Why Media Planning and Media Buying on Meta No Longer Work the Way They Used To

For a long time, Facebook Ads felt like a system you could reason with.

If you worked in performance marketing around 2014–2018, you were trained to believe that good planning and disciplined buying were the primary drivers of results.

You defined audiences.
You separated funnels.
You controlled bids, placements, and budgets.

When performance moved, it usually moved in ways you could explain.

That belief held because Facebook Ads executed instructions deterministically.
Human decisions sat at the centre of delivery.

Under the hood, delivery logic was simple and predictable. Audience rules constrained who could see an ad, bids resolved competition, and budgets capped exposure. Planning reduced uncertainty up front, while buying corrected deviations after launch.

For context, imagine a Germany-based ecommerce fashion brand that started running Facebook Ads around 2015. The team was small, performance-led, and selling across Germany first, then expanding into Austria, Switzerland, and the wider EU. Over time, they built a familiar Meta growth engine: acquisition for Germany, incremental expansion for DACH, then EU scale once fulfilment, returns, and customer service matured.

 

The Facebook Ads Era: How Planning and Buying Actually Worked

To ground this shift properly, imagine a European ecommerce fashion brand that began advertising on Facebook around 2015.

They sold apparel across multiple EU markets and scaled gradually.

In this Germany-based version, early traction came from predictable, repeatable patterns: seasonal collections, sale cycles, and product drops that naturally aligned with Meta’s demand curve. The account evolved in phases: Germany-first prospecting, Germany retargeting for cart abandoners, then separate EU market ad sets once shipping and returns were stable.

Media planning for this brand meant

• defining target customers by age, gender, and interests
• building country-specific interest stacks
• creating lookalike ladders from purchasers
• separating prospecting and retargeting
• structuring campaigns by funnel stage

The plan assumed that audiences defined demand. If the audience was correct, delivery would follow.

In practice, Germany planning was often more conservative and structured: tighter exclusions to protect frequency in DE, separate spend buckets for prospecting versus remarketing, and strict market splits because performance reporting and merchandising were run market by market.

Media buying for this brand meant

• manual bid changes to control CPA
• placement selection to protect brand quality
• daily budget reallocation toward winners
• frequent intervention to stabilise performance

Buying assumed that execution controlled outcomes. When results drifted, buyers stepped in and corrected them.

This worked because the platform behaved like a rules engine.
If the planner defined who, the buyer decided how much, and Facebook delivered inside that box.

The important point is that planning decisions constrained delivery early, and buying decisions resolved performance late. Humans controlled both ends of the system.

For Germany, this also meant predictable operating rhythms: Monday catch-up optimisations after weekend sales, heavier budget pushes around salary cycles and seasonal moments, and strict CAC guardrails because return rates and margins in fashion are operationally sensitive.

 

When the Same Brand Started Seeing Something Break

From around 2022 onward, this same fashion advertiser began experiencing patterns that no longer matched the old logic.

• Broad campaigns outperformed segmented ones
• Bid changes stopped moving performance reliably
• Creative variety mattered more than audience precision
• Scaling felt unstable despite “best practice” setups

The product had not changed.
The markets had not changed.
The team had not forgotten how to plan or buy media.

Operationally, instability showed up in subtle but consistent ways. Learning phases reset without obvious changes. Identical audiences behaved differently week to week. Ads that had not degraded creatively lost delivery without warning. Cost volatility appeared even when competition and demand were stable.

What changed was how Meta decides what to show.

A few “in-the-account” signals made the shift obvious to anyone watching closely:

• broad targeting began beating previously reliable interest stacks
• simplified structures started outperforming complex segmentation
• creative fatigue accelerated, with concepts burning out faster than before

Those weren’t random trends. They were symptoms of a new delivery architecture.

In Germany, the shift was amplified by the reality of signal quality and privacy constraints. As tracking clarity weakened and consent dynamics varied by browser and environment, deterministic audience assumptions became less reliable. What looked like “random volatility” in the dashboard often reflected the system simply having less deterministic input to work with.

