlocker - Automated Web Unblocker

yt-dlp supports custom proxy configuration through its --proxy flag, so once I had Bright Data set up, it was remarkably easy to just do this:

yt-dlp "VIDEO_URL" 
 --proxy "http://USERNAME:PASSWORD@brd.superproxy.io:22225" 
 --skip-download 
 --write-info-json 
 --write-auto-subs

It needs credentials, so I did have to sign up here, but that pretty much unlocked CAPTCHA solving, IP rotation, and retry logic with zero effort on my part, which meant I could run the script overnight without babysitting it.

💡 Using it for YouTube like this is not widely available yet — I had to reach out to Bright Data to actually get authorization. I can see a use case where you’re downloading thousands of actuarial videos for AI training — not just metadata — and then this would be essential.

Classification

From there, I built a lightweight classification system for each video:

• Which RLM series it belongs to (Half in the Bag, Best of the Worst, etc.)
• Any major franchises or topics mentioned (Star Wars, Marvel, Trek, etc.)
• Which cast members appear
• Notable personalities mentioned (writers, directors, actors; detected by scanning titles + tags + subtitles with regex)

For this, I combined automated tagging with manual correction — using an LLM was out of the question because anyone who’s actually watched RLM knows their videos are full of inside jokes and running gags that no model will ever understand.

Analyses I Ran

Once everything was clean, I ran analyses + comparisons on titles, duration, profanity density, upload timing/frequency, heatmaps, and video structure itself (via subtitles).

This isn’t exactly academic research; just a weekend mix of scripting, scraping, and curiosity. But the results were fascinating. RLM somehow breaks every metagaming rule YouTube tries to enforce and still wins.

Before we get into it, let’s quickly look at what kind of data we actually downloaded.

What YouTube Metadata Looks Like

Each video I pulled from the channel ends up as a pair of files:

20211231-rpSo4fu1rgM-Half in the Bag： The Matrix Resurrections.info.json  
20211231-rpSo4fu1rgM-Half in the Bag： The Matrix Resurrections.en.srv2

The first is a JSON metadata dump, the second is a subtitle file (autogenerated ones in absence of a channel-provided one; in YouTube’s SRV2 format). yt-dlp quietly can grab a lot more information than most people realize.

The JSON contains far more than you’d expect — not just title, views, and likes, but the entire playback manifest: format arrays with codec/resolution/bitrate details, thumbnail URLs, automatic captions in multiple languages, engagement metrics, and even the per-second audience-retention heatmap (the “popular parts” graph).

The SRV2 subtitle files (basically just XML) are equally useful: they’re timestamped dialogue blocks that let you reconstruct pacing, detect silence gaps, or even align jokes with spikes in the heatmap.

<text t="6080" d="7200" w="1" r="15" c="1">hey</text>
  <text t="6399" d="6881" w="1" append="1"> nostalgia</text>
  <text t="7120" d="6160" w="1" append="1"> fans</text>
  <text t="7919" d="5361" w="1" append="1"> neo</text>
  <text t="8320" d="4960" w="1" append="1"> and</text>
  <text t="8480" d="4800" w="1" append="1"> trinity</text>
  <text t="9040" d="4240" w="1" append="1"> are</text>
  <text w="1" t="9270" d="4010" append="1">
</text>
  <text t="9280" d="5600" w="1" r="15" c="1">back</text>
  <text t="9760" d="5120" w="1" append="1"> in</text>
  <text t="10000" d="4880" w="1" append="1"> joe</text>
  <text t="10240" d="4640" w="1" append="1"> dante&amp;#39;s</text>
  <text t="11120" d="3760" w="1" append="1"> matrix</text>
  <text t="11840" d="3040" w="1" append="1"> resurrections</text>
  <text w="1" t="13270" d="1610" append="1">
</text>

Each of these JSON + SRV2 combinations weighs about 1.5 Megabytes; multiplied by 611 videos (from RLM alone, that is. And ~2500 total), it’s enough structure to keep any analyst busy for weeks.

Also, these subtitles above give us a glimpse into RLM’s brand of humor — Joe Dante is not involved with Matrix Resurrections in any way, of course — but to explain the joke would take some doing. 😄 See why I said you can’t use an LLM to categorize this content?

All told, yt-dlp can get you a full behavioral dataset — well, once you get around YouTube fencing you off. That’s what the proxy layer is for.

Let’s get into the actual data now.

1. Audience Efficiency

On paper, RLM shouldn’t be able to compete with the giants. They have a fraction of the subscribers of MKBHD, Linus, or MrBeast — and yet, on a per-subscriber basis, their videos pull more weight than almost anyone in their league.

