On its earnings call on 29 July 2026, Meta described what its chief financial officer, Susan Li, called a paradigm shift in how its ad system works. Rather than scoring every possible ad for a person one by one, Meta said it was now using large language models to reason about the content of ads and a person's preferences together, to predict which ad suits them best. It calls this the Meta Generative Recommender.
It is the most significant change to how Meta matches ads to people since the move away from detailed targeting, and it is worth being precise about, because much of what has been written about it goes further than Meta did.
What Meta actually said
- Meta said it had put its first generative model into the retrieval stage of its ad system: the stage that narrows tens of millions of possible ads down to the few thousand that are then ranked.
- It reported that ad clicks on Facebook rose 8.3 percent and conversions 15.7 percent. It credited those gains to a combination: the Generative Recommender, early pilots of large language models, and improved models of what people are interested in, working alongside its existing ranking models.
- Mark Zuckerberg said that using wider context from people's organic and ad activity had brought significant gains in relevance and conversions on Facebook and Instagram.
Meta has not published an engineering post or paper about the Generative Recommender itself, so its architecture and how widely it runs are not public. Much of the trade coverage has attributed the whole 15.7 percent to it alone, and described it as a ranking system. Meta's own account is more careful than that.
What it builds on
The Generative Recommender is the latest in a line of systems Meta has described publicly, each taking a little more of the job away from the advertiser's settings and giving it to models that learn from behaviour and content.
- Lattice, from 2023: one model learning from many objectives and surfaces at once, rather than many small ones. Meta said it improved ads quality by about 8 percent in its early Instagram deployment.
- Sequence learning, from late 2024: learning from the order of what a person does, not just a list of features about them. Meta reported 2 to 4 percent more conversions on some segments.
- Andromeda, from December 2024: a new retrieval stage, built partly to cope with the volume of ads that AI tools were producing. Meta said more than a million advertisers had made over 15 million ads with its tools in a single month.
- GEM, Meta's generative ads model, described in November 2025: a large model that teaches the smaller ones that serve ads, which Meta said lifted conversions by 5 percent on Instagram and 3 percent in the Facebook feed.
- An adaptive ranking model of about a trillion parameters, launched on Instagram at the end of 2025, which Meta said lifted conversions by 3 percent and click-through by 5 percent for the people it served.
Meta is spending heavily to keep going. It expects capital expenditure of 130 to 145 billion dollars in 2026, much of it on the computing that these models need. In the second quarter of 2026 its ad impressions rose 14 percent and the average price per ad 12 percent.
One contradiction is worth knowing about. In March 2026 Li told Digiday that large language models were not yet a big part of Meta's core ranking. Four months later they were presented as central. The honest reading is that this moved quickly, and that Meta is still early in rolling it out.
What it changes for an agency
Meta has not published guidance specific to the Generative Recommender, but its wider guidance points one way. If a model reads what an ad says and shows, the creative itself becomes a large part of how the ad finds its audience.
- Creative diversity. Meta says ads that look alike are grouped and treated as variations of one creative. Genuinely different ideas, messages and formats reach different people.
- Clear creative. An ad that is obviously about one thing, for one kind of buyer, gives the system something to match. A vague brand message gives it very little.
- Signals. Conversions sent through the Conversions API, not just the pixel, give the models more to learn from.
- Patience with automation. Broad audiences, Advantage+ settings and consolidated campaigns give the models room to work. Frequent manual changes interrupt them.
There is also a privacy point clients may ask about. Since December 2025, Meta has used people's conversations with Meta AI to personalise content and ads in most regions, excluding sensitive topics. It is one more sign that the inputs to ad delivery are increasingly Meta's own understanding of people, not the advertiser's targeting.
What it does not change
The fundamentals of a good campaign are the same: an offer people want, creative that makes it clear in the first seconds, a landing page that works, and measurement that the client trusts. What has changed is where the skill goes. Less of it goes into audience settings, and more into making creative that a very capable system can match to the right people.
The ad mockup generator on this site shows one ad across Meta's placements with each app's text cuts, which is a quick way to check that a new creative idea reads clearly wherever the system chooses to show it.
