Why Meta Seems To Prioritise Some Of Your Ads More Than Others

TL;DR: Why And How Meta Chooses Which Ad Creatives To Push Hardest

‍

When you're running ads on its platforms, Meta's job is to maximise the results you obtain with the budget you give it. To do this, it employs a sophisticated AI system making millions of decisions about which of your ads people are most likely to find relevant and engaging, and that are most likely to drive the actions you're seeking: Andromeda which finds the right ads to show; GEM which helps Meta understand people's interests and behaviours so predictions can be made about the kinds of ad content they're likely to respond to; and ranking models, including Lattice which brings together information from different advertising activities to make better decisions about where to serve your ads (Facebook or Instagram) and in which format (Feed post or Reel).

‍

"Why Have We Wasted All This Time On Ad Creatives For Meta To Just Ignore Them?" (The Cry Of Advertisers Everywhere)

‍

We get it.

‍

You've invested hours into developing the creatives for your upcoming Facebook ad campaign - the graphics, photographs, and videos you plan to use to get attention and drive action - whilst writing the copy (words) that will accompany them, and so it's more than a little frustrating when it becomes obvious that Meta is pushing all your budget to one or two ads, whilst starving the others.

‍

We've all been there:

‍

You build your campaign in Meta Ads Manager, choosing your Objective and Performance Goal, setting a budget, and selecting your target audience. You then load the various creative assets you've lovingly crafted into your Ad Set, populate the Primary Text, Headlines, and Descriptions, then choose a Call To Action button.

After a final check to make sure there are no typos or structural issues that could prevent your ads from running, you hit 'Publish', then sit back and wait for Meta to review your campaign for compliance with its advertising policies and make sure it's technically sound and eligible to run.

‍

With your ads approved, you watch for early signs of progress, and you start to see your ads clocking-up impressions and clicks.

‍

Then, after a few days, you spot that four out of six ads are doing nothing for you.

‍

They're getting some spend, but hardly any impressions and few, if any, of the results you're looking for. It's a shocker, because you secretly had your own favourites and they're among the apparent 'flops'.

‍

It can feel like you've wasted a lot of time and effort, whether this is your first go or you're a seasoned pro.

‍

And it can almost seem arbitrary - especially when you've supplied two very similar versions of a graphic, for instance, and one gets prioritised whilst the other is starved of spend.

‍

However, as frustrating or mystifying as it may seem, it's actually the best thing for your campaigns if you want them to have the best chance of success.

‍

Choice Not Chance: You Guess, Meta Knows Best

‍

‍

If you're investing in ads on Facebook and Instagram, you need then to work, whether that's getting traffic to your website, generating leads, making sales, or securing donations (if you're a charity).

‍

When you're busy developing your creative assets, your focus is generally going to be on:

• Making your ads look appealing
• Conveying your message or offer

But, here's the thing: you, and literally everyone else who has ever devised a paid social ad to run on Meta's platforms, are largely going off instinct. And your decisions are subconciously influenced by your cognitive biases, whether you're making font choices or image selections.

‍

It's true, when you've run a lot of ads, your instincts become honed and you have examples of creative selections that have worked for you that you can adapt, and that give your work an edge.

‍

However, unlike Meta and the sophisticated machine-learning models it's built itself over the years, you lack something crucial: enough data to move your decisions from gut-driven to statistically-driven.

‍

This is exactly what's happening behind the scenes when you see Meta favour some ads over others.

‍

Meet Meta's 'Contextual Multi-Armed Bandit' Used To Identify Which Ads To Prioritise

Meta’s delivery algorithm is not designed to be a fair scientist, giving each of your ad creatives equal weight and opportunity; it is an economic profit maximiser.

‍

So, when you publish multiple ads inside an ad set, is objective is to extract the maximum number of desired actions for your budget, not just rotate your ads evenly.

‍

To do that, it uses an exploration-exploitation machine-learning model (a Contextual Multi-Armed Bandit). The moment an ad shows even a microscopic statistical edge, the algorithm aggressively shifts from "exploring" (testing variants) to "exploiting" (funneling spend into the frontrunners).

‍

Going off information that Meta has placed in the public domain, we know that it consists of at least these three components (although it's likely there's more to it that it it keeps secret for commercial reasons):

‍

Andromeda

‍

What it does: helps Meta sift through all the adverts available and identify those that are potentially most relevant to the person currently using Facebook or Instagram.

‍

GEM

‍

What it does: helps Meta understand people's interests and behaviour, using patterns in how they interact with content and adverts to predict which messages and images are most likely to catch their attention and prompt action.

‍

Lattice

‍

What it does: helps Meta bring together information from different advertising activities to make better decisions about which adverts are most likely to achieve the advertiser's objective, across its advertising ecosystem.

‍

‍

How Meta's Andromeda, GEM, and Lattice AI Models Work Together

‍

Imagine you're running a Christmas campaign as a small retailer selling handmade gifts.

You've supplied Meta with six different creative approaches:

• A product photograph on a white background

• A short product demonstration video

• A customer testimonial

• A Christmas gift guide

• A lifestyle photograph showing the product in use

• A limited-time Christmas offer

Each creative is effectively another potential option – or arm – that Meta can consider when deciding what to show someone.

‍

A traditional Multi-Amed Bandit problem involves choosing between several options, learning which ones produce the best results, and progressively favouring those that appear most successful.

‍

A Contextual Multi-Armed Bandit adds another dimension: the decision depends on the context in which the opportunity arises.