Before we begin, here are the November ’25 subscriber counts for these channels.

• RLM: 1.59M
• Cinemassacre: 3.95M
• MKBHD: 20M
• LTT: 16.6M
• MrBeast: 450M

First, let’s look at average views per video. The results are not surprising.

But the real story appears when you normalize views by subscriber count — or, viewership efficiency:

That’s…insane. Here’s the breakdown:

A channel with 1/12th the subscribers of MKBHD is pulling more per-video viewership efficiency than anyone else in the comparison — 4.5x more views per subscriber than MrBeast, whose reach is practically the platform’s upper bound, representing the farthest you can possibly get on YouTube.

RLM’s audience cannot be categorized as anything resembling a broad funnel. Their subscriber number actually understates their reach, because whoever subscribes simply watches everything.

It’s the opposite of the algorithmic sprawl you see in big channels, where subscribers are passive, half-forgotten ghosts from previous content eras. (I’m not even going to touch the can of worms that is paid, artificially inflated sub counts.)

But why is that? Why does a 1.5M-sub channel punch this far above its weight? Why do RLM’s viewers feel less like an audience and more like a culture that sustains itself?

To understand that, we have to look at value proposition —i.e. what makes people show up.

Code used in this section: https://gist.github.com/sixthextinction/d15da133622d68e00a5196d66448b9e7

2. The Value of Authenticity

If you’ve never heard of Neil Breen, you’re not alone.

He’s a one-man filmmaking machine — writer, director, star, editor, composer, and distributor of a series of aggressively earnest, awful, low-budget “outsider” films that feel like they were beamed in from a parallel universe. Double Down, Fateful Findings, Twisted Pair — each one terrible in different ways, each one defying logic (cinematic and otherwise!) in a way that’s hypnotic.

For most of YouTube, Breen wouldn’t even register as a data point. He’s too niche, too weird, too unsearchable.

And yet, RedLetterMedia’s videos about him routinely outperform their uploads about multi-billion-dollar franchises.

A Best of the Worst spotlight episode on a Breen film can pull as many views as one discussing Justice League or Doctor Strange. When you line up every RedLetterMedia video about Neil Breen against their uploads about Marvel or DC, Breen should statistically lose. Six videos about a niche outsider filmmaker shouldn’t hang in the same orbit as thirty-one videos about the most recognizable entertainment brand on the planet.

Right…?

Wrong.

Let’s look deeper, this time into retention data — that would be YouTube heatmaps — to see how these videos hold user interest over time. After all, RLM videos are frequently 60–90+ minutes long.

…and they continually hold viewer interest.

Remarkable. Now I was really curious. What actually stops Neil Breen, and consistently?

I extended my criteria to include the heavyweights in Star Wars and Star Trek again, and…yeah. That does it.

Only the might of the Mouse’s Star Wars machine, and the raging Star Trek fanbase can tame the juggernaut that is Neil Breen’s insanity, apparently…and it’s still close!

So what does this say?

On paper, this approach should break every rule of the YouTube metagame:

• There’s no search traffic for “Neil Breen.” In fact, most search traffic involving him only materialized after RedLetterMedia’s videos.
• Neil Breen isn’t the only niche, “so bad it’s good” trash they cover. All the ‘Spotlight’ episodes are like this — thoroughly cursed, bad movies made immensely entertaining by their discussion. Otherwise, there’s no trending topic hook here, no SEO juice. These are about as niche a topic as one could get.
• And as we’ve already seen, all their titles are deadpan, zero-clickbait (“Best of the Worst: Twisted Pair…” etc.) instead of algorithm-optimized keyword bombs.

But that’s the point. It’s not about Neil Breen himself.

When RLM covers someone like him, they’re not optimizing for discovery — they’re just making a video about a movie that they found weird and presented it as a panel of friends getting drunk and discussing a bad movie together.

And it works because they’ve trained their audience for authenticity.

To click not because of the subject, but because of the curators. Breen just happens to be a perfect stress test for that relationship: if viewers will show up for him, they’ll show up for anything. If the RLM crew can make awful, cursed movies a laugh riot to watch because of their commentary, they can do it for anything.

That’s a big part of what makes their channel antifragile. They will never need the next Marvel release, the next Star Wars meltdown, the next actor/director trending for a week — to stay relevant. Their Last Jedi/Rian Johnson video got 6.78M views — 3.8x their channel average — but had LOWER engagement rate (1.71% vs avg. 2.3%). This was algorithmic virality, not community devotion. Casual viewers clicked for Last Jedi controversy, but didn’t stay to discuss.

