[{"data":1,"prerenderedAt":572},["ShallowReactive",2],{"guide-meta-andromeda-creative":3},{"id":4,"title":5,"answer":6,"authorId":7,"body":8,"category":454,"ctaVariant":455,"dataset":454,"description":456,"examples":457,"extension":458,"faqs":459,"heroImage":484,"intro":485,"meta":486,"navigation":487,"path":488,"publishedAt":489,"seo":490,"sources":491,"stats":528,"stem":570,"updatedAt":489,"__hash__":571},"blog\u002Fblog\u002Fmeta-andromeda-creative.md","Meta Andromeda: Impact on Ad Creative (2026)","Meta Andromeda is the AI-powered retrieval stage of Meta's ads delivery system, announced by Meta engineering in December 2024 and built on NVIDIA Grace Hopper hardware. It screens tens of millions of candidate ads down to a few thousand before Meta's ranking models (Lattice, then the foundation-scale GEM) pick a winner, with retrieval model capacity Meta says grew 10,000x. Nothing forces you to rebuild campaigns, but delivery now rewards advertisers who supply genuinely distinct concepts, not near-duplicate variations: diverse angles get matched to diverse buyers, while a single hero ad limits how often the system can choose you.","likit-sae-lee",{"type":9,"value":10,"toc":440},"minimark",[11,16,20,27,30,33,36,103,106,110,113,116,119,123,126,129,137,199,202,206,209,212,215,219,222,225,229,232,235,241,244,247,270,273,277,280,313,316,319,323,326,329,405,408,412,415,418,421,425,428,431],[12,13,15],"h2",{"id":14},"the-part-of-metas-ads-system-andromeda-changed","The part of Meta's ads system Andromeda changed",[17,18,19],"p",{},"Meta's ad delivery runs as a funnel. At the moment a feed loads, the system holds tens of millions of live ad candidates that could fill the slot. No ranking model can score that many in the milliseconds available, so delivery happens in stages: a retrieval stage first narrows the pool to a few thousand plausible candidates, then heavier ranking models score those finalists, and the auction decides what appears. Andromeda is that first stage, rebuilt.",[17,21,22],{},[23,24],"img",{"alt":25,"src":26},"Diagram of Meta's three-stage ad delivery funnel: Andromeda retrieval screens tens of millions of live ad candidates down to a few thousand, the Lattice then GEM ranking models score those finalists, and the auction picks the single ad shown to one person.","\u002Fimages\u002Fblog\u002Fmeta-andromeda-creative-delivery-funnel.webp",[17,28,29],{},"Meta's engineering team announced it on December 2, 2024 as a personalized ads retrieval engine running on the NVIDIA Grace Hopper Superchip alongside Meta's own MTIA accelerator hardware. The job did not change: select from tens of millions of ad candidates down to the few thousand that move on to ranking. What changed is how much intelligence sits in that step. Older retrieval leaned on hand-engineered features and had to stay simple to stay fast. The new hardware let Meta push far larger neural networks into the stage, learning subtler patterns between people and ads on the fly.",[17,31,32],{},"Most advertisers never had a reason to know retrieval existed. The name escaped engineering circles because the rebuild was big enough for Meta to publish, big enough for its CFO to cite to investors, and visible enough in ad accounts that practitioners went looking for a cause.",[17,34,35],{},"The shift is easiest to see laid side by side. The dimensions below describe the direction of travel Meta's posts and earnings calls point to, not a switch that flipped on one date.",[37,38,39,55],"table",{},[40,41,42],"thead",{},[43,44,45,49,52],"tr",{},[46,47,48],"th",{},"Dimension",[46,50,51],{},"Older retrieval",[46,53,54],{},"Andromeda",[56,57,58,70,81,92],"tbody",{},[43,59,60,64,67],{},[61,62,63],"td",{},"What the stage holds",[61,65,66],{},"Tens of millions of candidates, narrowed with simpler logic",[61,68,69],{},"Tens of millions of candidates, narrowed by far larger neural networks",[43,71,72,75,78],{},[61,73,74],{},"What it leans on",[61,76,77],{},"Hand-engineered features, kept simple to stay fast",[61,79,80],{},"Learned patterns between people and ads, run on Grace Hopper plus MTIA hardware",[43,82,83,86,89],{},[61,84,85],{},"Primary lever for the advertiser",[61,87,88],{},"Audience definition and interest stacks",[61,90,91],{},"Breadth of genuinely distinct creative",[43,93,94,97,100],{},[61,95,96],{},"Model capacity (Meta's figure)",[61,98,99],{},"Baseline",[61,101,102],{},"10,000x larger",[17,104,105],{},"The right-hand column is design intent. The rest of this page is where the evidence for its