More insights.
The learning never stops.
Strategy, creative and numbers feed each other every week. What the loop teaches us lands in your inbox.
If you came here from the film, this is everything the five minutes didn't have room for. If you landed here cold, it works alone: the written version of how Meta actually decides what gets seen.
Everyone talks about the algorithm like it's a mysterious black box. It isn't. Meta has told us how it works, in engineering blog posts and Mosseri's Instagram Lives. Nobody puts it together in a way that helps you run a brand. That's this page's only job.
the one-line model
Every time someone opens Instagram, Meta asks one question: what will keep this person here longest? It can't read minds, so it predicts. For every post and every ad, it calculates how likely this specific person is to engage with this specific content, ranks everything by that probability, and shows the winners.
Organic and paid run on the same logic; only the predicted action changes. For a Reel, Meta predicts watch time, likes per reach, and sends per reach, and sends is the strongest signal of the three, because forwarding something to a friend is the highest endorsement a feed can produce. For an ad, the prediction is called Estimated Action Rate: the probability this person takes the action the advertiser paid for. Different names, one machine. The algorithm isn't mysterious. It's a prediction engine.
the machine
Here's the engineering problem Meta solves on every scroll. The decision of what to show runs in roughly 300 milliseconds, faster than a blink, but billions of ads are live in the system. No auction can rank billions of candidates in a blink. So the decision is split in two.
Stage one is retrieval. A system called Andromeda, shipped quietly in late 2024, scans tens of millions of eligible ads and narrows them to roughly a thousand candidates per impression. It's a neural network, and what it reads to make that cut is the creative itself: the visuals, the text, the offer structure, the emotional tone. It asks, in effect, what kind of person would respond to this, and whether the person scrolling is that kind of person.
Stage two is the auction. From those thousand, one winner, by the math in the next section.
The part most advertisers never internalise: if your ad doesn't pass stage one, it never enters the auction. Your bid doesn't matter. Your targeting doesn't matter. You're not even in the room. Andromeda decides who gets into the room, and it decides by reading the creative. Which is why your targeting settings have quietly become suggestions, and your creative has become the actual targeting.
the auction formula
Three parts. Your bid: what you're willing to pay. Estimated Action Rate: Meta's live prediction that this specific person takes this specific action on this specific ad, right now, recalculated per impression. Ad Quality: how good the experience is, measured before the click (do people watch, like, hide it) and after it (does the landing page deliver what the ad promised).
Highest total value wins. Not highest bid. Run the arithmetic and the strategy falls out of it: because the bid is multiplied by the prediction, a modest bid attached to a strong prediction beats a heavy bid attached to a weak one. You can pay less and win more if your creative is better. The reverse is equally true, and it's why throwing budget at a tired ad does nothing: if the predicted action rate is near zero, no multiplier rescues the product. The system is built, deliberately, so that money cannot buy its way past relevance.
the three mechanisms underneath
The signals. The predictions run on everything Meta observes: every like, every Reel watched to the end or skipped at one second, every ad clicked, every purchase on any site with a pixel installed, scroll speed, hover time, the hour you browse versus the hour you buy. Meta knows you browse shoes at 11pm but only buy at lunch. Think of it as a psychic bartender: it knows your order before you've said a word. Thousands of signals per person, updated in real time, which is why the same ad carries a different Estimated Action Rate for every single impression.
The learning loop. New content gets tested on strangers first. Reels are shown to a slice of non-followers to see if they hold up without the courtesy of familiarity; new ads enter a learning phase of roughly fifty conversions. During it, Meta watches who clicks, who converts, who bounces, and builds what's effectively a conversion profile: a pattern of the kind of person who responds to this creative. Then it scales by finding more of that pattern. Read that again, because it's the most underrated sentence in performance marketing: your early responders define who you reach at scale. If the first hundred buyers are discount hunters, Meta concludes your brand is for discount hunters and obligingly finds you thousands more. Your creative summons the early responders, the early responders set the profile, the profile becomes your audience. It compounds, in whichever direction it starts.