And they know that. They could churn out hundreds of controversial ragebait/hot take videos if they wanted to — After The Last Jedi they certainly had evidence that that content would get views, and with them being nerds who already had Plinkett as an established character for these takes, it would be easy — but that’s just not their brand. They never followed up on any of that.

If the YouTube metagame is about chasing what’s popular, RLM’s is about making things popular on their own terms.

That’s the real “Neil Breen Factor” — and the fact that they pioneered a whole platform format should be proof enough of that.

Code used in this section: https://gist.github.com/sixthextinction/a4d7025ce066f9e3dcc305f931e9258b https://gist.github.com/sixthextinction/9ec17940dc4aebf4e3b3de8e92c8381c

3. The ‘Video Essay’ Format

If you were to reverse-engineer RLM’s channel for “best practices,” the YouTube optimization handbook would have a stroke. Every single lever that’s supposed to drive growth — video length, titles, upload cadence, sponsor integration, even profanity filters — is pointed in the opposite direction.

Let’s start with duration.

Across 611 RLM videos, the average runtime clocks in at 41.6 minutes, with a median of 39.1. That’s longer than three typical MKBHD uploads combined, more than double a Linus Tech Tips episode, and almost triple a MrBeast video. They’re not optimizing for the short-form dopamine loop — at all.

Which makes all the sense in the world when you realize that RLM were pretty much the pioneers of the “video essay” format, starting with the Plinkett Star Wars reviews back in 2011–12. That style is heavily copied these days — and while I’m a big fan of Lindsay Ellis and Every Frame a Painting, their channels are mostly inactive and they were never cultural phenoms like RLM and (to a lesser extent) Cinnemassacre/AVGN were.

Press enter or click to view image in full size

With an average runtime of ~42 minutes (vs. ~12–16 min. for the rest), RLM viewers behave like they’re watching something a third that length.

Only a ~52% drop off in the first three minutes, compared to over 60% for MKBHD and 70% for MrBeast — creators whose videos are less than half as long. And even by the halfway mark, RLM’s viewership curve stays unusually flat (best final retention of the lot — 18.95%)

Where most channels show sharp decay — the typical “ski slope” pattern YouTube heatmaps are famous for — RLM’s retention graph looks more like a plateau. The audience doesn’t skim for highlights; they sit through the conversation. They don’t consume “content”, they’re just hanging out with familiar voices.

And that’s just the start.

Code used in this section: https://gist.github.com/sixthextinction/d0501230c4b345fe271e15834d39e76f https://gist.github.com/sixthextinction/b552717fa151275135a0bc5880ec1a01

4. The Anti-Clickbait Manifesto

If you look at title structure, RLM is allergic to clickbait. Here’s the criteria I came up with for quantifying this.

def analyze_title(title):
    """
    analyze a single title for optimization metrics
    """
    if not title:
        return None
        
    analysis = {
        'length': len(title),
        'word_count': len(title.split()),
        'has_caps': bool(re.search(r'[A-Z]{3,}', title)),
        'caps_percentage': sum(1 for c in title if c.isupper()) / len(title) if title else 0,
        'has_exclamation': '!' in title,
        'has_question': '?' in title,
        'clickbait_keyword_count': sum(1 for kw in CLICKBAIT_KEYWORDS if kw.lower() in title.lower()),
        'has_colon': ':' in title,
        'has_numbers': bool(re.search(r'\d+', title)),
        'has_year': bool(re.search(r'(19|20)\d{2}', title)),
    }
    
    # calculate clickbait score
    score = 0
    score += min(analysis['caps_percentage'] * 50, 20)
    score += (title.count('!')) * 10
    score += analysis['clickbait_keyword_count'] * 15
    score += 10 if analysis['has_question'] else 0
    score += 5 if analysis['has_caps'] else 0
    analysis['clickbait_score'] = min(score, 100)
    
    return analysis

Their titles average 44 characters and score an average clickbait rating of 7.8 out of 100 — far below the 17+ range of MKBHD and LTT.

Here’s the full title metrics.

Press enter or click to view image in full size

For Clickbait Score, lower is better

They rarely use all caps (5%), exclamation points (4%), or teaser-style phrasing. Pretty much all of their titles are like these:

• re:View — John Carpenter’s The Thing
• Best of the Worst: Kill Squad, Ryan’s Babe, and Demonwarp
• Half in the Bag: Hereditary

Series name, movie title(s), a year if needed. That‘s it. You can’t even perform sentiment analysis on these titles because they’re so cut-and-dry, so very boring and to the point.