consequences is strong and where it runs out.",[12,107,109],{"id":108},"what-meta-has-claimed-on-the-record","What Meta has claimed, on the record",[17,111,112],{},"The verifiable claims come from two places: the engineering announcement and earnings calls. The engineering post reports a 6% recall improvement to the retrieval system (it surfaces more of the ads it should), an 8% ads quality improvement on selected segments, a 100x improvement in feature extraction latency and throughput, and a 10,000x increase in model capacity for personalization. On Meta's Q4 2024 earnings call, CFO Susan Li told investors Andromeda \"has led to a 10,000x increase in the complexity of the models that Meta uses for ads retrieval\", as reported by AdExchanger in January 2025. The same call put Meta's generative AI ad tools at more than 4 million advertisers, up from 1 million six months earlier: a separate program, but part of the same AI-first delivery push.",[17,114,115],{},"Retrieval is half the story, and Meta rebuilt the ranking half on its own timeline. The two changes are easy to blur together, so it helps to keep them apart by name and date.",[17,117,118],{},"The supply side moved in lockstep. On Meta's Q3 2024 earnings call, reported by Marketing Dive in October 2024, the company said more than a million advertisers used its generative AI tools to create over 15 million ads in a single month, and that businesses using its Image Generation feature saw a 7% increase in conversions. Those numbers describe a separate program from Andromeda, but they point the same way: Meta is building both a delivery engine that personalizes across enormous creative variety and the tools to manufacture that variety at volume. That is the full list of what Meta has said about the engine itself. Everything else is interpretation.",[12,120,122],{"id":121},"how-andromeda-lattice-and-gem-split-the-work","How Andromeda, Lattice and GEM split the work",[17,124,125],{},"It is tempting to treat \"the algorithm\" as one black box, but Meta's own posts describe three named layers, each rebuilt at a different point and each with its own published figure. Keeping them straight is the difference between guessing and citing.",[17,127,128],{},"The first layer is retrieval, and that is Andromeda: announced December 2, 2024, it screens the tens of millions of eligible ads down to the few thousand that move on, and Meta credits it with an 8% increase in ads quality that, in Meta's words, increases the impact of Advantage+ automation and its creative generative AI tools.",[17,130,131,132,136],{},"The second layer is ranking, and here Meta moved twice. In a March 27, 2025 post, Meta described Lattice, a single unified ranking architecture that \"generalize",[133,134,135],"span",{},"s"," learnings across campaign objectives and surfaces in place of numerous, smaller ads models\", and credited it with raising ad quality by almost 12% and lifting conversions by up to 6%. Then, in a November 10, 2025 engineering post, Meta announced GEM, its Generative Ads Model, as the model now driving ranking. Meta calls GEM \"the largest foundation model for recommendation systems in the industry, trained at the scale of large language models\", reporting that since its Q2 2025 launch it delivered a 5% increase in conversions on Instagram and a 3% increase on Facebook Feed, and that Q3 2025 architectural refinements doubled how much performance it gains from each unit of added data and compute. The training scale is the tell of how serious the bet is: a 23x increase in effective training FLOPs using 16x more GPUs, with a knowledge-transfer framework Meta says is twice as effective as standard distillation.",[37,138,139,155],{},[40,140,141],{},[43,142,143,146,149,152],{},[46,144,145],{},"Layer",[46,147,148],{},"Model (announced)",[46,150,151],{},"Job in delivery",[46,153,154],{},"Meta-stated figure",[56,156,157,171,185],{},[43,158,159,162,165,168],{},[61,160,161],{},"Retrieval",[61,163,164],{},"Andromeda (Dec 2024)",[61,166,167],{},"Screen tens of millions of ads to a few thousand",[61,169,170],{},"+8% ads quality; 10,000x retrieval model capacity",[43,172,173,176,179,182],{},[61,174,175],{},"Ranking, unified",[61,177,178],{},"Lattice (Mar 2025)",[61,180,181],{},"One model across objectives, replacing many smaller ones",[61,183,184],{},"+12% ad quality; up to +6% conversions",[43,186,187,190,193,196],{},[61,188,189],{},"Ranking, foundation-scale",[61,191,192],{},"GEM (Nov 2025)",[61,194,195],{},"Score finalists at LLM training