 

The Real Shift: Meta No Longer Executes Your Plan

This is the most important mental reset.

Meta advertising no longer starts with your targeting.

It starts with its own interpretation of relevance and probability.

The role of the advertiser shifted from decision-maker to signal provider. Targeting, bids, budgets, and objectives stopped acting as instructions and started acting as inputs the system interprets.

This is why media planning and media buying feel fundamentally different today. Decisions that used to be made by humans are now resolved earlier, automatically, and at a scale humans cannot intervene in.

For the Germany-based fashion brand, this was the turning point: the team could still “do everything right” according to the old playbook, yet performance would not respond in the same predictable way. The platform was no longer executing their plan. It was interpreting their signals.

 

What Meta’s AI-Powered Ads System Actually Is

Meta advertising today is not a traditional auction engine with automation layered on top.

It is an AI-powered ads delivery system designed to make decisions probabilistically, at the impression level, using prediction and feedback rather than fixed rules.

Instead of executing a predefined media plan, the system continuously answers four questions:

• which ads are relevant enough to consider
• which of those ads is most likely to deliver value
• what happened after the ad was shown
• how future decisions should change based on that outcome

This loop runs continuously, across all surfaces, markets, and users.

To make this work at scale, Meta separated delivery into two distinct decision layers.

This separation is also why Meta Ads is no longer an “open manual optimisation environment” in the way it used to feel. Performance increasingly depends on understanding how the system evaluates inputs and learns patterns over time, not on how fast a human can tweak levers.

For a German ecommerce business, that learning loop also has an operational consequence: when you scale spend, you are not just scaling reach. You are scaling training data. That makes the interaction between conversion quality, returns, margins, and signal depth far more important than it used to be.

 

The Two Core Engines Inside Meta’s AI-Powered Ads System

The system is anchored around two real models, each responsible for a different part of the decision process.

They operate sequentially, not in parallel.

This is not a detail. It is the core reason media planning and buying feel different.

In the Facebook Ads era, the “auction” was the main story.
In the AI era, the auction is still there, but it is no longer the main decision-maker.

Most meaningful decisions now happen before the auction and inside the models.

A simple way to understand the two-engine structure is:

• Engine 1 decides eligibility
  Which ads from your account are even allowed to compete for a user in a moment
• Engine 2 decides priority and sequence
  Which eligible ad should be shown now, and how the next impression should be shaped based on what the system learns

If you only see Meta as “auction plus bidding,” modern performance will feel random.
If you understand eligibility, priority, and sequence, modern performance becomes explainable.

What matters is that these engines are not isolated. They form a loop:

→ Andromeda retrieves what can be considered
→ GEM decides what should happen next
→ outcomes train GEM’s predictions
→ GEM’s predictions feed back into what Andromeda retrieves in future moments

So “retrieval” and “ranking” are not two separate optimisations. They are one connected decision cycle.

For the Germany-based fashion advertiser, this is why the same creative can behave differently across time and markets: the engine is not just scoring an ad, it is learning sequences, context, and downstream outcomes based on the signals it receives from each market environment.

 

Andromeda: The Eligibility and Retrieval Engine

Andromeda is responsible for retrieval, not optimisation.

Its job is to decide which ads from a single advertiser are relevant enough to be placed into a shortlist for a user, right now.

That sounds simple. It is not.

Because what Andromeda is doing is replacing the old question, “Which audience did the advertiser select?” with a new question:

Which of this advertiser’s active ads best match what this user appears to care about in this moment?

Andromeda is effectively building a live shortlist from your account, impression by impression.

What “eligibility” actually means in practice

Eligibility is not a binary rule like “in audience” or “not in audience.”

It is a probability-weighted inclusion decision.

An ad can be:

• retrieved very frequently for a certain context
• retrieved rarely for another context
• retrieved almost never if the signals are weak or repetitive

So even inside one advertiser account, not all creatives have equal access to delivery.

This is why advertisers often say:
“Meta is not spending on my ad even though the targeting is fine.”

The targeting can be fine. The ad is not being retrieved.

