Between October 13 and 28, 2026, Meta's engineers published three technical posts about how ads are selected and ranked. They cover a central model called GEM, a new way of reading each ad called GenRecs, and a new version of Andromeda, the system that picks which ads compete for each person. Summaries circulating on social media promise immediate gains in your account, and some parts are not confirmed. As a Meta Ads agency, we read the sources, and here we separate what Meta said from what is being assumed.
What did Meta publish and when?
Meta published three entries on its engineering blog during October 2026. The first, dated October 13, introduces GEM, a foundation model for ads. The second, dated October 19, explains GenRecs, a way of turning ads and products into codes the system can compare. Finally, the October 28 post describes the new version of Andromeda, which picks candidates before another stage ranks them.
These dates come from PPC Land, Relevant Audience, and Segwise, which reviewed the posts. We could not open Meta's pages from our work environment, so we cite these outlets and not the original text. Every figure below was given by Meta, and no third party has audited any of them.
What does it mean that Andromeda reviews all ads?
It means Andromeda now scores every eligible ad, not only a shortlist. According to Meta, the new version evaluates "every eligible ad" across Facebook and Instagram with quality close to that of the ranking stage. In its 2024 version, the system narrowed tens of millions of candidates to a few thousand per request.
One condition changes the reading. According to Relevant Audience, targeting still applies before scoring. An ad that does not match your audience, your location, or your targeting rules never enters the pool. So "all your ads" means all those that pass that filter for a specific person, not every ad in your account.
It is also not confirmed that Meta treats each variant or each product as a separate piece. That is an interpretation some outlets make, not a Meta statement. What Meta does say is that AI creative tools added hundreds of millions of ads to the pool. Meta does not give the current size.
Does Meta's central model learn from all campaigns?
It learns from data across all of Meta's ad platforms, but Meta does not say it transfers what it learns from one campaign to another. According to Relevant Audience and PPC Land, GEM replaces separate systems for Facebook, Instagram, and Reels with one central model. That model predicts click rate and conversion rate together, and it feeds hundreds of smaller models that serve ads in real time.
Meta's text does state something useful for low volume sellers. Patterns learned from click data, which is plentiful, help predict conversions, which are scarce. That is the part that resembles the idea of "starting with more knowledge."
However, PPC Land points out that the text describes learning across surfaces and objectives, not across advertisers. It also does not mention small advertisers. That an account with few sales benefits more is a reasonable hypothesis from that outlet, and Meta did not confirm it.
How reliable is the 18.9% conversion figure?
It is a Meta figure with no baseline, so it does not predict your results. Meta said GEM delivered an 18.9% improvement in conversions and a 12% improvement in click through rate, across its ad surfaces and objectives. It did not publish what it is compared against, over what period, or whether it came from A/B tests or offline evaluations.
Other figures do not fully line up. Meta cited gains for targeted users of 5.8% on Facebook, 4.1% on Instagram, and 1.6% off platform, and it does not define that term. PPC Land recalls that in July Meta credited GEM and other models with 15.7% on Facebook, a number that is not reconciled with the new ones either.
For Andromeda, Meta gave 2.4% more Facebook conversions since 2025, also without a method. The practical conclusion is the one the sources recommend. Measure with your own sales, and do not attribute a one week change in your account to a system you cannot see.
Do new ads start faster?
They may start with better information, but the learning phase still exists. According to Segwise, GenRecs reads the text and images of an ad, a product, and a page, and turns them into short codes. Similar items share codes, so a new product can inherit what the system already knows about its category.
Meta says this helps newly created ads, niche products, and pages with little engagement. That does not mean a product uploaded today sells as if it had months of history. Segwise recalls that, according to Meta's help center, an ad set exits the learning phase after about 50 results in the week following its last significant edit. Meanwhile, cost per result is usually higher.
There are no clear numbers either. PPC Land found no performance results in the GenRecs post, and Segwise cites a small 0.35% conversion gain together with other models. That a new ad starts from a better point is plausible. How much better, nobody has measured publicly.
Is it already working in my account, and can I configure it?
It probably already has an influence, but you cannot see it or control it. Relevant Audience reports that Meta announced no new setting, option, or report in Ads Manager, and that it describes the benefit as "better performance automatically." Meta also did not say when the new Andromeda version reached production.
This has a concrete consequence. There is no switch for GEM or Andromeda. Your work concentrates on what you do control: the objective you choose, the data you send, the creative you upload, and the budget you assign. The system decides everything else.
What do you do with this as an advertiser?
These are recommendations from our Meta Ads agency, not Meta instructions. First, choose an objective that resembles your sale. If you optimize for clicks, the system looks for people who click. If you optimize for purchases or qualified leads, it looks for people who resemble those who buy.
Second, send clean and complete data. Check that the pixel and the Conversions API record purchases with their value and without duplicates. This matters more now because Meta turns on a new on-device measurement on November 12, as we explained in our pixel blog. Third, let the system work. Each major edit can restart learning, and a campaign with few results does not need daily changes.
Fourth, upload several ads with different ideas instead of small variations of the same one. If the system compares ads by what they show and say, two nearly identical versions add little information. This is our inference from how Meta describes GenRecs, not a confirmed rule. Fifth, compare the results Meta reports with your real orders before changing budgets.
Consider a typical case, not real data but an example of how it might look. A sportswear store in Texas launches a new product with three ads that tell different stories: comfort, price, and style. It chooses purchases as the objective, checks that each order's value reaches Meta correctly, and touches nothing during the first seven days. At the end of the week it compares the sales Meta reports with those in its store and decides with that difference which ad gets more budget.
When does it make sense to hire a Meta Ads agency?
When you invest enough that a data error or an unnecessary edit costs real money, and you do not have time to review it every week. A Meta Ads agency can audit the pixel and the Conversions API, define the right optimization objective, and plan the creative so the ads bring different ideas. It can also compare what Meta reports with your real sales. If you want us to review how your account is set up, you can contact the JP Director team.
This topic connects with others we have published. To understand the November 12 measurement change, read our blog on the Meta pixel and AI terms. And to see other AI features Meta turns on by default, review our analysis of product browsing and AI terms.
Frequently Asked Questions
What are GEM, GenRecs, and Andromeda?
They are three Meta systems involved in choosing ads. Andromeda picks candidates, GEM is a central model that predicts clicks and conversions, and GenRecs turns ads and products into comparable codes.
Does Meta learn from other advertisers' campaigns for mine?
Meta has not said so. Its post describes learning from data across all of its ad platforms, across surfaces and objectives, but it does not confirm that information passes from one campaign to another.
Does my new ad no longer need a learning phase?
It still needs one. Meta says GenRecs helps new ads, but its help center still indicates a learning phase of about 50 results in the week after the last significant edit.
Can I expect 18.9% more conversions?
No. It is a Meta figure with no baseline, period, or method, and it is not a per account result. Measure the effect with your own sales.
Last updated: November 6, 2026. This article will be updated if Meta publishes baselines, clarifies learning across advertisers, or adds controls in Ads Manager.