Code used in this section: https://gist.github.com/sixthextinction/599f4eb6e9954528983691e3dedfbe6d

5. Going Against The Attention Economy

If YouTube is an attention economy, RLM violates its first commandment: grab the viewer in the first 30 seconds or lose them forever.

When I analyzed subtitle timing and word counts, the difference was immediate.

Across 600+ RLM videos, their average words-per-minute (WPM) hovers around 158, slower than MKBHD (192), Linus (179), and even MrBeast (183). They’re not in a rush. There’s space for silence, for laughter, for someone to start a tangent and forget what they were saying — frequently played for laughs, especially with copious amounts of alcohol involved. 😅

Even their opening 30 seconds — the supposed “hook window” — average just 55 spoken words, compared to 80+ for most other creators. These videos are conversational, not performative. If anything — a strategy they’ve used frequently is to include long-running comedic skits before the review even starts. This *should* turn most people off. It does not.

I went further and analyzed subtitles for pause data to get a feel for their editing style, and the contrast sharpens. RLM averages 3.4 seconds between sentences and over 7 long pauses per video (defined as a gap over 5 seconds). Everyone else? Between 1.7 and 2.3 seconds, and barely a single long pause per upload. Video length differences don’t matter here — I’m only looking at subtitles and measuring from timestamp to timestamp.

Their speaking density (how much of the video actually contains speech) is around 95%, leaving 5% of the runtime as silence, laughter, or background noise. That sounds small, but for a ~42-minute video, it’s two minutes of pure dead air — something most creators would edit out instantly.

That rhythm builds intimacy — the sense that you’re watching real friends talk, not hosts performing for a retention curve.

What about upload cadence? They’re not grinding daily or even biweekly. RLM releases roughly 0.86 videos per week, or one every eight days. That’s slower than almost every other major creator — yet it works. That’s incredibly rare in YouTube’s algorithmic economy.

It should be noted that Cinemassacre’s output has slowed down considerably in recent years.

Even their upload timing is contrary to pattern. While most large creators cluster around Friday–Sunday (the “weekend watch window”), RLM spreads releases pretty evenly across the week — Monday through Friday look almost identical in volume. There’s no clear optimization curve here, really. Just consistency.

Code used in this section: https://gist.github.com/sixthextinction/28bc504371ebf2778c786a4ec25357dchttps://gist.github.com/sixthextinction/408d16123c58473003f87421e5730d71

6. The Real Strategy — Self Sustaining Pipelines

There is one thing they *do* pay attention to, however. RedLetterMedia is a channel that engineered three self-sustaining ecosystems.

• Half in the Bag is their consistency engine: 270 videos averaging 1.45M views each. This format keeps the channel alive in YouTube’s bloodstream — latest movies/shows, steady cadence, steady presence.
• Best of the Worst is the viral engine: 2.7M average views per episode, drawing the broadest audience through chaos, laughter, and novelty. There’s never a shortage of bad movies (from any decade, but mostly the VHS era — 80s and 90s) they can add value to.
• And re:View is the loyalty engine: the quiet center of the storm. It exclusively covers movies they like and felt like talking about that week, pulls 34% fewer views than Best of the Worst but drives 63% more engagement — 0.0263 vs. 0.0159.

Engagement Rate = (Likes + Comments) / ViewsThis is just a ratio I’m using that quantifies how actively the audience responds relative to how many people watched:* Higher engagement rate = More likes + comments per view* Lower engagement rate = Fewer likes + comments per view (more passive viewing)

Together, they form a self-balancing system: one pulls in casuals, one keeps the algorithm happy, one feeds the cult.

HITB vs BOTW vs re:View

And of course, we can use a radar chart here to represent each of the three series to find out how elastic they are.

Half in the Bag is their most diverse format as that covers latest movies. BOTW their most consistent. re:View is their most rigid format — it only includes movies/shows they liked.

The choice of content isn’t random (except perhaps re:View?). When you map their data by the decades of the films discussed, a clear gravitational pattern emerges.

The 1980s, 1990s, and 2000s form a flat plateau of peak performance — each decade averages roughly 2.0 million views per video, the most stable period in RLM’s orbit. These decades dominate their library.

Then the floor drops. The 2010s and 2020s, despite making up a massive portion of recent uploads (285 and 177 videos), average only 1.66M views — about 17% less reach.