Inside the fashion brand’s account

At any moment, the brand may be running:

• a seasonal collection campaign
• a sustainability-led message
• a price-driven promotion
• a styling or lifestyle narrative

In Germany, this might map to very practical business moments: winter outerwear drops, back-to-office cycles, sale periods, or premium capsule launches where brand tone matters as much as conversion rate.

Andromeda evaluates those creatives and asks if each one should enter consideration for the specific user.

To do that, it analyses:

• the visual meaning of the creative
• the semantic meaning of the copy
• product category and usage cues
• recent on-platform behaviour

This is not just “who likes fashion.”

It’s closer to:

• what style signals is the user responding to
• what shopping mode the user appears to be in
• what content pattern the user is engaging with
• what intent-like behaviour is emerging right now

What Andromeda changed in delivery

This was Meta’s first major AI overhaul because it flipped the logic from audience-first to creative-first matching.

With Andromeda, Meta increasingly evaluates:

• visuals, themes, hooks, language
• format signals (what type of asset this is)
• prior engagement patterns tied to similar signals

That is why the platform began rewarding setups with a larger opportunity pool. Broader campaigns with more creative inputs give the retrieval layer more options to match users and still satisfy campaign goals.

 

Creative Diversification as a Retrieval Mechanism

For the ecommerce fashion brand, this explains why creative variety became a structural requirement.

A trend-led visual, a sustainability narrative, a discount-driven message, and a styling-focused creative are not small variations.

They are distinct retrieval signals.

Each signal opens a different pathway into eligibility.

• trend-led signals connect to discovery and novelty behaviour
• styling signals connect to consideration and “how it looks” behaviour
• sustainability signals connect to values-based evaluation behaviour
• discount signals connect to price sensitivity and urgency behaviour

In Germany, diversification also protects the business from over-dependence on a single conversion narrative. If one message saturates, the account still has other eligibility corridors that can continue retrieving demand without forcing a full reset.

If the brand repeats one dominant message:

• retrieval narrows
• exploration stalls
• delivery plateaus

If the brand supplies multiple, clearly differentiated concepts:

• retrieval expands
• more demand pockets open
• delivery stabilises

Creative diversification is not experimentation.
It is how the system maps demand.

This also explains why “creative fatigue” started to feel faster for many advertisers. If the retrieval layer learns a concept is saturated for the audience contexts it maps to, eligibility can decline sooner than teams expect, even before classic performance indicators visibly degrade.

 

GEM: The Ranking and Decision Engine

Once Andromeda produces a shortlist of eligible ads, GEM (Generative Ads Model) takes over.

GEM’s role is to decide which ad should actually be shown, impression by impression.

GEM is not looking at one ad in isolation.

It is comparing eligible options and estimating which one creates the best expected value right now.

What GEM is actually ranking

GEM is essentially running a probability and value calculation for each eligible ad.

It blends:

• Estimated Action Rate
• expected value of the action
• creative and format performance patterns
• user experience signals
• pacing and competitive context

The key shift is that the system is not obeying fixed rules.

It is making probabilistic tradeoffs based on what it predicts will happen.

What “ranking” really means

In the old model, buyers assumed:
“If I set the audience correctly and bid correctly, delivery follows.”

In the ranking model, delivery is conditional on predicted outcomes.

Even if your bid is strong, the system can deprioritise you if it predicts:

• low action probability
• weak value signal
• inconsistent feedback history

That is why “perfect” buying control can fail.

Why GEM was a bigger shift than Andromeda

Andromeda answers: what can be shown.

GEM answers: what should be shown next.

That “next” is the part that changes media buying psychology. GEM is not only picking a winner in isolation. It is learning patterns across:

• ad sequences
• formats
• messaging arcs
• what users do organically versus after ad exposure
• what tends to work in combination across time

So ads are increasingly evaluated inside broader contextual journeys, not as single isolated impressions.

GEM also feeds predictions back into retrieval. In practical terms, the ranking system learns what combinations and sequences work, and that learning influences what Andromeda is more likely to retrieve for similar contexts in the future.

