Learn
The three steps Meta runs on every ad, in plain language: how the algorithm reads, shortlists, and ranks your creative, and the moves that help you win.
The video above walks through the whole system in nine minutes. This page is the study-guide version: the three steps Meta runs on every ad, what each step reads, and the practical move that follows from each one.
Updated 31 July 2026 to reflect Meta's Q2 2026 earnings call, where CFO Susan Li introduced Meta Generative Recommender, an LLM deployed into ads retrieval. Read the transcript →
How does the Meta ads algorithm work? Meta's ad system runs three steps. A foundation model reads and summarises every ad, a retrieval engine narrows tens of millions of candidates to a few thousand for each person, then a ranking model scores that shortlist in detail before the auction picks the final winner. In July 2026 Meta named a new system, Meta Generative Recommender, that puts an LLM inside retrieval itself: rather than scoring every ad individually, it reasons about the creative content and the person together to predict the best match.
Before your ad is shown to anyone, Meta's foundation model reads it in full: the copy, the visuals, the audio, the video. It works out what the ad is and who it suits, then stores a summary. There is no targeting form for this part. The creative itself is the input.
The practical consequence: everything in the ad is a signal you are sending, whether you meant to send it or not. A cluttered visual, a generic hook, or a mismatched voiceover all shape who the system thinks the ad is for.
Meta's retrieval engine scans those summaries every time someone opens the app and cuts tens of millions of ads down to a few thousand candidates for that person. This is the step Meta upgraded in July 2026: the company deployed the first generative model, an LLM, into ads retrieval, naming the new system Meta Generative Recommender. Instead of scoring each ad individually, the model now reasons about the creative content and the person together, predicting the best match before the ad ever reaches ranking.
Here is the part most accounts miss: the retrieval engine groups ads it reads as functionally the same, so twenty near-identical variations compete as one. We cover the mechanics in Meta's hidden Entity ID problem. If your ads are all variations of one idea, you enter this stage with one ticket, not twenty. And now that retrieval runs on an LLM that reads the creative itself, what is in the ad matters more than it ever has: it is not just the input to ranking, it is the input to being retrieved at all.
The ranking model, which Meta calls GEM for ads ranking and sequence learning, reads the shortlist closely and ranks each candidate on six dimensions, a framework known as POSTER:
The survivors go to Meta's ad auction, where the winner is the combination of your bid, how likely the person is to act, and how likely they are to genuinely enjoy the ad. A strong creative match regularly beats a higher bid.
Each step favours accounts whose ads are genuinely different from each other. Different personas, different emotions, and different formats give the retrieval engine more distinct candidates to pull and the ranking model more profiles to match. With Meta Generative Recommender now using an LLM to reason about creative content during retrieval, creative diversity is not just a ranking advantage. It determines whether the ad gets retrieved at all. Small tweaks to one concept do not create that range. This is why creative diversity is the working lever, and we keep a full guide on how to measure and improve creative diversity.
On its Q2 2026 earnings call, Meta introduced Meta Generative Recommender, the first generative model deployed into the company's ads retrieval system. CFO Susan Li described it as "a paradigm shift in how our ads system works."
Here is what changed, in plain language.
Until this quarter, retrieval worked by scoring each eligible ad individually against a profile of the person about to see an impression. The system asked "does this ad match this person?" for every ad, one by one. It was fast, it was mechanical, and it treated every ad as an independent data point.
Meta Generative Recommender does something different. Instead of scoring ads individually, it uses an LLM to reason about the creative content and the person's preferences together, and predict the single best ad for that person in that moment. Susan Li's words from the call: "Rather than scoring every possible ad individually, we are now using LLMs to reason about ad content and user preferences together, and predict the best ad for each person."
This is retrieval with context. The model reads what is in the creative, the visuals, the copy, the audio, the offer, and what it knows about the person, and it makes a judgement about the match. It is not asking "does this fit." It is asking "which of these is the right one."
