# How Meta Ranks a Brand-New Ad in 2026: GEM and GenRecs Explained Author: Angad Singh Author URL: https://segwise.ai/blog/author/angad-singh Published: 2026-10-29 Category: Strategy & Optimization Category URL: https://segwise.ai/blog/category/strategy-and-optimization Meta Title: Meta GEM and GenRecs Explained: How New Ads Get Ranked Meta Description: How Meta's GEM and GenRecs let a new ad learn from similar ads, what changed in 2026, and what to do about creative variety, structure, signals and catalogs. Tags: Creative Strategy, meta ad campaigns, Updates Tag URLs: Creative Strategy (https://segwise.ai/blog/tag/creative-strategy), meta ad campaigns (https://segwise.ai/blog/tag/meta-ad-campaigns), Updates (https://segwise.ai/blog/tag/updates) URL: https://segwise.ai/blog/meta-gem-genrecs-new-ads Meta picks your ads with three systems. Andromeda, the retrieval step, now rates all eligible ads and builds a shortlist for each person. GEM, a giant foundation model, teaches the smaller models that rank that shortlist. GenRecs reads each ad's text and images and turns them into shared "meaning" tokens that both steps use, so a new ad borrows what Meta already learned from similar ads. Your job: give it distinct concepts, clean value signals and accurate product data. Segwise is an AI creative analytics and creative generation platform that tags every element of every ad you run, shows which elements drive results across 15+ ad networks and MMPs, and helps your team turn those winning patterns into new statics, videos and AI UGC-style videos. ## TL;DR by situation - **If you run DTC or ecommerce ads:** variety now has to be variety a model can read: a different person, setting, product, message or format. Ten colour swaps of one ad likely look like one item to GenRecs. Send purchase values through Conversions API. - **If you run catalog ads with thousands of SKUs:** your titles, images and variant fields are likely the "description" GenRecs reads to place a new product next to similar ones. Fill size, colour, brand and extra images properly. - **If you run app installs or mobile games:** GEM is trained on mobile app install data and Meta lets app campaigns optimise for purchase value. Meta has not published app-specific GenRecs results yet. - **If you manage account structure:** Meta's own Help Centre still says fewer ad sets and fewer ads, each with diverse assets, beat lots of small ones. A new ad learns faster now, but the learning phase still exists. ## What are Andromeda, GEM and GenRecs, and how do they fit together? Think of a busy shop assistant who has seconds to pick one product for each customer walking in. - **Andromeda is the shortlist.** It's Meta's retrieval system. When Meta launched it [in 2024](https://engineering.fb.com/2024/12/02/production-engineering/meta-andromeda-advantage-automation-next-gen-personalized-ads-retrieval-engine/), Meta described retrieval as selecting "from tens of millions of ad candidates into a few thousand relevant ad candidates." The October 2026 version "scores every eligible ad with near ranking-level model performance at retrieval scale" and passes the top candidates on to ranking. - **GEM is the teacher.** GEM (Generative Ads Recommendation Model) is a huge model trained on behaviour across Facebook, Instagram, Reels and every major objective. It's too big to run on every ad request, so Meta transfers what it knows into the "hundreds of vertical models that serve ads in real time." Those smaller models do the actual ranking. - **GenRecs is the understanding.** GenRecs (Generative Recommenders) uses pre-trained multimodal models to read "each ad, product, and page, including its text and images," then compresses that into a short string of tokens. Similar items share tokens. Meta uses these tokens "for both retrieval and ranking." So there are two steps, retrieval then ranking, and GenRecs feeds both. Meta doesn't describe GenRecs as a separate stage. Its tokens are an input that Andromeda uses (alongside your targeting rules) when it builds the shortlist, and that the GEM-taught ranking models use when they pick the winner. The auction then weighs that prediction against your bid and ad quality. ![Diagram of how Meta picks an ad: Andromeda shortlists, GEM-taught models rank, the auction decides, and GenRecs feeds retrieval and ranking](https://prod.superblogcdn.com/site_cuid_clo00o2d0644641vqp8vh8w6cd/images/three-layers-1793253548272-compressed.jpg) Here is the same thing as a table you can share with your team. Layer Plain-English job What it reads What it means for you GenRecs (understanding) Describes what each ad or product _is_ Ad and product text and images, plus how people engaged with the items New ads and new SKUs start from a sensible guess Andromeda (retrieval) Builds the shortlist Your ad, the