scale",[61,197,198],{},"+5% conversions on Instagram; +3% on Facebook Feed",[17,200,201],{},"Read down that column and the \"both halves got smarter\" claim stops being a slogan. Andromeda decides what can be shown; Lattice and then GEM decide what is shown. Every figure is Meta's own, attached to a date, which is exactly the discipline the rest of this page asks of the practitioner readings below.",[12,203,205],{"id":204},"why-meta-is-betting-its-ad-business-on-ai-delivery","Why Meta is betting its ad business on AI delivery",[17,207,208],{},"The reason any of this is worth a creative strategy rethink, rather than a shrug, is the money sitting behind it. This is not a feature Meta bolted on. It is the core of how the company now sells advertising, and the figures from its own earnings calls make the scale concrete.",[17,210,211],{},"On the call covering Q4 2024 (reported January 29, 2025), Meta said its Advantage+ shopping campaigns had surpassed a $20 billion annual revenue run-rate, growing 70% year over year. By the Q3 2025 call (October 29, 2025), Meta put its end-to-end AI-powered advertising tools, the Advantage+ suite included, at a $60 billion-plus annual run-rate, with the number of advertisers using its video-generation tools up 20% over the prior quarter and average price-per-ad up 10% year over year. The same quarter, CFO Susan Li told investors that advertisers running lead campaigns through Advantage+ were seeing a 14% lower cost per lead on average than those who were not.",[17,213,214],{},"Those numbers matter to a creative team for one reason: they tell you the AI-first delivery model is permanent. When a system carries tens of billions in run-rate and is still growing, Meta keeps sharpening it rather than quietly rolling it back. So the implication for how you produce creative is durable, not a trend to wait out: adapting your creative supply to how this engine matches is aligning with the direction Meta has staked its ad revenue on, not chasing a quarter's algorithm tweak.",[12,216,218],{"id":217},"what-practitioners-infer-and-where-the-evidence-stops","What practitioners infer, and where the evidence stops",[17,220,221],{},"The most common practitioner readings are worth taking seriously. First, that creative now functions as targeting, because the system reads format, hook, and subject matter to predict who responds. Second, that near-duplicate variations buy little, because retrieval treats five crops of one concept as one idea. Third, that creative fatigue arrives sooner, because a sharper system finds your responsive audience faster and exhausts it faster. All three are plausible, consistent with how retrieval works, and reported across enough ad accounts to respect. None of them is a Meta statement.",[17,223,224],{},"Be more skeptical of three other claims. The dated \"Andromeda update\" that supposedly hit accounts on a specific day: the rollout was progressive across late 2024 and 2025, and Meta does not flag accounts, so a precise before-and-after date is a story, not a record. Fixed creative-count rules: the idea that you need 20 or more creatives per ad set circulates everywhere, and AdExchanger's January 2026 analysis names it a misconception, because when creative supply dramatically outweighs the budget, learning slows and delivery fragments. And the hope that a smarter system compensates for weak inputs: the same analysis is blunt that Andromeda does not change the fundamentals, it raises the bar on discipline.",[12,226,228],{"id":227},"why-distinct-creative-gets-retrieved-more","Why distinct creative gets retrieved more",[17,230,231],{},"The mechanism that makes creative breadth matter requires no leap of faith. A retrieval engine personalizes across whatever variety you give it. Meta built a system that can hold tens of millions of candidates and pick a personalized few thousand for every person who opens the app. The personalization it can do for your account is bounded by how many genuinely different things you have given it to retrieve.",[17,233,234],{},"Run one hero ad plus four resizes, and every buyer Meta considers gets matched against a single idea. Buyers motivated by price never see proof. Buyers who need to watch the product work never see a demo. Those people still scroll, and the slot goes to another advertiser whose pool contains an angle that fits them. A spread of distinct angles (problem-led, testimonial, offer, demonstration, founder-voice) gives retrieval different entries that map to different buyer