The Maze: Meta is explaining how its Facebook and Instagram ad systems learn what products mean, rather than relying only on each ad's past results. In an October 28 interview, the company described connected upgrades that help match new ads with people. Advertisers need no new setting to benefit, Meta says.
Source: Meta. Text, images and learned item representations become shared tokens for ad retrieval and ranking.
The Details:
New ads can borrow experience. A fresh campaign normally has little performance history. Meta's generative recommendation system, GenRecs, uses language models to read an ad's text and images. It represents attributes such as brand, colour, fit and style through shared compact descriptions called semantic tokens. That lets the system connect a new product with similar products it already understands. A new running-shoe campaign need not arrive as a complete stranger. Meta's ads infrastructure executive Matt Steiner says this can shorten the learning period. It does not eliminate the need to test how people actually respond.
The shortlist and the final choice are different jobs. When someone opens Facebook or Instagram, Meta must narrow a huge pool of eligible ads before choosing what to show. A new generative retrieval model adds recommendations alongside the existing route. Retrieval means making that shortlist; ranking means comparing its candidates in greater detail. Meta has also rebuilt its Andromeda retrieval system so more processing happens together on graphics chips, reducing time spent moving data. Its separate Generative Ads Model, or GEM, trains smaller ranking models. The updated version reads twice as much engagement history. GEM supplies learning; it does not pick every ad directly.
The performance claims come with dates. Meta's technical explanation reports a 0.35% lift in conversions, meaning completed actions such as purchases, from tokenization paired with GEM and another ranking model. The footnote dates that internal measurement to February 2026. This is not a fresh October test of GenRecs alone. Nor is a conversion necessarily a profitable order. The broader click and conversion gains discussed in the interview concern other periods and systems. They cannot be added together to estimate a merchant's expected return.
MarketMaze Take: Meta is moving more audience-selection work inside its own models. For businesses buying these ads, the useful change is that product meaning can travel between campaigns. The limit is equally commercial: a system that predicts who might respond still cannot decide whether the product, offer and resulting sale make economic sense.
The input matters as matching becomes automated. Meta's own example is a new season's trail shoe benefiting from learning about an earlier season. For advertisers in different markets using Facebook and Instagram, that explains why a new campaign need not start with zero useful information. It supports a sharper role for product and creative strategy, although it does not prove any particular creative treatment will improve sales.
Automation leaves the profit question open. Steiner says advertisers need not change settings and that models cannot set strategy. That combination matters: the platform can improve matching without handing businesses a new control. Merchants still need to judge whether the customers and orders it delivers justify their advertising spend. Meta's internal conversion figures do not answer that question for them.
Sources: AdExchanger, Meta, Meta, Meta