person's signals, your targeting rules Every eligible ad now gets scored, so more of your ads get a chance GEM to ranking models (ranking) Picks the best ad for this person Long history of what people clicked and bought across Meta Conversion signals and value matter more than ever For the deeper story of Andromeda itself, and how it changed creative strategy since 2024, read our [Meta Andromeda update creative strategy playbook](https://segwise.ai/blog/meta-andromeda-update-creative-strategy-2026). This post focuses on what is new: GEM, GenRecs, and what the October 2026 releases mean. ## Why can a brand-new ad start learning faster now? Because the system no longer treats a new ad as a blank ID. Before semantic tokens, a new ad was a new number with no history. The model had to wait for impressions and clicks to learn anything. Meta's [GenRecs Tokenization post](https://engineering.fb.com/2026/10/19/ml-applications/genrecs-tokenization-transforming-ads-recommendation-at-meta-with-stable-model-vocabulary/) explains the change with a shoe example: because tokens describe what an item is, "a newly launched 'trail shoe' inherits what the model has already learned about that category, rather than starting a new representation with no signal." Meta lists the new things this helps: "newly created ads, niche products, and pages with limited engagement." It also says the tokens let models "carry insights across campaigns," so a new campaign isn't starting from zero either. GEM adds a second boost. Meta says "representations learned on data-rich click tasks transfer into data-sparse conversion tasks." In plain terms: GEM learns from the huge volume of clicks across Meta and uses that to make better guesses about conversions, which are much rarer. A new ad with few conversions benefits from that. The measured effect is real but modest. Meta reports GenRecs Tokenization, paired with GEM and the Adaptive Ranking Model, "delivered a +0.35% lift in conversions," measured in February 2026. Meta measured that lift across surfaces. Small in percentage terms, it shows GenRecs improves the starting guess; it doesn't replace testing. ### Does the learning phase still matter? Yes. Meta's [learning phase article](https://www.facebook.com/business/help/112167992830700) still says an ad set leaves learning "after about 50 results in the week after the ad set's last significant edit," and that during learning, ad sets "usually have a higher CPA." Creating a new ad still triggers it. What changed is the quality of the first guess, not the need to learn. Read it this way: - A new ad that is clearly similar to your proven ads (same product, a new angle) should find its footing sooner, because it shares tokens with things Meta already knows. - A new ad that is genuinely new (new product, new audience, new message) still needs results to prove itself, and it may take the full learning period. - Constant edits still reset learning. Meta's advice is to edit only "when you have reason to believe that the edit should improve performance." ## What did Meta actually change in 2026? There wasn't one "update." Meta published several pieces in 2026. Here's what each one says, in date order, with only Meta's own numbers. Date What Meta published What changed Meta's reported result Jan 2026 [Newsroom post on 2026 AI performance](https://about.fb.com/news/2026/01/2026-ai-drives-performance/) Doubled the GPUs used to train GEM in Q4 2025, plus "a new, more efficient sequence-learning architecture" 3.5% more ad clicks on Facebook and more than 1% more conversions on Instagram in Q4 2025 Mar 31, 2026 [Adaptive Ranking Model](https://engineering.fb.com/2026/03/31/ml-applications/meta-adaptive-ranking-model-bending-the-inference-scaling-curve-to-serve-llm-scale-models-for-ads/) LLM-scale ranking model served in real time, routing each request to the right model size +3% conversions and +5% CTR for targeted users on Instagram since Q4 2025 Oct 13, 2026 [GEM 2026 generation](https://engineering.fb.com/2026/10/13/ml-applications/generative-ads-model-gem-scaling-frontier-ads-ai/) One central model instead of separate Facebook, Instagram and Reels models; 2.5x more model complexity; 2x longer engagement history in production 18.9% conversion improvement and 12% CTR improvement across ad surfaces and objectives Oct 19, 2026 [GenRecs Tokenization](https://engineering.fb.com/2026/10/19/ml-applications/genrecs-tokenization-transforming-ads-recommendation-at-meta-with-stable-model-vocabulary/) Ads, products, pages and pixels described with shared semantic tokens for retrieval and ranking +0.35% conversions (measured February 2026) Oct 28, 2026 [Next-generation Andromeda](https://engineering.fb.com/2026/10/28/ml-applications/meta-andromeda-next-gen-ads-retrieval-gpu-co-design/) Retrieval now uses a mixture of specialised "expert" models and scores all eligible ads at over one billion ads per second, 100x the 2024 system 2.4% more conversions on Facebook since 2025 A few details matter for advertisers. **Andromeda now rates all eligible ads.