Meta Andromeda: Why Creative Is Now Your Targeting (2026)

Last updated: July 1, 2026.

If your Facebook and Instagram ads behaved differently over the past year — audiences you carefully built stopped mattering, budget alone stopped fixing things, and one creative suddenly carried a whole campaign — you were not imagining it. Meta rebuilt the first stage of how ads are chosen, and the change quietly moved the single most important lever from who you target to what you make.

The short version: Meta’s ad system now runs on an AI retrieval engine called Andromeda. Before the auction ever happens, Andromeda reads the content of your creative — the image, the video, the hook, the message — and uses that to decide which few thousand of tens of millions of ads are even eligible to reach a given person. In 2026, your creative is your targeting. This guide explains what Andromeda actually is (with Meta’s own numbers), reconciles Meta’s optimistic claims with what advertisers actually measured, and gives you a concrete creative system to win under it.

What Meta Andromeda actually is

Most explainers skip the one fact that makes everything else click: Meta’s ad delivery is a two-stage pipeline, and Andromeda is stage one. According to Meta’s own engineering team (announced December 2, 2024): “Retrieval is the first step in our multi-stage ads recommendation system. This stage is tasked with selecting ads from tens of millions of ad candidates into a few thousand relevant ad candidates.”

In plain English: every time someone opens Facebook or Instagram, Meta cannot score all eligible ads against them — there are far too many. So a fast first pass (retrieval) picks a shortlist. Only the ads that survive retrieval move on to ranking and the auction. If your creative does not make the shortlist, your bid and budget are irrelevant — your ad was never in the running. That is the shift in one sentence.

The second stage in that diagram, GEM (Meta’s generative ads model), then ranks the shortlisted candidates before the auction — more on it later. The engineering leap in the retrieval stage is real: Meta built Andromeda as “a highly customized deep neural network with sublinear inference cost,” yielding a stated ~10,000x increase in model capacity for personalization, running on NVIDIA Grace Hopper Superchips and Meta’s own MTIA silicon. Meta reports the system produced a +6% recall improvement in retrieval and a +8% ad-quality improvement on tested segments. Those are the primary numbers — quoted directly, not paraphrased through a third party.

What Andromeda “sees” when it reads your creative

Here is the mechanism most guides leave vague. Andromeda does not just match keywords or interests. It uses machine learning to interpret the content of your ad — what is visually in the image or video, the structure of the hook, the language and claims, the format. It builds a rich representation of your creative and matches that representation to the people most likely to respond. In effect, your ad is the query, and the audience is the result.

Practically, that means the creative signals Meta can read now carry the targeting weight that manual audience settings used to:

• Visual subject and scene — what the product is, who is in frame, the setting, on-screen text.
• Message and angle — the problem you name, the promise you make, the proof you show.
• Format and pacing — UGC selfie-video vs polished studio spot vs static; the first three seconds.
• Language and market cues — which help Meta place the ad with the right people.

This is why two ads with identical targeting can reach completely different audiences: Andromeda routes each one based on what it contains. Your job shifts from describing the audience to Meta, to showing Meta — through varied creative — who you are for.

Why creative is now your targeting

Industry practitioners converged on the same conclusion once Andromeda rolled out. As one 2026 paid-social playbook puts it, “Creatives are now the largest performance lever… audience controls matter less. Creative quality, variation, and clarity matter more.” Another Andromeda field guide is blunter: “Your ad creative now does most of the targeting work. Audience settings still exist, but they play a much smaller role.”

The strategic consequence: broad targeting plus a diverse creative library now beats narrow, hand-built audiences in most accounts. You are handing Meta the raw material (many distinct creatives) and letting Andromeda do the matching it is now far better at than manual segmentation. There are exceptions — tightly local businesses, age- or geo-regulated categories, and retargeting warm audiences still benefit from explicit controls — but for prospecting, creative variety is the lever.

What Meta claims vs what advertisers actually saw

This is the part most articles will not tell you, and it is the most useful. Meta’s headline stat — widely repeated — is a +22% increase in ROAS. Read Meta’s wording carefully: it applies to “advertisers who did not previously use Advantage+ creative” when they “turned on its AI-driven targeting features.” It is a switch-on-the-feature lift for a specific cohort, not a promise that every account improved 22% because Andromeda exists.

What happened at the account level was messier. An independent analysis of roughly $834 million in ad spend across 3,014 advertisers by Confect found that, in aggregate, ROAS and conversion rates declined in the months around the rollout, and the accounts that had been top performers under the old system were hit hardest — a “great equalizer” effect as Meta’s automation compressed the gap between good and average accounts.

