If you’ve spent the past two years feeling like the algorithm has become unpredictable, erratic, or downright hostile, you’re not imagining it. But you’re also not interpreting the situation correctly.

What most creators call “the algorithm” is not an algorithm anymore.

Not in the old sense.

Not in the way social media once worked.

What Meta is running today is a fully rebuilt AI discovery architecture, one that has been quietly evolving since 2022. It has more in common with Google’s ranking engine than with the early Instagram feed. It behaves like TikTok, YouTube and Netflix, not like the social networks we grew up on. It interprets content the way humans interpret meaning, not the way machines processed hashtags in 2016.

The truth is: Instagram and Facebook no longer show you content based on who you follow. They show you content based on what the system predicts will keep you engaged. That prediction isn’t a guess. It is a complex chain of AI processes stitched together into a discovery engine that is constantly learning, recalibrating and refining itself.

Let’s go inside that engine — the architecture almost no one in the industry has taken the time to actually understand — and trace how Meta rebuilt its entire ranking pipeline from the ground up.

The reconstruction started in 2022 — and almost no one noticed

To understand today’s system, you need to go back to the moment the shift began. In early 2022, Meta started hinting — quietly, indirectly — that the old model of the social graph was no longer sufficient. This was the year TikTok’s growth forced every major platform to confront a difficult truth: users weren’t loyal to content from their friends. They were loyal to whatever the machine recommended next.

In earnings calls and investor meetings, Meta began using new language:

“unconnected content,”

“AI-powered recommendations,”

“interest-driven surfaces,”

“relevance modelling.”

These weren’t marketing phrases. These were signals that Meta had already started laying the bricks for a completely new engine that would sit beneath every part of the user experience — Feed, Explore, Reels, Ads, Shops, Recommendations, and later, Advantage+.

By late 2023, developer transparency notes started referencing the internal components: the Discovery Engine, multimodal embeddings, user embeddings, interest learners, and ranking systems that sounded more like something you would find in a machine learning research lab than a social media product.

By 2024, it became impossible to deny: Meta’s architecture had moved away from chronological or network-based systems and into a fully predictive AI-driven discovery framework.

And then in 2025, Andromeda launched — the final sign that the rebuild was complete.

But most people never connected the dots.

Meta’s discovery engine: the new heart of social media

The Discovery Engine is the beating heart of the new system. It is the component responsible for deciding what any individual sees at any moment. Unlike traditional feed ranking systems, the Discovery Engine doesn’t rely on your social relationships — it relies on probability modeling.

Its purpose is simple but transformative:

to predict which piece of content will keep you on the app longest.

And it can’t make that prediction using simple engagement metrics. It needs a deep, textured understanding of both users and content — deeper than human reviewers or manual tagging could ever provide. This is why Meta moved to AI-led embeddings.

To understand relevance today, you need to look at what the Discovery Engine is actually ingesting and interpreting.

User embeddings: the machine’s model of you

Let’s start with the user side of the equation, because this is where Meta’s approach is the most misunderstood.

A user embedding is not a profile.

It’s a mathematical representation of your behaviour, preferences, emotional patterns, viewing habits, micro-engagement signals, and temporal routines.

It tracks:

• the kind of content you respond to,
• the pace you scroll through the app,
• your emotional appetites at different times of day,
• the audio energy you prefer,
• the way you behave when you’re tired,
• how you engage in private (saves, rewatches, shares),
• the topics you drift toward or away from,
• the formats you abandon quickly,
• your average “attention window,”
• your tolerance for slow or fast content.

This embedding updates constantly — not weekly or monthly, but multiple times per session. It’s a rolling behavioural fingerprint.

The system isn’t interested in who you follow. It’s interested in what you are primed for. And it interprets your behaviour far more accurately than you could describe it.

This alone breaks every old rule of social media.

Multimodal content understanding: how Meta reads your Instagram and Facebook posts

This is where creators underestimate the system the most. To match content to users, Meta must understand the content with extraordinary precision. And so they’ve built multimodal AI models that consume every possible input:

The visual layer

It detects objects, environments, lighting, micro-expressions, motion, scene composition, colour palettes, spatial cues, dominant themes, and visual “energy signatures.”

The audio layer

The system parses tone, pace, emotional weight, music intensity, speech patterns, and even the type of background noise that correlates with user preference clusters.

The textual layer

It reads on-screen text, captions, voice transcripts, and keywords — but not the way a keyword engine does. It interprets sentiment, intent, coherence, and semantic clarity.

The contextual layer

It looks at your past posts, your niche, your message consistency, your brand signals, your funnel patterns, your keyword drift, and even the behavioural reactions your content historically triggers.