** Meta says the new system "scores every eligible ad with near ranking-level model performance at retrieval scale." Its user side is now a set of specialised "expert" models, fed by engagement history, cohort signals, "LLM-derived semantic features" and cross-surface signals. A router sends each candidate ad to the experts most relevant to it. Meta also notes that generative AI tools helped advertisers create "hundreds of millions of new ads," which is why retrieval had to scale. **GEM became one brain.** The 2026 GEM "jointly predicts clicks and conversions across all surfaces, including Facebook and Instagram, and across every key advertiser objective from off-site conversions to mobile app installs." Before this, in Meta's words, "each surface and optimization objective previously had its own independent prediction system." **Your targeting rules still run.** Andromeda still does "index matching (i.e., enforcing advertisers' targeting rules)" before scoring. Targeting isn't ignored. It decides which ads are eligible; the models decide which eligible ad wins. ## How many creatives and how much variety do you need now? Fewer ads than most people think, but more _different_ ones. This is where advice online often contradicts Meta itself. Meta's [ad volume article](https://www.facebook.com/business/help/2720085414702598) says plainly: "too many ads can result in worse performance," because each ad delivers less often and fewer ads exit learning. Its fix is: "Decrease ads per ad set, but maintain diverse creative assets per ad set. One ad can contain multiple (up to ten) creative assets." GenRecs explains why _diverse_ matters. Similar items share tokens. So if you upload ten versions of one ad with a different background colour, the model likely sees them as close neighbours of the same thing. You've added count without adding new options for the system to match to new people. That's our reading of how tokens work, not a Meta statement, but it lines up with Meta's "diverse creative assets" wording. ### Rule of thumb: how to size your creative mix These are Segwise rules of thumb, built on Meta's learning-phase number. Adjust them to your account. 1. **Count the ad sets you can feed.** Divide your weekly optimisation events by 50. An account with 210 purchases a week can support about 4 ad sets that reliably exit learning. More ad sets than that, and some will sit in "Learning limited." 2. **Run 3 to 6 distinct concepts per ad set,** not 30 near-copies. Put variations of each concept inside one ad as assets (Meta allows up to ten). 3. **Make each new concept differ on at least 2 of 5 things a model can read:** the person or character on screen, the setting, the product shown, the core message or offer, and the format (static, video, UGC-style video). A new headline on the same image is a variation, not a concept. 4. **Replace, don't pile on.** When a concept fatigues, swap in a new concept instead of adding more ads to the ad set. For a step-by-step version of this testing loop, see our guide on [how to beat Meta's Andromeda algorithm with creative strategy](https://segwise.ai/blog/beat-meta-andromeda-algorithm-creative-strategy). > **Tip:** Before you call something a "new concept," describe it in one sentence without mentioning colour, font or button text. If the sentence matches a live ad, it's a variation. ## Should you consolidate campaigns and drop manual audiences? Mostly yes, with clear exceptions. Meta's [combine ad sets article](https://www.facebook.com/business/help/2419480091640105) gives the reasoning with a simple example: three ad sets aimed at people interested in dogs, dog toys and dog collars "could decrease performance because your total audience size is split." Combining them gives each ad set more results, so it leaves learning sooner. For reporting, Meta suggests using breakdowns (such as age, gender, country or region) instead of splitting ad sets. Our reading, not a Meta statement: with GEM learning across surfaces and Andromeda scoring every eligible ad, splitting by narrow interests likely adds less than it used to. Meta's own pages make the narrower point that split audiences fragment learning. **Keep separate campaigns or ad sets when:** - **You only want new customers.** Meta's [customer lifecycle strategy](https://www.facebook.com/business/help/1740191093676743) lets a Sales campaign "Acquire new customers" and exclude existing buyers. Meta warns this "may increase your cost per result," so judge it on new-customer CPA, not blended CPA. Keep it apart from your all-customer campaign. - **The goal or the value is different.