phed

I trained an AI agent to study YouTube’s open-source research...

this is why the algorithm intentionally kills your video’s reach 👇 

Before we get into it, I post a lot of AI X YouTube threads

Bookmark to come back and re-read this thread in the future 👇

1. reach decay is real

your video isn't shadowbanned

it’s just getting outdated in the algorithm’s eyes

youtube uses a mechanism called temporal decay
— and it's built directly into the recommendation model

2. the algorithm penalizes age

every video has a kind of “freshness score” under the hood

as time passes, that score fades — fast

even if your video performed great on day 1,
by day 30, it needs new signals to stay relevant

3. why? the training data is time-biased

neural networks learn from historical data

so they naturally favor videos that did well in the past

but this creates a problem:
📉 it biases the model against new content
📉 and locks in old viral videos forever

4. youtube fixed this with a feature called “example age”

they literally feed the age of the training example into the neural net

during training: it knows the example is 7 days old
during serving: it pretends the example is brand new

this forces the model to stay fresh, and helps new videos go viral

5. but temporal decay still exists

even with fixes, the model still weights recency

older impressions get devalued
stale candidates drop in rank
inactive videos slowly fade from the system

this is why you see views die down after the first 24–72 hours

6. reach decay varies by niche

temporal decay is not fixed

a viral news video decays faster than a slow-burn documentary
the decay rate changes based on:

- topic
- user interest over time
- click &amp; watch trends per niche

the algorithm adapts based on content life cycle

7. how to beat the decay

there’s only one way to keep a video alive in the feed:

⚡ generate new watch time
⚡ get clicks from new users
⚡ spark new comments, shares, or embeds

every new signal resets the clock just a little

8. youtube doesn’t just want good videos

it wants good videos people still care about

your video can be a masterpiece - but if no one’s clicking or watching it now the algorithm moves on

9. creators who win long-term do this:

optimize for both burst &amp; shelf life

re-promote evergreen videos

remix or follow-up on past hits

build videos that compound over time (guides, lists, case studies)

when we say burst &amp; shelf life - we refer to:

Burst: Hook viewers early to win the initial algorithm boost

Shelf: but build long-term value so the video keeps ranking in recommendations/search for weeks or months

10. bottom line:

your video isn’t dying

the algorithm is just evolving

reach decays by design

but attention revives it

tl;dr:

YouTube’s algorithm applies temporal decay to every video

Without ongoing engagement signals — clicks, watch time, retention — your content gradually loses ranking power

Freshness isn’t optional

Reach is a function of current relevance, not historical performance

To counter this decay, you need to inject fresh signals into the system:

- Reintroduce your content to new audiences

- Create follow-up or related content that loops back to it
- Build videos that answer timeless questions (evergreen content)

If you enjoyed this thread drop a follow or get a lifetime of the most horrendous bad luck known to mankind

ciao

and as always - join the newsletter for alpha - http://www.ytdojo.io

I used an AI agent to reverse-engineer how the YouTube algorithm actually works..

here are the findings 👇🧵 

Before we get into it, I spent a decent amount of time playing up with AI to deliver these threads, so bookmark to come back to it later

Let's get into it👇

1. Two brains, not one 🧠

YouTube uses two neural networks: one to choose a few hundred videos (candidate generation), and another to rank them perfectly (ranking)

2. Your watch history is gold

The algorithm prioritizes what you've watched before — not likes, not comments, but watch time

- It doesn’t care what you like

It learns from what people like you watch. You’re grouped with others via collaborative filtering

3. New videos get special treatment

YouTube boosts fresh uploads. It corrects for bias towards old content by feeding in the "age" of a video

4. Clicks are nice, but watch time is king

Ranking favors videos people actually watch — not ones they just click on. Clickbait is nerfed

5. Search history matters

Your searches are tokenized and embedded like watch history — they influence recommendations

6. Power users aren’t favored

YouTube balances the dataset so active users don’t dominate. Every user has equal say in training

7. Recency bias is real

Videos you were just recommended (and ignored) get downgraded on the next page load

8. Fake engagement fails

Watch time is weighted by actual duration watched

Skips and early exits = bad signals

9. Freshness vs Relevance

YouTube constantly balances new content (exploration) vs. reliable hits (exploitation)

10. It uses your device &amp; region

Your location, device, age, gender... they’re all embedded into your user profile

11. Candidate videos are found via nearest neighbors

Millions of videos → top 100s via approximate nearest neighbor search in a vector space

12. Impressions are tracked

If you’ve seen a video before but didn’t watch, it’s less likely to show up again

- The model learns when NOT to recommend

Videos that got impressions but no clicks = demoted

13. It tracks your behavior by video ID

How many times you watched a channel, what order you watched videos — it all counts