motivations, which is exactly the matching a 10,000x model was built to do. The bar for each new ad follows from this: does it give the system a genuinely new reason to pick you, for someone your current ads miss?",[17,236,237],{},[23,238],{"alt":239,"src":240},"Side-by-side comparison: one hero ad resized five ways reads as a single idea so most buyers go unmatched, while five genuinely distinct angles each match a different buyer motivation, giving retrieval five separate ways to choose your ad.","\u002Fimages\u002Fblog\u002Fmeta-andromeda-creative-variation-vs-diversification.webp",[17,242,243],{},"That lever has a name worth internalizing: the split between variation and diversification, because most creative work that feels productive is actually variation. Variation refines one idea: a new crop, a swapped button color, \"shop\" in place of \"buy\". Diversification replaces the idea: a different reason a different person buys. Meta has not published the dimensions retrieval uses to judge sameness, so resist any precise similarity threshold quoted as fact. The practical test holds without a number: if two ads would persuade the same person for the same reason, the system gains little by holding both.",[17,245,246],{},"A reliable way to force real diversification is to vary the inputs that change who an ad is for, not just how it looks. Three inputs do most of the work:",[248,249,250,258,264],"ul",{},[251,252,253,257],"li",{},[254,255,256],"strong",{},"Awareness level."," A problem-aware buyer needs the problem named; a product-aware buyer needs proof and a reason to choose you now. Those are different ads, not different crops.",[251,259,260,263],{},[254,261,262],{},"Objection."," Price, trust, effort, and \"will it work for me\" are distinct walls. Each objection answered is a distinct concept.",[251,265,266,269],{},[254,267,268],{},"Format and proof type."," A founder talking to camera, a customer testimonial, a side-by-side demonstration, and a clean offer graphic carry different evidence to different temperaments.",[17,271,272],{},"Cross two of those and you have a brief that produces genuinely distinct creative on purpose. The same product, written as \"problem-aware buyer, trust objection, testimonial\" and again as \"product-aware buyer, price objection, offer graphic\", yields two ads the system reads as two options.",[12,274,276],{"id":275},"a-five-angle-playbook-from-a-single-product","A five-angle playbook from a single product",[17,278,279],{},"To make that concrete, take one product, a daily greens supplement, and expand it into five distinct briefs (each crossing an awareness level with an objection) instead of five crops of one hero shot.",[281,282,283,289,295,301,307],"ol",{},[251,284,285,288],{},[254,286,287],{},"Problem-led, problem-aware buyer."," Open on the 3pm energy crash and the reach for coffee. The angle is the problem, not the product, reaching people who have not yet decided a supplement is the answer.",[251,290,291,294],{},[254,292,293],{},"Demonstration, \"will it work for me\" objection."," Show the scoop, the glass, the morning routine, the thirty-day change in how someone feels. This answers the buyer who believes the category but doubts this product.",[251,296,297,300],{},[254,298,299],{},"Testimonial, trust objection."," A real customer in their kitchen explaining what changed. This carries social proof to the skeptic that no founder claim can.",[251,302,303,306],{},[254,304,305],{},"Founder or UGC, identity objection."," The founder on why they formulated it, or a creator who fits the buyer's self-image. This reaches people who buy from brands that feel like them.",[251,308,309,312],{},[254,310,311],{},"Offer graphic, price objection."," A clean subscription discount or bundle, aimed at the price-sensitive buyer ready to act.",[17,314,315],{},"Five briefs, five buyer motivations, one product: the difference between feeding retrieval five entries and one entry five times.",[17,317,318],{},"None of this is hypothetical. Real ad examples from a crowded supplement category already run the full spread across competing brands: NMN3+ Pro leans on testimonial proof (a mum's wrinkles easing after drinking it), FlexiGold puts its founder on camera to argue that joint pain is not only about joints, and Befree takes the price lane with an eye-health drink at RM20 for new buyers. Those are representative angles, not a live feed, but the lesson transfers to any market. Three different reasons to be picked, three different buyers