** A campaign optimising for purchase value and one optimising for installs or leads are learning different things. Don't merge across objectives. - **The economics differ by platform or market.** Meta's [Advantage+ app campaigns](https://developers.facebook.com/docs/app-ads/advantage-app-campaigns) accept only location and language targeting and "do not support operating system targeting." If iOS and Android users are worth very different amounts to you, a manual app campaign per platform may still make sense. - **You need a clean test.** If you're proving a new offer or landing page, a separate test with a holdout is easier to read than a blended ad set. - **Each split still clears about 50 results a week.** If it does, the split costs you little. If it doesn't, merge. If you saw ROAS slip after consolidating, our breakdown of [why ROAS drops after Andromeda and how to fix it](https://segwise.ai/blog/meta-andromeda-roas-drop) walks through the usual causes. ## Which inputs matter more now? When the models do more of the matching, what you feed them decides what they optimise for. Four inputs carry the most weight. ![Checklist of four inputs Meta's models learn from: conversion events, value of each sale, customer goal and post-click quality](https://prod.superblogcdn.com/site_cuid_clo00o2d0644641vqp8vh8w6cd/images/inputs-checklist-1793253550793-compressed.jpg) **1\. Conversion events that reach Meta.** GEM leans on conversion data that is sparse to begin with. If your events are missing or delayed, the model has less to learn from. Meta's [value optimisation article](https://www.facebook.com/business/help/296463804090290) says value signals need "Conversions API (required) and Meta pixel." **2\. The value of each sale.** Meta's own example: with a USD 200 budget, optimising for purchases got four sales worth USD 250 (ROAS 1.25), while optimising for value got three sales worth USD 500 (ROAS 2.5). If your orders vary a lot in size, send the value. Meta also supports predicted lifetime value (about two weeks to set up) and profit (about one week). **3\. Whether you want new or all customers.** Covered above. Pick on purpose, because the model will chase whichever you choose. **4\. What happens after the click.** Meta's auction scores [ad quality](https://www.facebook.com/business/help/430291176997542) using "feedback from people viewing or hiding the ad" and checks for things like "withholding information" and "sensationalised language." Meta notes that "an ad that's more relevant to a person could win an auction against ads with higher bids." Its [relevance diagnostics](https://www.facebook.com/business/help/403110480493160) point you to "creative assets, post-click experience or audience targeting." And Meta has said since 2018 that businesses with poor customer feedback, especially "inaccurate shipping times" or products that are misrepresented, can have their ad delivery reduced. Honest shipping times and return policies are an ad performance input, not just a customer service one. ## How does this change catalog ads for large-SKU brands? Large catalogs are where Meta says GenRecs helps most. Meta says tokenization "maximizes the efficiency of high-volume campaigns by uncovering deeper semantic patterns across their vast product catalogs." For "tail" products that get few impressions, the models can now "generalize across them via shared content structure." And its example is a fashion-style case: when a sportswear brand launches next season's trail shoe, the new product and its ads "build on what the model learned from last season's shoe, even before anyone sees the new ad." ### How does product selection work in catalog campaigns now? Meta has renamed Advantage+ catalog ads to [ads with product media](https://www.facebook.com/business/help/397103717129942). They show "the most relevant products or services from your catalogue based on their interests, intent and actions online." In Sales campaigns, product data is on by default if you have a catalog, and even ads that don't show product images "can still benefit from your product data." Put that next to GenRecs, and product selection for a large fashion catalog likely works like this: Andromeda scores every eligible product ad for a person, GenRecs tokens tell it which products are similar, and a brand-new SKU gets placed near the products it resembles. The product's text and images decide who its "neighbours" are. ### What should you fix in your product feed? Meta's [catalog best practices](https://developers.facebook.com/docs/marketing-api/catalog/best-practices) give the fields. How they connect to GenRecs is our reading, not Meta's statement: Meta says GenRecs reads each product's "text and images," and these fields are where that text and those images come