Meta's Generative Recommender: what Gulf advertisers should know

If you advertise on Facebook or Instagram, the way Meta picks which of your ads a person sees has changed again. On its 29 July 2026 earnings call, Meta's chief financial officer Susan Li described what she called a paradigm shift in its ad system. Instead of scoring every candidate ad for a person individually, Meta said it now uses large language models to consider what an ad contains and what a person prefers at the same time, then predict the ad that suits them best. Meta calls this the Meta Generative Recommender.

For a Gulf brand spending on Meta, it matters. It is the biggest change to how ads are matched with people since Meta moved away from detailed targeting. It is also easy to misread, because much of the commentary claims more than Meta has.

What Meta has confirmed

• Meta has placed its first generative model in the retrieval stage of its ad system. Retrieval is the step that cuts tens of millions of possible ads down to a few thousand, which are then ranked.
• On Facebook, Meta reported ad clicks up 8.3 percent and conversions up 15.7 percent. It attributed that to several things working together with its existing ranking models: the Generative Recommender, early pilots of large language models, and better models of what people are interested in.
• Mark Zuckerberg said that drawing on wider context from people's organic and ad activity had produced significant gains in relevance and conversions across Facebook and Instagram.

Meta has not published an engineering post or research paper on the Generative Recommender, so its architecture and how broadly it runs are not public. Much trade coverage has credited the entire 15.7 percent to this one system and described it as a ranking system. Meta's own account is more careful on both points, and any agency briefing you on it should be too.

The systems that came before it

This is not a sudden break. Meta has publicly described a run of systems, each of which moved a little more of the decision-making away from the advertiser's settings and towards models that learn from behaviour and content.

• Lattice (2023). One model learning from many objectives and surfaces together, in place of many smaller ones. Meta said its early Instagram deployment improved ads quality by about 8 percent.
• Sequence learning (late 2024). Models that learn from the order in which a person does things, rather than a flat list of features about them. Meta reported 2 to 4 percent more conversions on some segments.
• Andromeda (December 2024). A new retrieval stage, built partly to cope with the volume of ads being made with AI tools. Meta said that in a single month more than a million advertisers had created over 15 million ads with its tools.
• GEM (described November 2025). Meta's generative ads model: a large model that teaches the smaller models that actually serve ads. Meta said it raised conversions by 5 percent on Instagram and 3 percent in the Facebook feed.
• An adaptive ranking model of about a trillion parameters (Instagram, end of 2025). Meta said it lifted conversions by 3 percent and click-through by 5 percent for the people it served.

The spending behind this is considerable. Meta expects capital expenditure of 130 to 145 billion dollars in 2026, much of it on the computing these models depend on. In the second quarter of 2026, its ad impressions grew 14 percent and the average price per ad rose 12 percent.

One inconsistency is useful context. 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 being presented as central. The fairest reading is that things moved quickly and that Meta is still early in the rollout.

Where Gulf brands should put their effort

Meta has issued no guidance specific to the Generative Recommender, but its broader advice points one way: when a model reads what an ad says and shows, the creative does much of the work of finding the audience. For a business running Meta ads in the UAE or across the GCC, that suggests four priorities.

• Genuinely different ideas. Meta says ads that look alike are grouped and treated as variations of a single creative. Distinct messages, angles and formats reach distinct people.
• Ads that are clearly about something. One product, one benefit and one kind of buyer give the system something to match against. A broad brand message gives it very little.
• Stronger signals. Conversions passed through the Conversions API, and not only the pixel, give the models more to learn from.
• Room for automation. Broad audiences, Advantage+ settings and consolidated campaigns give the models space to work. Frequent manual edits interrupt them.

If an agency runs your account, ask: how many genuinely different concepts are live, is the Conversions API in place, and how often are campaigns edited by hand?

Privacy may also come up. Since December 2025, Meta has used people's conversations with Meta AI to personalise content and ads in most regions, with sensitive topics excluded. It is another sign that ad delivery now runs more on Meta's own understanding of people than on the targeting an advertiser picks.

What stays the same

The basics hold. A campaign still needs an offer people want, creative that makes that offer clear in the first few seconds, a landing page that works, and measurement that everyone involved trusts.

What changes is where effort pays off: less in audience settings, more in creative that a very capable system can put in front of the right people. Before a new concept goes live, mock it up in each placement Meta might choose and check it reads clearly wherever it lands.

If you would like a clear view of whether your Meta creative and tracking are ready for this way of matching ads, ALXO would be glad to talk it through with you.