who stop scrolling. The advertiser who fields several of those reasons at once is the one retrieval can choose most often.",[12,320,322],{"id":321},"read-and-refresh-by-angle","Read and refresh by angle",[17,324,325],{},"The reporting change matters as much as the production change. When retrieval matches distinct angles to distinct buyers, per-ad cost per result reflects the audience slice that one ad was paired with, not the angle's true worth, and the same concept can read strong in one ad and weak in a near-twin. So tag every live ad with the angle it represents and read performance at the angle group: sum spend, results, and cost per result by tag. The decision you are making is which buyer motivations to keep supplying, and that lives at the angle level, not the ad ID.",[17,327,328],{},"Refresh works the same way: by angle, on triggers. Creative fatigue arrives sooner under a sharper retrieval stage, because the system finds and exhausts an ad's best-matching audience faster. Meta publishes no fatigue window, and practitioner estimates vary widely, so run a cadence you can sustain and let live signals override the calendar.",[37,330,331,347],{},[40,332,333],{},[43,334,335,338,341,344],{},[46,336,337],{},"Signal to watch",[46,339,340],{},"Practitioner trigger",[46,342,343],{},"What it means",[46,345,346],{},"Response",[56,348,349,363,377,391],{},[43,350,351,354,357,360],{},[61,352,353],{},"Frequency",[61,355,356],{},"Above roughly 3-4x in a buying window",[61,358,359],{},"The same people are seeing the ad repeatedly",[61,361,362],{},"Introduce a new angle, not a recrop",[43,364,365,368,371,374],{},[61,366,367],{},"Cost per result",[61,369,370],{},"Up more than about 20% from baseline",[61,372,373],{},"Efficiency is decaying",[61,375,376],{},"Refresh the tired concept",[43,378,379,382,385,388],{},[61,380,381],{},"Click-through rate",[61,383,384],{},"Falling while frequency rises",[61,386,387],{},"Classic fatigue signature",[61,389,390],{},"Retire the ad, promote a tested winner",[43,392,393,396,399,402],{},[61,394,395],{},"Standing cadence",[61,397,398],{},"Every 2-4 weeks regardless",[61,400,401],{},"Keeps the pool fresh ahead of decay",[61,403,404],{},"Ship a few genuinely new concepts",[17,406,407],{},"Treat the thresholds as rules of thumb, not Meta policy. The pattern that matters is the combination: frequency up and results down at once is fatigue, and the fix is a new reason to be picked, not a fresh export of the same one.",[12,409,411],{"id":410},"the-creative-volume-question-answered-honestly","The creative-volume question, answered honestly",[17,413,414],{},"Most readers arrive at this topic with one practical question: how many creatives do I actually need? The honest answer starts with what is not true. There is no Meta-published number, and the rule of thumb that you need 20 or more creatives per ad set is a misconception, because when creative supply dramatically outruns budget, learning slows and delivery fragments. Counting your way to a magic number is the wrong frame.",[17,416,417],{},"A more useful anchor comes from 2026 practitioner guidance, which lands around 10 to 15 conceptually distinct assets per Advantage+ campaign, with the emphasis squarely on distinct: different angles (UGC, studio, text-only, testimonial, offer) rather than minor variations of one concept. Treat that as directional, a vendor playbook range, not Meta policy. The figure is a sanity check, not a target to hit at the cost of quality, and it sits comfortably with the learning-phase math already covered: a campaign with a single broad ad set wants enough distinct concepts inside it to give retrieval real choice, but not so many that each ad starves before the ad set clears its roughly 50 events. Match the count to the budget, not to a number you read.",[17,419,420],{},"The same practitioner sources put creative fatigue under heavy spend or in smaller audiences as soon as 10 to 14 days, which is why the count is never a one-time decision. You are not stocking a shelf; you are feeding a queue. And the pressure on that queue is rising fast. Adobe data cited in 2026 industry coverage has the majority of marketers anticipating content demand to grow more than 5x between now and 2027, driven by personalization, cross-platform campaigns, and the need for more variations per audience segment. Andromeda's appetite for genuinely distinct creative meets a content-demand curve bending upward, which is precisely why \"how many creatives\" stops being a planning question