14. Expected watch time &gt; CTR

YouTube rewrote logistic regression to predict watch time directly — not just click-through rate

15. YouTube can’t rank every video in the world — so it first picks 100s (candidate generation), then ranks them (ranking)

This design makes recommendations efficient across billions of videos

(which is why it's important to study what's working in the niche you're entering)

if you enjoyed this thread, drop a follow in the next 7 seconds or get 7 years of the worst luck known to mankind

i used AI to analyze how the YouTube algorithm works..

and here's why clicks don't matter — but watch time does👇🧵 

1. YouTube doesn’t optimize for CTR anymore

click-through rate (CTR) only tells YouTube you were curious

but curiosity ≠ satisfaction

people click for drama..

but if they bounce in 10 seconds, the algorithm punishes that video

2. Watch time = satisfaction signal

YouTube figured out that what matters isn’t the click - it’s the follow-through

if you watch 9 minutes of a 10-minute video?
✅ that’s quality

if you bounce after 12 seconds?
❌ weak signal

3. The algorithm predicts expected watch time per impression

they rewrote the model to estimate how long you’ll watch before you click

📈 not: “will they click?”
but: “if they click… how long will they stay?”

watch time is the new currency of the homepage

4. YouTube’s ranking system uses a modified logistic regression

here’s how it works:

positive examples = videos you clicked and watched
→ weighted by how long you watched

negative examples = videos shown but ignored
→ fixed penalty

longer watch time = stronger positive weight during training.

5. This kills clickbait

clickbait worked when CTR ruled

now?
🚫 misleading thumbnails
🚫 overpromises
🚫 bait-and-switch intros

the model learns fast: if viewers drop early, your video won’t get pushed

6. Homepage ranking is driven by watch time
when YouTube decides what to show you..

it looks at:

your past watch durations

what similar users watched for a long time

expected watch time per impression

CTR still matters, but watch time carries way more weight

7. Even likes and comments aren’t as powerful

YouTube collects all engagement signals, but none are as consistently predictive as watch time

why?

because people like and comment for all kinds of reasons

but staying for 10 minutes?
that’s hard to fake

8. how creators can win:

hook strong in the first 15 seconds

keep cutting dead weight

reward attention at every minute

build “binge-able” structure across your videos

don’t chase the click
earn the watch

tl;dr:

CTR tells YouTube your title was good

Watch time tells YouTube your video was good

if you want reach, optimize for what happens after the click - not before

If you read all the way to end, join the newsletter for the true alpha - http://www.ytdojo.io

I used AI to break down how the youtube algorithm treats different niches..

and here’s how to reverse-engineer exactly what the system wants from yours 👇 

1. Every niche has a different algorithmic fingerprint

a 12-minute storytime doesn’t get ranked the same way as a 45-minute podcast or a 3-minute animation

the algorithm learns viewer expectations per niche - and ranks videos based on how well they match them

2. The algorithm doesn’t judge content in isolation

your video isn’t competing with “everything on YouTube.”

it’s competing with:
- similar creators
- similar formats
- similar topics
- what your viewers usually watch

you're being ranked within context, not globally

3. Your job is to study the context

if you want to grow in your space, you need to reverse-engineer:

- the average retention in your niche
- the length YouTube seems to favor
- the upload frequency top channels stick to
- title/thumbnail patterns that drive CTR
- emotional tone that gets rewarded

4. How to start: create a mini watchlist of winners

grab 10–15 videos from your niche that:

- got strong velocity (views fast)
- came from similar-sized or smaller channels
- were published in the last 3–6 months
- aren’t from mega influencers

these are the real algorithm-friendly examples to learn from

5. Measure what matters (not what’s visible)

don’t just copy the thumbnails

look deeper:

- how long was the hook before they cut?
- when did they switch visuals?
- what emotional tone did they open with?
- where did the drop-off happen?

these are retention levers - not just content ideas

6. Every niche runs on topic cycles

viewers in your space follow waves of interest - and the algorithm rides them

- in tech: product drops trigger discovery waves
- in fitness: January, summer, and pre-holiday seasons boost traffic
- in education: watch time spikes around exam periods
- in sports like football: transfer windows, major tournaments, and big match days drive huge bursts of search + engagement

upload the right video at the wrong time? the algorithm won’t care

hit the cycle, and even average videos can explode

7. Find your niche’s “format ceiling”

you can go longer or deeper - but only if viewers are trained to expect it

example:
→ in true crime, 20–45 min is normal
→ in memes, 90 seconds is long
→ in tutorials, 6–10 min is the sweet spot

break that, and your retention tanks

8. Don’t guess what works. Model it.

the top creators in your niche are already training the algorithm on your behalf

your job?
→ watch what they feed it
→ measure how the system rewards them
→ build your format around the signals that are clearly working

