Who Decides Who Sees Your Meta Ad?

Meta Ads as a five-station assembly line: how your ad goes from tens of millions of candidates to one screen, why so few buy, and what you can change.

Meta Ads, explained. Every time someone opens Instagram or Facebook, Meta fires up an assembly line that decides which ad to show them. That's the core idea. It isn't magic, and it isn't "the algorithm" in a bad mood: it's a line with five stations, and once you understand all five, you understand why one campaign sells and another just collects clicks.

We'll follow one person through the whole line. Her name is Megan, she lives in Austin, and she opens Instagram at nine at night, after dinner.

Isometric assembly line with five numbered stations: an avalanche of ads enters on the left and a single one exits on the right toward a person's phone. — The five stations your ad passes through before it reaches a screen.

Station 1: the inventory

The moment Megan opens the app, tens of millions of ads could, in theory, be shown to her^[]. Yours is one more card in that avalanche. Nobody, not even Meta, can calmly weigh tens of millions of options for each person every time they open the feed. That's why the next station exists.

Station 2: the first filter (Andromeda)

The first machine is called retrieval, and Meta's engine for it is called Andromeda. It has one job: out of those tens of millions of candidates, keep a few thousand that make sense for Megan^[]. It doesn't pick the winner; it builds the shortlist.

An ocean of millions of ad cards flows into a violet glass machine, and out the other side comes a small cube holding a few thousand cards. — Retrieval: from tens of millions of candidates to a few thousand.

If your ad doesn't get through this filter, it doesn't matter how good it is: Megan will never see it. Meta reports that Andromeda improved recall at this stage by 6% and ad quality by 8% on selected segments^[]. Those are Meta's numbers about its own platform, not what your account should expect.

Station 3: the ranking (GEM)

Now each of those thousands of ads gets a score. The question the model answers isn't "does Megan like this?" but "how likely is Megan to do what this advertiser said matters?" If you optimized for clicks, the question is about clicks. If you optimized for purchases, it's about purchases. What you choose to optimize defines the question.

A row of ad cards on a conveyor belt; a bar of a different height floats above each one, and a robotic arm lifts the card with the tallest bar toward a phone. — Ranking: every candidate gets a score and the most likely one moves up.

The core model that does this scoring is called GEM. Meta reports that it delivered an 18.9% conversion improvement and a 12% CTR improvement across its surfaces and objectives^[]. Again: that's the result Meta states for its whole platform.

Station 4: the auction

The top-scoring ads go to an auction, and one of them lands in Megan's feed. This is where your bid comes in, but it has arrived late to the party: it competes only against the ads that survived the previous three stations.

Station 5: the return signal

This is the station almost everyone forgets, and it's the one that depends most on you. Megan sees your ad, does nothing that night, and buys three days later, over WhatsApp or in your store. If that sale makes it back to Meta, the line learns it got Megan right, along with everyone like her. If it doesn't, as far as the line is concerned that sale never happened.

A sale at a counter sends a cyan beam back to the machine, which lights up; below, a gray dotted path breaks halfway and never arrives. — The sale that comes back teaches the line. The one that doesn't teaches it nothing.

And coming back isn't enough: it has to come back in a way Meta can tie to a person. According to Meta's documentation, only the events it manages to match are used for attribution and ad delivery optimization; the better the match, the better the results^[].

The problem: almost nobody reaches station 5

Here's the uncomfortable part. The line processes millions of impressions, but only a small fraction of people end up in a purchase or a lead that really closes^[]. It's a factory that makes millions of parts and gets very few quality reports back. With so few examples of "this one worked," learning who to show a purchase to is hard, especially for expensive products or ones people rarely buy.

What's new: a map of interests, in levels

In July 2026 Meta published research to attack exactly that problem. It's called Hierarchical Interest Representation: a layer that sits before ranking, designed to improve deep-funnel conversion optimization^[].

The idea, in plain English: organize interests from general to specific. Sports, then running, then marathon running shoes. Very few people bought marathon shoes last week, so the model has few examples at that level. But it has plenty at "running" and at "sports." If the system understands that those levels are connected, it can use what it learned at the top to guess better at the bottom, even if Megan has never bought anything like it.

A tree of glass nodes that runs from a general trophy to tennis and cycling branches, and from there to leaves with a medal, a sneaker and a stopwatch; a cyan light climbs the branches and spreads among the leaves. — When a level has few buyers, the model learns from the level above.

Now the honest part. This is research, not a button you can switch on in your account. The article reports no conversion or ROAS gains. Its only figure is that, in testing, training ran 30 times faster than the previous version^[]: training speed, not the performance of your ads. And Meta says the layer is designed to plug into GEM, Andromeda and its Adaptive Ranking Model "eventually"^[]. That's a plan, not a fact.

What it isn't

It isn't a punishment. If your ad isn't being delivered, it's almost never that Meta "has you blocked"; usually it didn't pass the filter or didn't win the ranking.
It isn't manual targeting. Your audiences are a hint for the line, not an order. The system decides.
It isn't something you turn on. Hierarchical interest representation is published research, not an option in Ads Manager.

What you can change at each station

You don't control the line, but you do control what goes in and what comes back.

At the inventory, enter with options that are truly different. Changing the button color doesn't give the line another option; changing the angle (price, status, convenience, urgency) does. That's our working judgment, not a Meta figure. To get those angles, Psydience tells you what motivates each type of audience and Marcus generates up to twenty variants of an ad per cycle.

At the ranking, ask the right question. Optimize for the event that counts (the purchase or the qualified lead), not the click, and send the value of each sale so the line can tell a small ticket from a big one.

At the return signal, make sure no sale gets lost. Use the Pixel together with the Conversions API (CAPI), which sends events from your server, with a shared identifier so Meta counts each sale once, and watch your match quality. If your sales close on WhatsApp, ChatIA sends that conversion back to Meta, and setting up CAPI with deduplication and purchase value is work for our data analytics and MarTech team.

And when you measure, don't ask only Meta. The ROAS you see in Ads Manager is calculated by Meta. To know how much Meta contributes compared with your other channels, Pitonisa does marketing mix modeling (MMM).

The full trip

Inventory: tens of millions of possible ads.
Filter (Andromeda): a few thousand remain.
Ranking (GEM): each one gets a score based on what you optimized for.
Auction: the finalists compete and one reaches the screen.
Return signal: the sale makes it back to Meta, or it gets lost.

Meta is already great at finding the Megan who will buy. What it can't do is learn from a sale it never heard about.

Which of the five stations do you think your account is stuck on? Tell us; the answer is almost always in the fifth. And if you want a second pair of eyes, we'll review your Meta conversion signals in a diagnostic session: events, match quality and what you're optimizing for today.