Meta Ads algorithm: what's left for you to decide

Over the past three years Meta has rewritten the architecture that decides which ad lands in front of you four times over. The practical consequence for anyone running campaigns is that part of the craft has moved. There are fewer levers inside the panel, and they weigh less. What you feed the system weighs more.

That's neither good news nor bad. It's a change in where the hard work sits.

What changed under the hood, briefly

Andromeda is the retrieval engine. It sits ahead of the auction and cuts tens of millions of eligible ads down to a few thousand candidates, which only then get scored properly. GEM is the generative model Meta describes as the central brain of the system, distilling what it learns into the other models. On top of that runs a ranking layer that brings LLM-scale models inside the auction's response time.

That's the two-minute version, and I'm stopping there on purpose. The mechanism is interesting, but it isn't the useful part for your week.

The useful question is this one. If the system decides more and more on its own, what's left for you to decide?

Targeting isn't dead, it changed weight

Every time Meta automates something, someone announces the death of targeting, and every time it's an oversimplification. Geography, age, custom audiences, exclusions and business constraints all still do work. A company that only delivers in Rome doesn't leave the whole country open hoping the system works it out.

What changes is the return on manual work. For years the edge lay in hand-building the audience combination that performed best. Once the system gets good at finding it by itself, that edge thins out, and the difference between two accounts stops living there.

Ask for the wrong thing and you'll get it

This is the part I see underrated most.

When you give the system more freedom, being precise about what you're asking it to find becomes decisive. A campaign that brings a hundred leads at ten euros each looks like a good result, right up until you find out eighty of them never answer the phone. In that case Meta did exactly the job you asked for. You told it to find people who fill in a form, and it found them.

The problem isn't the algorithm. It's that you handed it a definition of success that was easier to measure than to defend.

The cure isn't a setting buried in the panel. It's feeding back the information that arrives later, once you know who was genuinely interested and who wasn't. On what makes a lead worthless before it even arrives, I've written a separate note.

A weak offer stays a weak offer

There's an uncomfortable consequence in all this, and it has to do with excuses.

While targeting was a manual craft, "we didn't find the right audience" was a usable explanation almost every time. As the system gets better at finding the people most likely to respond, that sentence covers fewer and fewer cases. What's left uncovered is the real stuff, and it's harder to say to a client.

That the price doesn't survive comparison. That the landing page loses half the people who reach it. That the message speaks to a problem nobody feels. That the product doesn't have enough demand in that market.

A recommendation system, however large, can estimate probabilities. It can't manufacture a reason to buy where there isn't one. I've argued the same thing about AI in marketing generally.

Why knowing people matters more, not less

It sounds like a contradiction, and it's the direct consequence.

The more the machine handles distribution and matching, the more the work left to you is the work the machine can't do in your place. Understanding why someone would buy. What holds them back. Which alternative they're weighing while they scroll. How much they already know about the problem, and which promise they find credible coming from you rather than a competitor.

Meta can cross billions of signals to find correlations. But someone has to build an offer and a message that mean something to a real person, and that someone isn't inside the panel.

The two questions that stay yours

On the operational levers the list is short and well known. The material you upload and how distinguishable it is. The event you ask the system to optimise for. The value you declare for each conversion. The information you send back about what happened after the click. On the first of those, creative variance and what actually counts as a different ad are worth their own read.

What's left are two questions no settings checklist covers, and they're where I start.

Is there a real reason to buy from this client rather than someone else? If the answer is price and the price doesn't survive comparison, no signal will fix the campaign. That question gets asked before you open the panel, not after three disappointing weeks.

Do I actually know who I'm talking to? What they fear, what they're already considering, how much they know about the problem, which promise they find credible coming from you. The system can find you the audience. You still have to give it something worth showing them.

Frequently asked questions

Is Andromeda Meta's new algorithm? No. It's the retrieval engine, the piece that selects candidates before the ranking systems get involved. Calling it "the algorithm" lumps together things that sit at different points in the chain.

Does interest targeting still work? Yes, but it carries less weight in the final result than it used to. It still does the job of setting the perimeter, meaning where you can sell and who you don't want to talk to. It does less as a technique for finding the right people inside that perimeter.

If the system decides more, why not just let it do everything? That's the wrong conclusion. The more operational decisions it takes, the more the few left to you matter, meaning objective, offer, message and definition of result. The fewer levers you have, the less you can afford to get them wrong.

How many creatives do I need? There's no number that works for everyone. What counts isn't the file count, it's how many different reasons to care you're putting into circulation.

What changes for someone who does this for a living? The value shifts from execution inside the panel towards strategy, offer, measurement quality and knowing the customer. The part being automated is the part you used to learn from tutorials.

Read also

More notes on automation, signals and creative:

If you want to work out whether your campaigns are giving the system the right signals, book a free consultation. We'll look at the optimisation event, the data you send back and your creative variance, and find where the value is leaking.

If Meta's algorithm decides almost everything, what's left for you to do