‍

For example, the best advert for someone browsing Instagram Reels on a mobile phone might be different from the best advert for someone scrolling through Facebook Feed. Someone who has previously engaged with gift-related content might respond differently from someone who has shown an interest in the product category itself.

‍

This is where the capabilities and sophistication of Meta's advertising system becomes evident.

‍

Rather than simply identifying one universally winning advert, the system is trying to predict which of your creative and copy selections is most likely to produce the desired outcome for each individual advertising opportunity, taking account of all the available contextual signals, in real-time.

‍

Now, imagine the scale of it all; Facebook and Instagram each have around 2 billion Daily Active Users (DAU) globally, all spending just over an hour a day scrolling (on average). Based on typical consumption rates, that means an active user will scroll past 300-600 separate pieces of content every day across both platforms combined.

‍

Every single one of the interactions users have with all this content is noted by Meta. So we're talking between 600 billion and 1.2 trillion interactions PER DAY. That's an awful lot of data that Meta is able to use to make decisions that are more scientific than any single advertiser could ever hope to do.

‍

Better still, the system can then use the outcomes of advertising interactions as further learning signals, improving future predictions and delivery decisions.

‍

In essence, it's an ongoing process of prediction, selection, observation and learning.

‍

‍

How Meta's AI Evaluates Your Ads In Practice

‍

‍

The system determines which ads get prioritised and which get starved across three distinct phases:

‍

1. Pre-Auction Evaluation (Before Serving Significant Impressions)

‍

It may sound odd, but the way Meta's ads work, it takes your budget then bids for impressions in an auction, where you're essentially competing against other advertisers for a share of the impressions available.

‍

Even before your ads enter competitive bidding, Meta processes the creative files through its internal multimodal AI models (Computer Vision, Natural Language Processing, and audio transcription). In doing so, it:

• Scans your creative
• Interprets the content
• Maps it against billions of past ads

The Prior Probability effect: If historical data across, say, the charity sector indicates that, relative to the others you've supplied, one style of your ad creatives typically generates higher dwell time (how long someone spends looking at an ad on screen) and lower bounce rates on the destinaton website receiving clicks from those ads, the algorithm begins with an inherently higher initial prior probability for that type of ad.

‍

Placement Liquidity (Asset Formatting): Ads that provide native aspect ratios across all high-volume placements (1:1 Feed, 9:16 Stories/Reels, 4:5 Mobile) receive lower auction entry barriers. If one creative fits cleanly across all placements without awkward letterboxing, Meta gives it early distribution priority because it is cheaper to place.  

‍

2. The First 200–500 Impressions: Micro-Engagement Signals

‍

Once the campaign goes live, Meta feeds a tiny sample of impressions into auctions containing users who match your target demographic. At this stage, Meta monitors real-time behavioural signals that occur long before users take any action:

‍

Thumb-Stop Rate / Dwell Time: When users scroll at normal speed through Facebook or Instagram, Meta's client apps measure whether their scroll decelerates over your ad. If users linger for 1.5 to 3 seconds on one of your ads instead of flicking past it in 0.4 seconds, it scores a major early boost in terms of Ad Quality.

‍

Intent-Weighted Click-Through Rate (CTR): Meta does not treat all clicks equally. A click from a user whose historical profile rarely converts is discounted. If one of your ads receives clicks from users whose behavioural profiles show frequent donations, e-commerce transactions, or high landing page dwell times, Meta immediately inflates that ad's Estimated Action Rate (EAR).

‍

Early Bounce Rate & Dwell on Destination: Meta tracks how quickly visitors bounce back from your landing page. If users who click a link in your ad immediately hit the "back" button because the landing page requires reading a long narrative, Meta penalises that creative's relevance score. If clickers land and immediately scroll or take the desired action on your page, the relevance multiplier surges.  

‍

3. The "Total Value" Auction Equation (Why the Bias Snowballs)

‍

Meta determines which ad wins each individual auction slot using this formula:

‍

Total Value = (Advertiser Bid x Estimated Action Rate) + (Ad Quality / User Value)

‍

The Feedback Loop: If one of your ads earns an early Estimated Action Rate of 2.5% and another scores 1.2%, the one with the higher EAR can win an impression auction with a significantly lower bid.

‍

The Winner-Take-All Trap: Because an ad with a higher EAR wins auctions more cheaply, the algorithm rapidly routes the next £10 of spend to it to preserve campaign efficiency. More spend brings more data, which confirms its high conversion rate, which drops its auction cost even further.

‍

Within 6 to 12 hours, Meta's machine learning determines that serving certain of your ads represents an unnecessary auction loss, effectively "starving" them down to pennies.

‍

‍

Meta Is Better At Picking Winners, So Let It

‍

So, there you have it. Meta favours some of the ads you upload in your Facebook or Instagram campaigns because it has so much more insight into what works than you or indeed anyone running ads on its platforms.

‍

There's simply no denying it; you'll design ads that you think will work, whereas Meta knows which ones will.

‍

Is it a problem that out of six creatives, Meta only chooses to push two?

‍

Not if you want your campaign to have the best chance of success, efficiently spending your budget where it can get the best results.

‍

Is it a waste of time developing multiple creative variations?

‍

No, not at all - think of it as integral to the whole "exploring → exploiting" process. Without the different options, neither you or Meta would stand much chance of learning anything useful.

‍

Trust the process, it works!

‍

Have you found this blog useful and informative? Please show it some love by sharing it on your socials and with anyone you thing would also benefit from it. Want hands-on help building and running ad campaigns on Facebook and Instagram? Just ask!

‍