Meta built the raw material for this on the organic side first. Every public Reels and Feed post on Instagram now passes through an LLM that analyses it across dimensions from topic to tone, and those signals feed ads ranking and recommendations. Facebook surfaces are next. The same LLM that understands what a post is about is now being pointed at ads.
The practical consequence is the same one this page has been making since it was published, only stronger: what is in the creative is now more of the input than it has ever been. Retrieval used to run on an ad's performance history. It now runs on what the ad actually contains, read by a model that reasons about content rather than tallying keywords.
Three things follow:
The auction still runs the same way it always has. The winner is still the combination of your bid, how likely the person is to act, and how likely they are to genuinely enjoy the ad. What changed is the step before the auction, retrieval, and what it runs on. The input is the creative now, not the performance history. That was the direction of travel when this page was first published in July 2026. Meta Generative Recommender is the confirmation.
Source: Meta Q2 2026 earnings call, CFO Susan Li prepared remarks. Read the transcript →
Meta never shows you this scorecard. Ads Manager will not tell you that one ad was read as humour for busy parents and another as nostalgia for athletes. Wilow rebuilds the picture from the outside: AI Creative Tagging labels every ad in your account the way the system reads it, and the Creative Leaderboard ranks what is actually winning. Start free, no credit card required, and see the profile the algorithm has built for your account.
Want the full theory? Read the complete deep-dive on The Keyword: how the Meta algorithm decides which ads you see.
Meta's ad system runs three steps on every ad. First, a foundation model reads and summarises the creative: visuals, audio, copy, everything. Second, a retrieval engine narrows tens of millions of ads to a few thousand candidates for each person. In July 2026 Meta deployed an LLM into this step, naming it Meta Generative Recommender. The model now reasons about the creative content and the person together rather than scoring each ad individually. Third, a ranking model, GEM, scores that shortlist in detail across six dimensions (persona, offer, format, creative type, emotion, and audience readiness), then the auction picks the winner. The full walkthrough is in the video above.
Meta has moved from interest-based targeting to content-based retrieval. Instead of advertisers picking who sees their ad, the system reads the creative itself and matches it to people. In July 2026 Meta deployed an LLM called Meta Generative Recommender into retrieval, which means the system now reasons about the creative content and the person together rather than mechanically scanning summaries. The creative is the targeting input: what you show determines who sees it. The practical consequence: creative diversity matters more than audience settings. Accounts that run genuinely different ads (different formats, emotions, personas) give the retrieval engine more distinct candidates to pull. Accounts that run variations of one idea enter the system with one ticket, not twenty.
Creative fatigue is the most common cause. Meta's algorithm shows your ad to people it predicts will respond, but after 2-4 showings to the same person, engagement drops sharply. The fix is not more budget. It is more creative diversity. The brands that scale on Meta run ads that are genuinely different from each other (different hooks, formats, emotions), not small variations of one concept. When each ad reaches a different audience, the system has more room to deliver. The Creative Diversity guide linked above covers how to measure and improve this.
Not universally, but video tends to earn more delivery because it generates more engagement signals: longer view times, more saves, more shares. However, a strong image ad that generates high engagement will beat a weak video ad. The format is not the deciding factor. The deciding factor is whether the ad matches what a real person wants to see in their feed right now. Video is a better bet for most brands because it gives you more seconds to earn that match, but a compelling static image with a sharp hook can outperform a mediocre video every time.
Creative diversity is the spread of genuinely different ads in your account: different hooks, formats, emotions, creative types (UGC, testimonial, demo, founder story), and personas. Meta's retrieval engine groups ads it reads as similar, so running twenty variations of the same concept counts as one ad in retrieval. With Meta Generative Recommender, the LLM Meta deployed into retrieval in July 2026, the system now reasons about the creative itself, which means what is in each ad determines whether it gets retrieved at all, not just how it ranks. The brands that scale on Meta run ads that are meaningfully different from each other, not just different in small details. Wilow measures creative diversity automatically across every ad in your account. The Creative Leaderboard shows you the spread, and the Outliers chart flags the few ads doing all the work. Start free, no credit card required.