and becomes a production-velocity one. The teams that win are not the ones who guess the right number; they are the ones who can reliably ship the next distinct angle before fatigue takes the last one.",[12,422,424],{"id":423},"the-posture-that-wins-breadth-at-constant-quality","The posture that wins: breadth at constant quality",[17,426,427],{},"Pull the threads together and the posture is simple: widen your pool of genuinely distinct concepts, read results by angle, refresh on triggers, and never let quality slip to hit a count. The free Meta Ad Library shows the active ads of any advertiser, a fast way to see which hooks and formats your category leans on and where the gaps you could own sit. And quality stays constant, because Meta's claims are about matching, not rescuing: a weak fifth angle is still a weak ad, and no retrieval gain turns it into a winner.",[17,429,430],{},"Peak season does not rewrite any of this. During a high-demand window (Black Friday in one market, the 11.11 and 12.12 mega-sales in another) the offer does more of the persuading, so a clear deal is the biggest single lever, but the delivery logic is unchanged: distinct concepts still reach distinct buyers better than one promotion in five aspect ratios. Diversify with purpose around the offer (a gift angle, an urgency angle, a bestseller angle, a first-timer angle) rather than running the same banner everywhere, and load that breadth before demand peaks, because the fatigue clock runs faster when frequency is highest.",[17,432,433,434,439],{},"Do this honestly and the constraint stops being strategy and becomes production: research angles, write each as a distinct brief, make the image and video, and ship to Meta on a schedule. That is a velocity problem, and it yields to process and tooling. A platform like AdPlay.ai keeps the research, generation, editing, and Meta launch in one place, which shortens the loop, but any workflow that reliably turns one more distinct angle into one more live ad each week is doing the job. If campaign setup is what you need next, the walkthrough in ",[435,436,438],"a",{"href":437},"\u002Fblog\u002Fhow-to-run-a-facebook-ad","how to run a Facebook ad"," covers it. The delivery system is ready for whatever breadth you can feed it.",{"title":441,"searchDepth":442,"depth":442,"links":443},"",2,[444,445,446,447,448,449,450,451,452,453],{"id":14,"depth":442,"text":15},{"id":108,"depth":442,"text":109},{"id":121,"depth":442,"text":122},{"id":204,"depth":442,"text":205},{"id":217,"depth":442,"text":218},{"id":227,"depth":442,"text":228},{"id":275,"depth":442,"text":276},{"id":321,"depth":442,"text":322},{"id":410,"depth":442,"text":411},{"id":423,"depth":442,"text":424},null,"neutral","Meta Andromeda is the AI retrieval engine behind Facebook and Instagram ads. What Meta has claimed since late 2024, and why distinct creative wins delivery.",[],"md",[460,463,466,469,472,475,478,481],{"question":461,"answer":462},"What is Meta Andromeda?","Andromeda is the retrieval stage of Meta's ads delivery system, rebuilt with AI and announced by Meta engineering on December 2, 2024. Retrieval is the step that screens tens of millions of live ad candidates down to a few thousand before Meta's ranking models score them and the auction picks a winner. The rebuild runs on NVIDIA Grace Hopper hardware plus Meta's own MTIA accelerators, which let Meta replace simpler, hand-tuned retrieval logic with far larger neural networks that personalize which ads even get considered for each person.",{"question":464,"answer":465},"What is the difference between creative variation and creative diversification?","Variation is the same idea dressed differently: a square and a vertical crop of one image, or the same hook with the word swapped from buy to shop. Diversification is a genuinely different idea: a price-led ad and a testimonial and a how-it-works demo, each speaking to a different buyer motivation. Retrieval personalizes across whatever distinct entries you give it, so five variations of one concept read as roughly one option, while five diversified concepts give the system five separate ways to match a person. Variation refines a winner; diversification is what actually widens delivery.",{"question":467,"answer":468},"When did Meta Andromeda roll out?","Meta's engineering team announced Andromeda on December 2, 2024, and the post described it as already deployed across Facebook and Instagram. Trade press coverage describes the impact spreading progressively through 2025, with