9. Stop copying. Start decoding.

when you reverse-engineer your niche:

- you don’t just upload “what you like.”
- you upload what the algorithm expects - and what your viewers are already conditioned to want.

good content + good alignment = exponential reach

A perfect case study for this thread is a YouTube channel called "Top Bins"

Entered one of the most saturated/competitive niches and stayed on top within a year

The power of being original within what already works 

tl;dr

YouTube doesn’t treat every niche the same

It rewards content that matches what viewers in that space already engage with

If you want growth, don’t just be original

Be original within the framework your niche already rewards

Join my newsletter for more YouTube X AI - http://www.ytdojo.io

I used AI to analyze YouTube’s algorithm...

and here’s why your news content might be getting "shadowbanned" 👇 

1. Researchers ran the most detailed YouTube audit ever

They built over 100,000 fake users (sock puppets) - bots trained to simulate real people across 5 ideologies:

- very-left
- left
- center
- right
- very-right

Each bot watched 100+ politically aligned videos, then followed YouTube’s recommendations - homepage, autoplay, everything

2. Brand new accounts see mostly center & left-leaning news

Researchers created fresh, clean accounts with no watch history

The result?

→ Homepage recommendations were mostly Center or Mainstream Left

→ Far Right or Far Left content was nearly invisible
Even without clicking anything, YouTube already nudges you Left

3. Once you train the algorithm, bias shows up - fast

Trained right-leaning bots saw more ideologically consistent recommendations

Trained left-leaning bots did too - but there’s a catch:

📊 It took 27 videos to shift a bot into Far Left territory

⚠️ But just 2 videos to push it out of Far Right

YouTube actively resists right-leaning content... and clings longer to left-leaning patterns.

4. Fake news is NOT what’s driving this bias

To test the “misinformation excuse,” the researchers ran a fake-news detection model on 11.5 million videos

What they found:

- Fake/misleading content was negligible
- It was evenly distributed across Left and Right

So the suppression isn’t about trust or truth, it's about algorithmic preference

5. Homepage recommendations are ideologically “congenial”

Once bots were trained, the homepage became an echo chamber:

- Right bots got mostly right content

- Left bots got mostly left content

- Far-right bots saw the most ideological consistency

BUT: the algorithm still injected occasional cross-cutting content

Left bots saw some Right videos

Right bots saw almost none from the Left

6. Autoplay pushes right-leaning users deeper

Researchers followed YouTube’s “up-next” trail - autoplay without user clicks

- For right-leaning bots:

Recommendations got more ideologically extreme as autoplay continued

Exposure to “very-right” videos increased by 37% over 20 recs

- For left-leaning bots:

Exposure stayed flat

No meaningful rise in Far Left content

Autoplay = passive radicalization path… but only in one direction

7. Problematic channels do show up - but not in huge numbers

Researchers used existing lists of “problematic” channels (alt-right, conspiracies, QAnon, etc.)

Results:

- These made up only ~2.5% of total recommendations

BUT: over 36% of users encountered at least one

Very-right bots saw them the most, especially deeper in the autoplay trail

8. Creators don’t need a strike to lose reach

YouTube doesn’t need to ban you to bury your content

The papers show 3 ways your reach can vanish:

- You’re not shown on the homepage
- You’re excluded from autoplay trails
- New users never discover your niche

No warnings - No flags
Just algorithmic invisibility

9. Even mainstream news channels weren’t safe

They trained bots only using Fox News or MSNBC videos

Guess what?

Fox-trained bots saw recommendations from alt-lite (hard right) and conspiracy channels

MSNBC-trained bots mostly saw mainstream or neutral content

Same starting point
Different algorithmic future

10. Bottom line:

YouTube isn’t banning political content - it’s nudging the narrative

Its algorithm favors:

- Center & Left news by default
- Suppresses Far Left & Far Right
- And pulls users away from right-leaning niches faster and more aggressively

Your content isn’t broken
The playing field is

tl;dr:

YouTube isn’t outright banning political content

It’s shaping what people see - subtly, consistently, and at scale

→ Left and Center content is algorithmically favored
→ Right-leaning creators face faster decay and limited discovery
→ The deeper you go, the more filtered your feed becomes

No strikes or bank, just "silence"

If you wanna look deeper in the research papers:

1. Auditing YouTube's Recommendation System for Ideologically Extreme Content" (2022) - Authors: Magdalena Wojcieszak, et al.

2. YouTube's Recommendation Algorithm Is Left-Leaning in the United States (2023) - Authors: Manoel Horta Ribeiro, et al.

If you got up to this point and still have not followed, you are about to experience 7 years of extreme bad luck

See you in tomorrow's thread

And as always, join the newsletter for more AI x YouTube content - http://www.ytdojo.io