For the Germany-based fashion advertiser, this is why constant small edits started producing weaker results. If the system is learning longer-term journey patterns, frequent resets interrupt the very pattern recognition it is trying to build.

 

How the Two Engines Changed Planning and Buying Together

This separation of retrieval and ranking is the structural change most teams miss.

Planning now influences eligibility through creative and signal design.
Buying now influences ranking stability through pacing and restraint.

In the Facebook Ads era:

• planning defined who could see an ad
• buying decided which ad would win

In the current system:

• planning shapes which ads can be considered
• buying protects the system’s ability to rank accurately

If planners don’t supply diverse eligibility signals, ranking has no good options.
If buyers over-control delivery, ranking loses the room to learn.

This is also why fast testing cycles and constant edits feel less reliable now. When long-term patterns matter more, frequent resets can interrupt pattern recognition and cause the system to relearn basic relationships instead of building momentum.

 

Declared Intent vs Latent Intent

Traditional planning assumed intent was declared:

• interests
• demographics
• funnel stages

Meta’s AI-powered system optimises for latent intent:

• inferred readiness
• contextual timing
• behavioural signals

This explains why interest stacking weakened, funnel splits fragment learning, and impression-level optimisation outperforms stage-based planning.

Planning moved from asking who the user is to understanding where the user is in intent space.

 

The Feedback Loop That Determines Scale

GEM learns only from outcomes it can observe.

For European ecommerce advertisers, this increasingly depends on:

• Conversions API
• server-side purchase confirmation
• first-party transaction data
• value signals

If feedback is delayed, incomplete, or flattened, probability estimates degrade. That degradation shows up as volatile performance, unstable scaling, and sudden delivery drop-offs.

This is not a tracking detail.
It is the quality of the learning signal.

This is also why conversion volume and consistency matter more than ever. Without enough stable outcome data, the system struggles to detect trends, sequence effects, and pattern reliability.

 

False Efficiency: The New Failure Mode

By default, the system optimises toward Estimated Action Rates.

For an ecommerce fashion brand, this means that without strong value signals, the system may prioritise:

• discount buyers over full-price buyers
• low-margin orders over high-margin ones
• one-time purchasers over repeat customers

Platform metrics can improve while business quality declines.

This is false efficiency.
The system is doing exactly what it is trained to do.

 

Budget Liquidity and Learning Stability

Probabilistic systems require room to explore.

When the fashion brand fragments budgets by country, funnel stage, or product category, learning density collapses. Each split reduces the system’s ability to compare outcomes and adjust probabilities reliably.

When budgets are consolidated, eligibility improves, ranking stabilises, and scale becomes sustainable.

Liquidity is not about spending more.
It is about allowing the system enough freedom to learn.

Budget also behaves like a signal. If daily budgets are too low relative to the conversion event, the system cannot generate enough consistent outcome data per learning cycle to detect patterns. High-intent events like purchases generally require more spend per learning cycle than upper-funnel actions like clicks or engagements, simply because the outcome is rarer and noisier.

For the Germany-based advertiser, this is often where finance and performance collide: low budgets feel “safe” in the short term, but they restrict the system’s ability to learn, which can make performance less stable and scaling harder when the business actually needs growth.

 

What Media Planning and Media Buying Are Now

Media planning now means

• defining outcomes the system should learn toward
• supplying diverse, intent-rich creative signals
• ensuring clean, value-aware feedback
• designing structures that preserve liquidity

Media buying now means

• pacing rather than steering
• restraint rather than control
• protecting learning rather than forcing outcomes

Execution skill matters less than system understanding.

In practice, this shift also changed how experienced teams read performance:

• less focus on daily spikes
• more focus on rolling windows (several days)
• less reactive editing
• more emphasis on stable learning periods where the system can recognise patterns rather than constantly resetting them

 

The Reality

Meta advertising is no longer a platform where humans guide delivery directly.

It is a probabilistic system where:

• Andromeda decides eligibility
• GEM decides prioritisation
• outcomes train future decisions

Media planning and media buying still matter.
They now operate above the system, not inside it.

That is the structural shift most teams are still struggling to internalise.