How Meta's Machine Learning Optimises Ad Delivery

Meta's ad system runs on machine learning that touches every part of delivery — who sees your ad, when, where, and at what bid. For UK advertisers, understanding how that ML actually works is the difference between fighting the system and feeding it.

This post explains the inputs, outputs and feedback loops behind Meta's ML, and what you can do to make it work in your favour.

The basic flow

Every time someone opens Facebook, Instagram, Messenger or an Audience Network app, Meta's system runs an auction in milliseconds. For your ad to win, the ML system has to:

1. Predict whether this user will take your desired action
2. Compare that prediction against your bid
3. Compare your "total value" against every other advertiser's
4. Decide which ad to show

The first step — predicting user action — is where ML lives. Everything else is auction mechanics.

What signals the ML uses

Meta's predictions are built from dozens of signals, including:

• The user's recent app and site behaviour
• Their interaction history with similar ads
• Profile data (age, location, declared interests)
• Device, OS, time of day
• Connection speed
• Engagement patterns by surface (Feed, Reels, Stories)
• Historical conversion rates for similar audiences
• Pixel and CAPI events from your site
• Catalogue data and product matches
• Creative format and historical performance

It then weighs these against the predicted likelihood of your specific action — purchase, install, lead, view.

The learning phase, explained

When you launch a new campaign, Meta marks it "Learning." This means the ML doesn't yet have enough feedback to confidently predict. During this window:

• Delivery is more exploratory
• CPA is more volatile
• The system tests across different audience slices
• Bidding is less precise

The campaign exits learning when it gets ~50 optimisation events in a 7-day window. After that, the ML has stable patterns and CPA tightens.

If you edit anything significant — budget, creative, audience, bid strategy — the learning resets. That's why "do not touch in week one" is the most important advice in Meta advertising.

How the ML uses your pixel data

Every pixel event you fire becomes training data for Meta's ML. The more events the system sees, the better its predictions get for users like the converters. This is why:

• A pixel firing 1,000 Purchase events/month outperforms a pixel firing 50/month
• CAPI is critical (it recovers events lost to ad blockers and iOS)
• Sending value parameters lets the ML optimise for revenue, not just count
• Custom events let you signal quality (e.g., "demo booked" vs "form fill")

Estimated Action Rate (the hidden number)

Inside the auction, every ad gets an Estimated Action Rate (EAR) — Meta's prediction of how likely a specific user is to take your desired action if shown your ad. EAR is multiplied by your bid to get "total value." Higher EAR = win more auctions at lower bids.

Things that boost EAR:

• Strong creative variety
• Relevant copy
• Recent positive engagement on similar ads
• Clean tracking signal
• Good ad quality scores

Things that lower EAR:

• Creative fatigue
• Bad feedback (hides, complaints)
• Slow landing pages
• Mismatched audience signals

This is why two ads with the same bid can have wildly different reach.

The feedback loop

Meta's ML improves continuously based on three feedback channels:

1. Conversion events (pixel, CAPI, SDK)
2. Engagement signals (clicks, dwell time, hides, reactions)
3. Quality signals (page experience, complaints, ad quality reviews)

The ML uses this feedback to refine predictions for everyone in your funnel — and adjacent audiences with similar patterns.

FAQ

How long does it take Meta's ML to "learn" my product?

Roughly 7-14 days for a healthy account, longer if you're under 50 conversions per week.

Why do CPAs spike when I edit a campaign?

Editing usually resets the learning phase, returning the campaign to exploratory delivery.

Does Meta's ML know what's in my image?

Yes. Computer vision models scan images for objects, settings, colours, faces and brand elements. This affects who sees the ad.

Does the ML penalise certain content?

Yes. Complaints, hides, low-quality landing pages, misleading claims and broken policies all reduce delivery.

Can I see what the ML thinks of my ad?

Indirectly — via the Quality, Engagement and Conversion ratings in Ads Manager. They're aggregated indicators of EAR and feedback.

Can I bypass Meta's ML?

No. Even manual targeting still runs through ML for delivery decisions.

What's the best way to feed the ML?

CAPI, value-based optimisation, fresh creative variety, and avoiding edits in the learning phase.

What you actually control

You can't tweak the ML directly, but you do control:

• Creative variety and refresh frequency
• Tracking quality and signal completeness
• Bid strategy and budget
• Country, language and basic audience constraints
• Landing page speed and experience

Everything else is the system's job. Trying to override it with manual fiddling usually costs more than it saves.

Pix-Vu and the ML

The single biggest thing you can give Meta's ML is creative variety. Pix-Vu generates 20+ ad variations from one upload, formatted for every placement. More variety means the ML can find what works faster, exit learning sooner, and stabilise CPA. Try it at pix-vu.com.