GEM, the companion ranking model, reaching broad impact by Q4 2025. There was no single switch-flip date for individual ad accounts, which is why claims that an Andromeda update hit a specific account on a specific day are stories rather than records.",{"question":470,"answer":471},"Did Andromeda replace audience targeting?","No. Your targeting controls still exist and still constrain delivery. What changed is how much work the system does inside whatever audience you set: retrieval reads signals from the creative itself (format, hook, subject matter) alongside behavioral data to predict who will respond. Practitioners widely report that detailed interest stacks matter less than they used to and that creative carries more of the matching. Meta has not declared targeting obsolete, but the trend toward broad targeting plus diverse creative predates Andromeda and accelerated with it.",{"question":473,"answer":474},"How often should I refresh creative under Andromeda?","Meta publishes no cadence, so treat this as a practitioner rule, not policy. Set a standing schedule (a few genuinely new concepts every two to four weeks) and also refresh on triggers rather than the calendar alone: frequency climbing past roughly 3-4x in a buying window, or cost per result rising more than about 20% from its baseline while click-through falls. Those two signals together signal fatigue. The fix is a new angle, not a recrop, because a smarter retrieval stage finds and exhausts an ad's best-matching audience sooner than the older system did.",{"question":476,"answer":477},"What is the difference between Andromeda, Lattice, and GEM?","They are three layers of the same delivery system, each named and dated by Meta. Andromeda is retrieval (announced December 2024): it screens tens of millions of candidate ads down to a few thousand. Lattice is the unified ranking model Meta credited in March 2025 with raising ad quality by almost 12% and lifting conversions by up to 6%, replacing many smaller models. GEM, the Generative Ads Model announced in November 2025, is the foundation-scale ranking model Meta says drives a 5% conversion increase on Instagram and 3% on Facebook Feed since its Q2 2025 launch. Andromeda decides what can be shown; the ranking layer decides what is shown.",{"question":479,"answer":480},"How many creatives should I run in an Advantage+ campaign?","There is no Meta-published number, and the old rule of 20 or more per ad set is a misconception: when creative supply dramatically outruns budget, learning slows and delivery fragments. A 2026 practitioner range is 10 to 15 conceptually distinct assets per campaign, framed as directional, not Meta policy. The real test is distinctness, not count: distinct angles (UGC, studio, text-only, testimonial, offer) give retrieval separate ways to match a person, while near-duplicate variations read as one. Match the volume to your budget so each ad set still clears the roughly 50-event learning phase.",{"question":482,"answer":483},"Is Andromeda the reason my performance dropped?","Possibly, but it is nearly impossible to isolate. Meta ships delivery changes continuously and does not flag which accounts got what, when. Before attributing a drop to the algorithm, rule out the usual causes: creative fatigue (rising frequency, falling click-through), a budget too small to exit the learning phase, seasonality, tracking breakage, or a landing page change. Treat confident claims that Andromeda cut a specific account's ROAS by a specific percentage on a specific date as unverified.","\u002Fimages\u002Fblog\u002Fmeta-andromeda-creative-hero.webp","The term keeps surfacing in performance reviews and marketing newsletters: results moved, someone says Andromeda, and nobody in the room is sure whether that names a real system or a convenient scapegoat. It is real. Andromeda is the rebuilt retrieval stage of Meta's ads delivery engine, announced in December 2024, and it is one reason creative strategy advice changed tone through 2025 and 2026. This explainer separates what Meta has actually stated from what practitioners infer, then gets to the part you can act on: why a delivery system this good at matching rewards breadth of genuinely distinct creative, and what that means for how you produce ads.",{},true,"\u002Fblog\u002Fmeta-andromeda-creative","2026-10-16",{"title":5,"description":456},[492,496,500,503,507,510,513,516,519,522,525],{"label":493,"url":494,"year":495},"Meta Engineering, Meta Andromeda: Supercharging Advantage+ Automation with