What actually changed for advertisers

More ads per ad set — and they must be different

The old best practice of 3–6 ads per ad set is gone. As one practical Andromeda guide notes, ad sets “typically contained 3–6 ads… Meta recently raised the amount of ads you can have in an ad set to 150.” High-performing accounts now run on the order of 15–20+ conceptually distinct creatives per ad set. The key word is distinct: 20 color variants of one concept is not diversity. Andromeda wants genuinely different angles, formats, and messages so it has real options to route.

Simpler account structure

Because the algorithm does the matching, fragmenting budget across many narrow ad sets starves each one of the conversions it needs to learn (Meta’s learning phase still wants roughly 50 optimization events in 7 days per ad set). Most accounts now consolidate into fewer, broader ad sets — often anchored by Advantage+ Shopping — and let creative do the segmenting.

Creative is the default format that lasts

UGC-style video has become the default for durable performance. As one 2026 creative-trends analysis observes, “UGC-style video… has become the default creative format for high-performing Facebook advertisers” — it consistently shows the longest active lifespans in ad databases. That does not mean polished creative is dead; it means authentic, native-feeling video is the reliable backbone.

The Concept Diversity Matrix: a framework for Andromeda-era creative

“Make more creatives” is useless advice without a system. Here is the one we use at Servo, and you can steal it. Instead of counting ads, map concepts across two axes — angle (the reason to buy) and format (how it is delivered). Fill the grid, and you naturally produce 15–20 genuinely distinct ads that give Andromeda real range.

Each “Yes” is a distinct concept to produce. A filled matrix like this yields ~14–16 conceptually different ads from five angles and four formats — exactly the diversity Andromeda rewards, without producing near-duplicates. Rotate the weakest concepts out on your refresh cadence and add new angles as you learn which ones the algorithm favors. This is the mental model behind Servo’s Creative DNA and Winner Patterns, which identify which visual, copy, and structural attributes drove your past wins so your next matrix is informed, not guessed.

Diagnosing creative fatigue with Meta’s new metrics

Under Andromeda, when performance drops it is almost always a creative problem, not an audience problem — and Meta added Ads Manager diagnostics to prove it. Learn to read three:

• Creative Fatigue — flags when an ad’s audience has seen it too often and response is decaying. Rising frequency alongside rising CPM and falling CTR is the classic signature.
• Creative Similarity — warns when your ads are too alike, which is exactly what starves Andromeda of routing options. High similarity is a signal your “20 ads” are really one concept in costume.
• Top Creative Themes — shows which themes are actually driving results, so your next matrix leans into what works.

Signal quality still matters: Pixel, CAPI, and EMQ

Creative decides who is eligible, but Meta still needs clean conversion signal to learn what “good” looks like — especially after iOS-era tracking loss. Run the Meta Pixel and the Conversions API together with proper deduplication, and push your Event Match Quality as high as you can — a strong target is commonly cited as roughly 7+ out of 10. Weak signal starves even a great creative library of the feedback Andromeda uses to route it. Think of it as: creative is the input, conversion signal is the grade, and Andromeda learns from both.

Where this is heading: GEM and fully AI-generated ads

Andromeda handles retrieval; a companion system, GEM (Meta’s generative ads model), is reshaping ranking, and analysts argue GEM is an even bigger shift than Andromeda. The trajectory is explicit: Meta aims to let brands fully create and target ads with AI by the end of 2026 — upload a product and a goal, connect a bank account, and let Meta generate the creative, pick the audience, and run the campaign.

That future makes the present advice more important, not less. As the machine takes over targeting and delivery, the human edge concentrates in creative strategy and judgment — which angles are true to your brand, which proof is credible, which concepts to feed the system. The advertisers who win the AI era are the ones who get very good at directing creative, fast.

How Servo helps you win the creative game

Servo is built for exactly this shift. Its AI Creative Audit scores ads across multiple quality pillars — hook clarity, design, brand alignment, Meta compliance, and more — before you spend, so weak concepts do not waste retrieval. Creative DNA and Winner Patterns tell you which attributes drove past wins, so your Concept Diversity Matrix is data-informed rather than guessed. And Autopilot surfaces the fatigue signals and suggests which concept to rotate in next — every recommendation shows its impact estimate and evidence, and you approve it. If you are weighing tools, our honest comparisons, the 2026 automation-tools guide, and Servo pricing lay out the options fairly. And because ad costs are the other half of this equation, see our companion piece on what Facebook ads actually cost in 2026.