Entity IDs. When you upload an ad, Andromeda doesn't file it as a lone file. It reads the visuals, the copy, the offer type and the tone, and clusters it with ads that say essentially the same thing, under one Entity ID. And each Entity ID gets one ticket to the auction. One. Upload fifty variations of your winner and they don't compete for fifty slots; they share a single ticket, cannibalising each other inside one cluster. This is the mechanism that quietly killed the old playbook of flooding the account with variations and letting the algorithm pick. What earns tickets now is genuine diversity: a different format, a different angle, a different persona, a different leading benefit. Each genuinely distinct idea gets its own cluster, its own ticket, its own shot at finding its audience. The question stopped being how many ads you have. It's how many genuinely different ideas you have.
three tests to run this week
1. The cost cap truth test. Pick your supposed winner. Duplicate it into a fresh campaign with a cost cap at your target CAC, broad audience, no interest stacks. Let it run 72 hours. If it spends and converts at that cost, it's genuinely strong: the algorithm found buyers at your price. If it barely spends, the mechanics are telling you something precise: Meta only enters auctions it predicts it can win at your cap, and it found none worth entering. The ad was never winning. It was buying conversions at whatever price lowest-cost bidding could find them. Cost cap is truth serum.
2. The diversity audit. List your live ads. Group them by idea, not by file: same angle, same format, same persona is one idea, whatever the file names say. If thirty ads collapse into three groups, you hold three auction tickets, not thirty, and most of your account is competing with itself. Next week's brief goes to a group that doesn't exist yet.
3. The early-signal check. Open your newest ad's first fifty conversions and look at who they are: order values, products bought, new or returning. That cohort is who Meta now believes your brand is for, and it's about to find you more of them. Judge your creative by who it summons, not just what it costs.
a note for this exact week
Festive auctions sharpen everything above. Every brand in the country is bidding at once, so attention costs 40 to 70 percent more than a normal month. The machine rewards fresh creative and taxes stale creative hardest exactly when impressions are priciest, and a campaign restructured now pays peak prices for its learning phase: the most expensive fifty conversions of the year. So ride proven winners, don't rebuild. And circle November 10 to 30: CPMs fall off a cliff after Diwali while wedding and self-purchase intent carries on, and most advertisers go quiet because the budget's spent. The cheapest good window of the half belongs to whoever is still talking.
the glossary
Andromeda: Meta's retrieval system since late 2024. A neural network that reads creative and decides which ads enter the auction at all.
EAR (Estimated Action Rate): Meta's per-person, per-ad, per-moment prediction of action. The biggest lever in the formula, and it's set by your creative.
Entity ID: Andromeda's cluster for ads that say the same thing. One cluster, one auction ticket.
Learning phase: Roughly the first fifty conversions, where Meta builds its profile of who your ad is for.
Conversion profile: The pattern of person Meta learns from your early responders, then scales.
CPMr: CPM times frequency: what you actually pay to reach a unique person. The leading indicator; rising CPMr predicts CPA trouble weeks early.
NCCPA: Total paid spend divided by new customers. Catches a brand quietly recirculating old buyers while ROAS still looks fine.
what to ask your agency on monday
Show me the FBU breakdown for our product. If there isn't one, the briefs are guesses.
Why did our last winner work? If they can't tell you why something worked, they can't make it work again.
Which persona and awareness level is next week's upload briefed against? If the answer is "we'll test some creatives," there's no map.
What number do you manage us to besides Meta's ROAS? If the dashboard number is the only number, read our article on why ROAS lies.
take it
None of this is a trade secret. The advantage is that we run it every week, for every account, without exception. The rest of the system, the research weeks, the brief structure, the compounding method, is in the other articles on this site. If you want it run on your brand, book a call.