All of this becomes a content embedding — a structured mathematical object that represents the meaning, tone, purpose, and emotional signature of your post.

This is why copy-paste content no longer works.

Why niche-hopping kills reach.

Why inconsistent brands flatline.

The machine needs clarity to match you accurately.

And inconsistency tells the machine:

“I don’t know who this is.”

And if the platform doesn’t know who you are, it cannot show you widely.

Embedding-based ranking: the prediction layer

Once the system has both embeddings — the user and the content — the ranking engine performs a similarity match. This is where the Discovery Engine becomes frighteningly powerful.

It compares the vector representation of the content embedding to the vector representation of the user embedding and predicts the probability of:

• attention
• completion
• saving
• rewatching
• sharing
• following
• clicking
• meaningfully engaging
  

This isn’t a simple “does this user like this topic?” question.

It’s “Does this content align with this person’s behavioural and emotional state at this exact moment?”

That is the sophistication that makes modern discovery so addictive.

It’s also why content feels like it finds you now.

Creators are no longer pushing content outwards.

Content is being pulled inwards toward users by the machine.

Meta interest learner: the adaptive brain

If embeddings are the system’s memory, Meta Interest Learner is the part that learns in motion.

MIL monitors how users evolve over time and adjusts their interest profiles dynamically. It identifies micro-shifts in preference — your sudden interest in self-development posts, your temporary obsession with dopamine detox content, your two-week phase of watching morning routine videos at 6am, your evening preference for lower-energy scrolls.

MIL doesn’t wait for stable patterns.

It predicts emerging ones.

This is why your feed changes so dramatically over days or weeks.

The system isn’t reacting.

It’s anticipating.

MIL ensures the Discovery Engine is not static. It keeps recalibrating what a user wants before they consciously know they want it.

The paid layer: Advantage+, creative matching, Andromeda

One of the most profound changes — and one marketers have barely begun to understand — is that paid and organic are no longer separate ecosystems.

Advantage+ uses the same relevance and prediction models as organic discovery. Ad creative is evaluated semantically and emotionally before budget is even considered. Weak organic signals weaken ads. Strong organic signals improve ads.

And then came Andromeda — the launch that quietly marked the completion of Meta’s cross-platform discovery unification.

Andromeda links together:

• user embeddings
• content embeddings
• interest prediction
• ranking surfaces
• organic signals
• paid signals
• long-term engagement patterns
• cross-format behaviour (Feed, Reels, Stories, Explore)
• ad prediction models

It is the backbone of the entire Meta ecosystem.

Once Andromeda went live, Instagram and Facebook began behaving like one unified recommendation machine instead of separate apps.

Engagement still matters, but only as training data

Here’s the part that confuses creators:

engagement matters, but not in the way they think.

It no longer determines the success of a post.

It trains the system.

A save becomes a signal of meaning.

A rewatch becomes a marker of emotional alignment.

A share becomes a marker of relational strength.

A long view becomes a signal of message clarity.

A fast scroll becomes a signal of mismatch.

But none of these numbers are judged in isolation.

They’re interpreted through patterns:

Who saved?

• Who watched twice?
• What type of user reacted?
• What behavioural profile matches these reactions?
• What emotional signature is being reinforced?
• Which cluster does this push you into?
• How does this alter your embedding?

This is why the same engagement number can have wildly different outcomes for different creators. The Discovery Engine isn’t judging quantity. It’s interpreting meaning.

The result: relevance is the new currency

In today’s architecture, relevance is everything.

Relevance determines reach.

Relevance determines ad performance.

Relevance determines discovery.

Relevance determines whether the platform sees you as “a good match” for audiences it wants to retain.

Relevance is not a metric Meta gives you — because they cannot afford for you to game it. But it is the closest social analogue to Google’s early ranking system.

You could call it a relevance score.

Meta just doesn’t show it to you.

But this is the logic that drives the entire Discovery era — and most creators have no idea it exists.

So why does this matter

Because SDO Is the Discipline That Reveals the Invisible Layer

Everything inside Meta’s new system — every embedding, every algorithmic inference, every behavioural prediction — points to one unavoidable fact:

You are no longer creating content for your followers.

You are creating content for the discovery engine’s ability to categorise you.

And that is exactly why Social Discovery Optimisation (SDO) exists.

SDO is the framework that helps creators understand:

• what the model thinks they are,
• why their content is being matched (or mismatched),
• when they drift off-topic or off-emotion,
• why certain posts spike and others tank,
• when their relevance erodes,
• how to strengthen semantic clarity,
• how to align emotionally with their ideal audience,
• how to sequence posts through the funnel the machine already understands,
• how to become predictable — in the way the algorithm rewards.

It’s not about gaming the system.

It’s about training it.