Updated 1 August 2026. Source: Meta Q2 2026 earnings call transcript.
Why the Meta algorithm works this way
Meta sells attention, and attention only lasts while people enjoy what they see. The platform has to keep users interested or they leave, and it has to deliver results for advertisers or they stop spending. Advertising brings in roughly $200 billion a year, so the ad system is tuned to balance those two pressures at once: show ads people respond to, and reward the advertisers who earn that response.
The three steps, from creative to auction
Picture Meta's ad inventory as a library holding tens of millions of books. Three steps decide which book ends up in front of you: foundation models read every ad and build a short summary, a generative AI system retrieves the most relevant ads for each person, and a ranking layer picks the winner before an auction sets the price.
Step one: foundation models read every ad
Before anyone opens the app, Meta's AI models read every ad an advertiser uploads: the text, the images, the audio, the video. Each model works out what the ad is and who it suits, then stores a short summary. Meta runs this process for every advertiser, building a library of tens of millions of summarised ads.
One of the models that does this work is GEM, Meta's ads ranking model. GEM reads each ad across multiple dimensions, building a profile that later stages of the delivery system use to decide who sees what. Meta describes GEM's role as "ads ranking and sequence learning." It ranks; it does not retrieve.
Step two: Meta Generative Recommender finds the right ad
When a person opens the app, Meta does not mechanically score every possible ad against a user profile. It now uses Meta Generative Recommender, an LLM-based system that reasons about the creative content and the person together, then predicts the single best match.
CFO Susan Li described the change on Meta's Q2 2026 earnings call as "a paradigm shift in how our ads system works." Her prepared remarks explained the mechanic directly: "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, not mechanical scoring. The system looks at what the ad actually says and shows, not just a summary of who clicked on something similar before. For advertisers, that means the creative itself carries more weight than it did under the old retrieval model.
Every public Reels and Feed post on Instagram now passes through an LLM that analyses it across dimensions from topic to tone. Those signals feed both ads ranking and content recommendations. Facebook surfaces are next.
Step three: ranking and the auction
With a shortlist of the most relevant candidates, Meta's ranking models look closely at each ad. GEM, the same model that summarised the library, now reads each candidate in full. Its ranking layer weighs not only what the ad contains but how it feels. Those signals map to a simple framework, POSTER, six dimensions Meta reads in every ad:
In practice your creative becomes your targeting, because Meta's ranking models read all six of these signals to decide who should see the ad. Wilow scores your own ads across these same six dimensions, so you can see the profile Meta builds in private.
Where this is heading: predicting the response
Meta is taking the same idea further with TRIBE v2, a model it describes as a digital twin of the human brain. The aim is to predict how someone will respond to what they see and hear, then match an ad to the people most receptive to its emotional pull. A model like that only works with creative that carries emotion, and with different ads that carry different emotions.
Creative diversity: who you talk to, and how
Meta's own guidance is direct about this: diverse creative reaches audiences a single ad never will. Diversity works on three axes.
Iteration is not diversification
There are two ways to produce new ads, and Meta draws a hard line between them. Creative iteration means small tweaks to the same ad, like a new button colour or a different call to action on the same design. Creative diversification means genuinely different ads for different people and different angles. Meta is explicit that iterating on a single concept is not enough to unlock new audiences, and that tweaks should not stand in for true diversification.
The auction decides the final winner
A strong match still has to win an auction. Meta sells each slot to the highest effective bid, so the final decision weighs three things at the same time: how likely the person is to enjoy your ad, how likely they are to take the action you want, and your bid measured against everyone else. Meta's ad auction documentation frames the winner as that combination. The bid matters, but it does not win on its own. A weaker match priced high can still lose to a stronger match, and a strong match can lose if the bid is capped too low.
The Meta algorithm keeps its scorecard hidden
None of this surfaces in Ads Manager. There is no screen that tells you one ad was read as humour for busy parents and another as nostalgia for athletes. Meta builds that picture privately, which leaves advertisers guessing at the exact thing the algorithm rewards.
See your Creative Diversity Score
Wilow reads every ad in an account and tags it the way Meta's system would, by audience, angle, and emotion, then scores how diverse the creative actually is. It turns the hidden scorecard into something an advertiser can act on. Check your Creative Diversity Score, free.
Strong creative is only half the equation. A diverse, well-read set of ads still underperforms if the campaign is built wrong inside Ads Manager, which is the subject of the next guide.