from. - **Titles:** name the product type and the attribute shoppers choose on (for example "Women's waterproof trail running shoe, navy"), not just a style code. The title is likely part of the text the model reads. - **Variant fields:** Meta asks that "the variant field's (such as size or color) value is provided for every product variant" and that variants share one `item_group_id`. That tells the system these are one product, not ten. - **Brand and IDs:** provide `gtin` or `mpn` plus `brand`. - **Images:** add extra images with `additional_image_link` (up to 10). Clean product shots likely help the model see what the item is. - **Descriptions:** Meta suggests details such as size, volume and origin. Write material and fit here. ### How should you restructure catalog campaigns? Rule of thumb: split product sets only where the business truly differs, such as margin band, price tier or a new-season launch you want to push. Don't split a catalog into dozens of tiny product sets "for control." Each one has to earn its 50 results a week, and GenRecs already groups similar products for you. ## Does this apply to app install campaigns and mobile games? Partly, and here's exactly what Meta has said. - **GEM: yes.** Meta trained the 2026 GEM on "deep funnel optimization (e.g., off-site conversion, mobile app install)" and says it predicts across objectives "from off-site conversions to mobile app installs." - **Andromeda: yes.** It retrieves for all ads across Facebook and Instagram. - **Value: yes.** Meta's value optimisation "is also available to optimise app promotion campaigns for purchase values." - **GenRecs: not stated.** Meta's GenRecs post names "ads, products, pages, and pixels." It doesn't mention app installs or games specifically, and it reports no app-specific results. Your app ads are still ads, so they're likely covered, but Meta hasn't published proof. For game studios, the practical points are the same: concepts that differ in what a model can see (gameplay moment, character, hook, format), clean post-install events through your MMP or SDK, and a purchase value signal if you run value campaigns. Meta's developer docs list up to 50 creative combinations for app campaigns. ## How Segwise helps you feed the system the right variety GenRecs rewards variety the model can read. The hard part is knowing which of your ads are really different and which are the same concept in a new colour. That's an element-level question, and it's what Segwise is built for. ![Segwise Creative Analytics view showing Echo ads grouped into concepts by shared tags, with spend share and CPI](https://prod.superblogcdn.com/site_cuid_clo00o2d0644641vqp8vh8w6cd/images/segwise-concept-clusters-1793253552117-compressed.jpg) - **See which concepts are really distinct.** [Creative tagging](https://segwise.ai/features/creative-tagging) labels the hook, person, setting, product, message and format of every ad, video and playable. Asset clustering groups ads that share footage or overlapping tags, so you can spot ten "new" ads that are one concept. - **See which attributes win.** [Creative analytics](https://segwise.ai/features/creative-analytics) maps every tag to ROAS, CPA, IPM and CTR across Meta and your other networks, with the option to use your MMP data as the source of truth. - **Make new concepts from what wins.** [Creative generation](https://segwise.ai/features/creative-generation) helps your team draft new statics, videos and AI UGC-style videos from your own winning patterns, ready to edit and export to Meta's specs. - **Check before you spend.** [Pre-flight testing](https://segwise.ai/features/preflight-testing) scores up to 10 creatives out of 100 against your own winners, including how novel each one is versus what's already saturated in your account. - **Know when to replace.** [Fatigue tracking](https://segwise.ai/features/fatigue-tracking) alerts you by email or Slack when a concept starts to decline, so you swap in a new concept on time. For game studios, Segwise also plays each playable ad end to end and tags its mechanic, hook scene, end card and CTA, so playables get the same element-level read as video and statics. **When you don't need it, and what it doesn't do.** If you advertise only on Meta with a handful of live ads, you can group concepts by hand and read Ads Manager breakdowns; a tool earns its place once you run dozens of ads across several networks (a rule of thumb). Segwise doesn't change how Meta ranks or delivers your ads, can't see Meta's GenRecs tokens (its tags are a proxy for what Meta's model reads), doesn't buy media, and doesn't fix your Conversions API or value-signal setup. Find which of your ads are truly different Segwise tags every creative element against performance and helps you build new concepts from your winners [Start free trial](https://ua.segwise.ai) [Book a demo](https://cal.com/segwise-bb/segwise-demo?overlayCalendar=true) ## A one-week checklist to act on this 1. Count weekly conversions per ad set. Merge any ad set that can't reach about 50 a week. 2. Group your live ads by concept. Cut near-copies and keep 3 to 6 distinct concepts per ad set. 3. Brief 2 new concepts that differ on at least 2 of the 5 readable attributes. 