Meta's Andromeda Is Making Your Ads a Black Box. Here's What We're Doing About It.

What Is Andromeda?

Andromeda is Meta's AI-powered ad delivery and ranking system. It uses machine learning to predict the probability that a given user will take the action an advertiser is optimising for — a purchase, a lead, a video view — and ranks ads in the auction accordingly. It's the engine behind Advantage+ Shopping Campaigns, Advantage+ Audiences, and the broader shift toward letting Meta determine targeting instead of advertisers building it manually.

The practical effect: you define a budget, a creative, and an objective. Andromeda decides who sees it.

The Black Box Problem

Before Andromeda and the shift toward broad AI delivery, advertisers had meaningful control. You could build structured Custom Audiences, create tight interest-based ad sets, run controlled creative tests against defined segments, and have reasonable confidence about what was driving performance.

With Advantage+ and Andromeda-powered delivery, much of that control is removed. You see costs, reported conversions, and a limited set of breakdowns. You can't easily answer:

• Which users is this ad actually reaching?
• Is my broad audience capturing people who would have converted anyway (cannibalising organic conversion)?
• Is the system optimising for the users most likely to take the intended action, or the users most likely to interact cheaply?

Meta's reporting shows aggregated outcomes. The underlying delivery logic is opaque. That's the black box.

The Commercial Incentive Problem

Meta's commercial incentive is to sell impressions efficiently. Advantage+ campaigns, which give Meta more control over delivery, allow the system to explore a wider audience and capture more auction inventory. This is, structurally, better for Meta's revenue than tightly controlled manual campaigns.

This isn't a conspiracy. It's an alignment problem. The objectives of the algorithm are not always identical to the objectives of the advertiser. Understanding this doesn't mean you should avoid Advantage+ — it means you should test it critically rather than accept the default framing.

When Broad Targeting and Advantage+ Works Well

Broad targeting and Advantage+ deliver strong results when:

• You have substantial purchase history and pixel data for the algorithm to model from (hundreds of conversion events per month)
• Your product has mass-market appeal and you're in an acquisition-first phase with high creative volume
• Your creative is strong, varied, and updated regularly — the algorithm has something to work with
• You're running enough conversion volume for Smart Bidding to function reliably (Meta's own recommendation is 50+ conversion events per ad set per week)

When the Old Way of Targeting Still Wins

We're still running manual audience structures and outperforming broad Advantage+ for a specific type of client. That type is:

• Niche products or services with a small, well-defined target audience that the algorithm doesn't have enough data to infer
• High-ticket B2B or specialist services where the conversion volume is low and Smart Bidding can't build a reliable model
• Clients with strong first-party data — email lists, past purchasers, warm social followers — where the audience definition is itself the targeting intent
• Re-engagement campaigns where you want to reach a specific, defined group rather than a lookalike expansion

For one client in a specialist professional sector, a tight Custom Audience of previous enquirers combined with a lookalike from a 1,200-person buyer list consistently outperforms any Advantage+ setup we've tested. The algorithm simply doesn't have a large enough signal pool to infer who this buyer looks like from general platform data.

What We Actually Do

Our approach is to test both, measure incrementally, and not assume that broad wins because Meta recommends it. We run structured experiments — Advantage+ versus manual, broad versus defined audience — for the same offer and creative. We measure by actual revenue against actual spend, not by Meta's reported ROAS (which has its own reliability issues without robust CAPI and Consent Mode setup).

We also apply scrutiny to what's inside Advantage+ reporting. Breakdowns by age group, placement type, and device give partial signal about where delivery is actually going. It's not the full picture, but it's better than accepting the aggregate number at face value.

Frequently Asked Questions

What is Meta Advantage+ Shopping Campaigns?

Advantage+ Shopping is Meta's AI-automated campaign type for ecommerce. Instead of manually defining audiences, budgets per ad set, and placement, the system makes most of those decisions automatically. It can perform well for ecommerce advertisers with strong purchase history, but removes significant control from the advertiser.

Should I use Advantage+ Audiences for my campaigns?

If you have high conversion volume and a broad potential audience, Advantage+ Audiences is worth testing. If you're selling a specialist product or service, or your conversion volume is low, manual Custom Audiences and Lookalikes from well-defined seed audiences will typically perform better. Run a structured test to find out which applies to your account.

Why does Meta report more conversions than my actual orders?

Meta's default attribution window includes 1-day view-through conversions — someone who saw your ad without clicking and later purchased. This inflates reported conversions significantly. Set your attribution window to 7-day click only for a more conservative and accurate view, and cross-reference with your actual order data or server-side reporting.