the Next-Gen Personalized Ads Retrieval Engine","https:\u002F\u002Fengineering.fb.com\u002F2024\u002F12\u002F02\u002Fproduction-engineering\u002Fmeta-andromeda-advantage-automation-next-gen-personalized-ads-retrieval-engine\u002F","2024",{"label":497,"url":498,"year":499},"Search Engine Land, Inside Meta's AI-Driven Advertising System: How Andromeda and GEM Work Together","https:\u002F\u002Fsearchengineland.com\u002Fmeta-ai-driven-advertising-system-andromeda-gem-468020","2026",{"label":501,"url":502,"year":499},"AdExchanger, What Meta's Andromeda Update Actually Changes (and What It Doesn't)","https:\u002F\u002Fwww.adexchanger.com\u002Fdata-driven-thinking\u002Fwhat-metas-andromeda-update-actually-changes-and-what-it-doesnt\u002F",{"label":504,"url":505,"year":506},"AdExchanger, Q4: Meta Minted Money and Improved Its Monetization","https:\u002F\u002Fwww.adexchanger.com\u002Fplatforms\u002Fq4-meta-minted-money-and-improved-its-monetization\u002F","2025",{"label":508,"url":509,"year":495},"Marketing Dive, Meta's Generative AI Tools Find Purchase With Users and Advertisers","https:\u002F\u002Fwww.marketingdive.com\u002Fnews\u002Fmeta-platforms-q3-2024-earnings-generative-ai-in-marketing\u002F731585\u002F",{"label":511,"url":512,"year":495},"Westwood One, Marketers Vastly Understate the Sales Effect of Creative (NCSolutions\u002FNielsen)","https:\u002F\u002Fwww.westwoodone.com\u002Fblog\u002F2024\u002F04\u002F22\u002Fmarketers-vastly-understate-the-sales-effect-of-creative-and-significantly-overestimate-the-impact-of-targeting\u002F",{"label":514,"url":515,"year":506},"Meta Engineering, Meta's Generative Ads Model (GEM): The Central Brain Accelerating Ads Recommendation AI Innovation","https:\u002F\u002Fengineering.fb.com\u002F2025\u002F11\u002F10\u002Fml-applications\u002Fmetas-generative-ads-model-gem-the-central-brain-accelerating-ads-recommendation-ai-innovation\u002F",{"label":517,"url":518,"year":506},"Meta for Business, AI Innovation in Meta's Ads Ranking Driving Advertiser Performance","https:\u002F\u002Fwww.facebook.com\u002Fbusiness\u002Fnews\u002Fai-innovation-in-metas-ads-ranking-driving-advertiser-performance",{"label":520,"url":521,"year":506},"Marketing Dive, Meta's AI Bets Supercharge Marketing Efficiency and Costs","https:\u002F\u002Fwww.marketingdive.com\u002Fnews\u002Fmeta-ai-bets-supercharge-marketing-efficiency-costs\u002F804238\u002F",{"label":523,"url":524,"year":506},"Meta Q3 2025 Earnings Call Transcript (via The Motley Fool)","https:\u002F\u002Fwww.fool.com\u002Fearnings\u002Fcall-transcripts\u002F2025\u002F10\u002F29\u002Fmeta-platforms-meta-q3-2025-earnings-call-transcript\u002F",{"label":526,"url":527,"year":499},"Modern Marketing Institute, How to Exit the Meta Ads Learning Phase","https:\u002F\u002Fwww.modernmarketinginstitute.com\u002Fblog\u002Fhow-to-exit-the-meta-ads-learning-phase-fast-and-start-scaling-profitably-in-2026",[529,533,536,539,543,547,550,554,558,562,566],{"label":530,"value":531,"source":532},"Live ad candidates Andromeda screens down to a few thousand","Tens of millions","Meta Engineering, 2024",{"label":534,"value":535,"source":532},"Increase in ads retrieval model capacity with Andromeda","10,000x",{"label":537,"value":538,"source":532},"Ads quality improvement on selected segments at deployment","+8%",{"label":540,"value":541,"source":542},"Advertisers using at least one Meta generative AI ad tool by Q4 2024","4 million+","AdExchanger, 2025",{"label":544,"value":545,"source":546},"Ads made in a single month with Meta GenAI tools by 1M+ advertisers","15 million+","Marketing Dive, 2024",{"label":548,"value":549,"source":546},"Conversion lift Meta reported for ads using its Image Generation","+7%",{"label":551,"value":552,"source":553},"GEM conversion lift since Q2 2025 launch, Instagram \u002F Facebook Feed","+5% \u002F +3%","Meta Engineering, 2025",{"label":555,"value":556,"source":557},"Ad-quality gain Meta credits to its unified Lattice ranking model","+12%","Meta for Business, 2025",{"label":559,"value":560,"source":561},"Advantage+ shopping annual run-rate, Q4 2024 (+70% YoY)","$20B+","Fifth Person (Meta Q4 2024 call), 2025",{"label":563,"value":564,"source":565},"Meta end-to-end AI ad tools annual run-rate, Q3 2025","$60B+","Marketing Dive, 2025",{"label":567,"value":568,"source":569},"Share of a brand's advertising sales lift attributable to creative","~49%","NCSolutions\u002FNielsen via Westwood One, 2024","blog\u002Fmeta-andromeda-creative","eVfOY-JdellPFVh1P6csprZxgKlRnzC5N64gslMv-mc",1786093698209]