4. Check that purchase values reach Meta through Conversions API, then test value optimisation. 5. Decide, per campaign, whether you want new customers or all customers. 6. Audit your catalog: titles with product type and key attribute, size and colour on every variant, brand, extra images. 7. Fix post-click basics: honest shipping times, a clear returns policy, and a landing page that matches the ad. ## FAQs Q: What is the difference between Andromeda, GEM and GenRecs? A: They are three parts of how Meta picks an ad. Andromeda is the retrieval step: in 2024 Meta described it as narrowing tens of millions of ads to a few thousand candidates, and its October 2026 version scores every eligible ad before passing the top candidates to ranking. GEM (Generative Ads Recommendation Model) is a large foundation model that teaches the hundreds of smaller models that rank those candidates in real time. GenRecs (Generative Recommenders) reads each ad's and product's text and images and turns them into shared semantic tokens, which Meta uses in both retrieval and ranking. Q: Does GenRecs mean a new Meta ad skips the learning phase? A: No. GenRecs gives a new ad a better starting guess by relating it to similar ads and products Meta already knows, which Meta says helps newly created ads, niche products and pages with little engagement. But Meta's Help Centre still says an ad set leaves the learning phase after about 50 results in the week after its last significant edit, and that CPA is usually higher while learning. Plan for the learning phase as before. Q: Does Meta read the text and images in my ads? A: Yes. Meta's GenRecs Tokenization post says pre-trained multimodal models read each ad, product and page, including its text and images, and combine that with how people engaged with the items. The result is a short set of tokens, and similar items share tokens. That is why the visible content of an ad, not just its ID or targeting, now shapes who it is matched with. Q: Do near-duplicate ads hurt performance on Meta? A: They usually add little. Meta's Help Centre says running too many ads at once can mean worse performance, because each ad delivers less often and fewer ads exit learning. Its advice is fewer ads per ad set with diverse creative assets, up to ten assets in one ad. Because similar items share GenRecs tokens, colour or headline swaps of one ad likely read as one item. Put variations inside one ad and save new ads for new concepts. Q: How many ad sets can my Meta budget support? A: A simple rule of thumb: divide your weekly optimisation events by 50, the number of results Meta says an ad set usually needs in a week to exit learning. An account with 210 purchases a week can support about four ad sets that reliably exit learning. If you run more ad sets than that, combine them and use breakdowns for reporting, as Meta's Help Centre recommends. Q: Do I need Conversions API for Meta value optimisation? A: Yes. Meta's Help Centre says that to use maximise value of conversions you need to send value signals through Conversions API, which is required, and the Meta pixel. Value optimisation is also available for app promotion campaigns optimising for purchase value. Predicted lifetime value takes about two weeks to set up and profit optimisation about one week, according to Meta. Q: Does GEM work for app install campaigns and mobile games? A: Meta says the 2026 GEM was trained on deep funnel data including mobile app installs, and that it predicts clicks and conversions across every key objective from off-site conversions to mobile app installs. Meta's GenRecs post names ads, products, pages and pixels but does not report app-specific results. For games, the practical steps are the same: distinct concepts, clean post-install events and a purchase value signal. Q: How should a large-SKU fashion brand set up catalog ads now? A: Fill in the product data that likely supplies the text and images GenRecs reads: titles that name the product type and key attribute, size and colour on every variant with one shared item_group_id, brand plus gtin or mpn, and extra images (Meta allows up to 10). Meta now calls these ads with product media and turns product data on by default in Sales campaigns. As a rule of thumb, split product sets only where margin, price tier or a launch truly differ. --- This blog is powered by Superblog. Visit https://superblog.ai to know more. ---