1980s → 262 videos, 1990s → 180 videos, 2000s → 149 videos

But this isn’t a bad thing — the engagement tells another story. Even as newer decades draw fewer views, their engagement rate holds steady around 0.022–0.023, suggesting fewer casuals, more commitment.

They’ve figured out a content dichotomy that feeds their channel in two different ways. The 80s–00s are either the decades people either

a) love to remember, or

b) love to laugh at how bad that VHS/home video era was, and the glut of abysmal content it incentivized being produced.

Modern Hollywood — the 2010s–20s — are the ones people *actually* love to debate.

The Genre Equation

They use the same strategy for genre, too.

• Action/sci-fi movies pull the biggest audiences — this is mostly the 80s/90s VHS trash they cover, but also many modern superhero movies would be categorized as sci-fi — it’s the mass-market magnet.
• Horror is self explanatory — pretty much the same as action, fuels reach in exactly the same way.
• The presence of Comedy in tags isn’t exactly the genre itself. The tag exists largely on videos where the review/commentary itself is comedic: a bunch of drunk Milwaukee friends riffing on forgotten VHS tapes. “Comedy” here marks the workhorse videos — the backbone of the channel’s output — filled with meta jokes, long-running bits, and sketchy intros. These are the videos that sustain engagement even when the movie being covered is disposable.

We can map engagement rate atop this, too, and we’ll see that while Drama attracts the fewest views, it commands the highest audience engagement (~2.8%), confirming that its smaller audience is far more invested.

That’s easily explained by the kind of “drama” RLM chooses to talk about. These aren’t Oscar-bait or self-serious film-school projects. They’re:

• David Lynch films they genuinely love, handled with care but without any film-snob reverence.
• Prestige releases like Oppenheimer or Inherent Vice.
• Forgotten gems like George A. Romero’s Martin in re:View, or indie dramas like Pig, Dinner in America, and Strange Darling from the Mike & Jay Talk About series.

These are conversations, not content. The kind that invites comments.

When we cross genres with decades, you see this clearly.

Their 1980s and 1990s coverage is dense with Action, and Horror titles, the most crowd-pleasing genres in the dataset.

Meanwhile, the recent decades they cover (2010s–2020s) lean toward Horror, Sci-fi, and Drama, genres that bring more conversation than clicks.

Code used in this section: https://gist.github.com/sixthextinction/c75303ff06311ad737b21e189fc84204

Other YouTube Metrics Defied — Profanity, Sponsors

I’ll put some other findings here that don’t really fit into the other categories.

YouTube’s internal moderation system quietly penalizes videos with high profanity density — something creators like Linus Tech Tips, MKBHD, MrBeast avoid almost completely.

RLM clocks in at a frankly hilarious 0.4 profanities per minute, or roughly twice Cinemassacre’s rate. And somehow, nobody minds. In fact, fans barely notice. The same audience that criticizes forced profanity in AVGN never bats an eye at RLM, because it doesn’t feel forced; it feels lived-in. Not a gimmick, not performative. It feels like how people actually talk when the cameras aren’t supposed to be rolling.

And finally: sponsorships. RLM has none.

Across all 608 videos, there’s only one sponsor mention — and it was a sarcastic NordVPN skit. Compare that to LTT (38% sponsored), MKBHD (47%), and MrBeast (8% — though this is misleading when he has entire videos that are in service of the sponsor), and you start to see the point: RLM’s videos are the product. There’s nothing to sell except the experience itself.

Conclusion

So what do we learn from ~2500 videos?

That RLM’s success is simply…authenticity. The algorithm rewards chasing trends; RLM rewards staying curious. Their audience is here because — sure, perhaps they’ll cover the latest Marvel movie — but mostly because these guys will spend 90 minutes drunk-dissecting an awful shot-on-video mafia movie by someone named Vitaliy Versace that no one else would ever unearth, and make it funnier than any corporate content farm ever could.

The data confirms what their viewers already knew: genuine beats optimized. You can’t A/B test your way into authenticity. The flat retention curves, the loyal rewatches, the complete absence of clickbait panic — it all points to an audience that trusts RLM to be interesting, not topical.

If YouTube is an economy of attention, RedLetterMedia has built a self-sustaining currency by never selling out. The algorithm didn’t make them. They made themselves algorithm-proof by simply giving a shit about weird movies and being funny while doing it.

Hey, so this is the first in a series of data-driven deep dives I’m doing this month — forensic teardowns of things that are interesting, or things that shouldn’t work but do. If you want to see what else I find buried in data, follow along. More weird experiments soon. ✌