Frequently Asked Questions

What is Meta Andromeda?

Andromeda is Meta’s AI-powered ad retrieval engine — the first stage of ad delivery. Announced by Meta’s engineering team in December 2024, it filters tens of millions of eligible ads down to a few thousand relevant candidates for each person before ranking and the auction, using a deep neural network with roughly 10,000x more model capacity than the prior system.

Why did my Facebook ads change in 2026?

Because Andromeda moved the decisive signal from audience settings to creative content. The system now reads your ad to decide who is eligible to see it, so hand-built audiences carry less weight and creative variety carries more. Accounts that relied on narrow targeting often saw results shift; accounts with strong, diverse creative tended to hold up better.

Is detailed targeting dead on Facebook in 2026?

Not dead, but demoted for prospecting. Broad targeting plus a diverse creative library now outperforms narrow manual audiences in most prospecting campaigns, because Andromeda does the matching from your creative. Detailed controls still help for local businesses, regulated categories, and retargeting warm audiences.

How many ads should I run per ad set under Andromeda?

More than the old 3–6, and they must be genuinely different. High-performing accounts commonly run 15–20+ conceptually distinct creatives per ad set (Meta raised the ceiling to 150). Focus on distinct angles and formats — the Concept Diversity Matrix above — not near-duplicate variants, which Andromeda cannot meaningfully differentiate.

Did Andromeda actually improve performance?

It depends on who you were. Meta reports a 22% ROAS lift for advertisers switching on Advantage+ creative. But independent analysis of $834M in spend across 3,014 advertisers found aggregate ROAS and conversion rates dipped after rollout, with former top performers hit hardest. The system rewards creative volume and diversity and erodes old manual-targeting advantages.

Why are my Facebook CPMs rising in 2026?

Two reasons stack. First, auction competition and CPMs have risen industry-wide. Second, under Andromeda, rising CPM on a specific ad is often a creative-fatigue signal — the audience has seen it too much. Check Meta’s Creative Fatigue metric; if it flags, rotate in a fresh concept rather than just raising budget. See our guide to Facebook ad costs in 2026 for benchmarks.

Does UGC outperform polished creative under Andromeda?

On average, yes for durability — UGC-style video shows the longest active lifespans and has become the default backbone for high-performing accounts. But the real winner is diversity: run UGC alongside statics, carousels, and founder-style video so Andromeda has range. Polished creative still earns its place, especially for brand and offer angles.

Do I need to disclose AI-generated content in Meta ads?

Meta has moved toward labeling and, in some cases, requiring disclosure of AI-generated or significantly AI-altered content. As AI creative tools become standard, set your content-disclosure fields correctly per Meta’s current policy at publish time. Tools like Servo set the AI-generated disclosure automatically where required.

What is GEM, and how is it different from Andromeda?

Andromeda is retrieval (choosing the candidate shortlist); GEM is Meta’s generative ads model reshaping how those candidates are ranked and, increasingly, created. Together they point toward Meta’s stated 2026 goal of fully AI-generated and AI-targeted ads. Andromeda changed who sees your creative; GEM is changing how ads are built and scored.

---

Sources and methodology: Primary technical figures are quoted directly from Meta’s Engineering blog on Andromeda (Dec 2, 2024). Meta’s 2026 fully-AI-ads goal was first reported by the Wall Street Journal and covered by eWeek. Real-account performance data is from Confect’s analysis of $834M in ad spend. Practitioner guidance and the GEM framing are drawn from Search Engine Land, Social Media Examiner, Logical Position, Admove, Foxwell Digital, and Adligator. The Concept Diversity Matrix is an original Servo framework.

About the author: Pauls Rubenis is the co-founder of Servo, an AI-powered Meta Ads workspace. He holds an MSc in Digital Strategy and AI Management and Meta Blueprint Media Buying Professional certification, and has managed performance campaigns for e-commerce, SaaS, and lead-gen brands across the Nordics, Benelux, DACH, and Baltic regions. Follow Servo on LinkedIn.

Disclosure: This guide references Servo, a product built by the author. We have aimed to describe Meta’s Andromeda system accurately, cite primary sources, and present both Meta’s claims and independent real-account data. Technical details are accurate as of July 2026; verify current Meta policy and metrics in Ads Manager.