What Is Andromeda? Meta's AI Ranking Engine Explained for Advertisers

TL;DR: Andromeda is Meta's AI-native ad ranking and retrieval system that evaluates semantic relevance — how well your ad content matches a user's interest graph — not just bid price. It means creative content is now a first-class ranking signal. Advertisers who understand what Andromeda rewards will outperform those still optimizing purely for bid and historical CTR.

What Andromeda Actually Is

When Meta's engineers say "Andromeda," they're referring to a retrieval and ranking architecture that replaced parts of the old ad auction stack. It was announced publicly in 2023 as part of Meta's AI infrastructure push and has been scaled progressively across Facebook and Instagram ad delivery.

The core mechanism: Andromeda uses large language model (LLM) embeddings to encode both ads and users into a shared semantic space. Instead of matching ads to users purely by bid price and predicted click-through rate, the system asks a more nuanced question: does the content of this ad align with what this user actually cares about?

That shift has real consequences for how ads perform — and for how you should be thinking about creative strategy.

Before Andromeda, Meta's ad ranking was dominated by a formula combining three factors: your bid, your estimated action rate (predicted CTR or conversion probability), and your ad quality score. Get the bid right and maintain a decent quality score, and your ad moved through the auction competitively.

Andromeda adds a retrieval stage before the auction. The system narrows a massive candidate pool of potentially eligible ads down to a manageable set using semantic similarity. Only ads that pass this semantic filter enter the full ranking auction. Which means if your ad is semantically mismatched to a user's interest graph — regardless of bid — it may not reach the auction at all.

For advertisers, this is a meaningful change in the rules of competition.

The Technical Architecture: Retrieval Then Ranking

Meta has published partial technical details about Andromeda in their AI research blog. The architecture follows a two-stage pattern common in large-scale recommendation systems: retrieval then ranking.

Stage 1 — Retrieval: Given a user in a specific context (feed, Reels, Stories), the system needs to quickly identify which ads from a pool of hundreds of millions are relevant candidates. Andromeda uses approximate nearest neighbor (ANN) search over dense vector embeddings to do this. Both ads and users are encoded as vectors in the same high-dimensional space; ads closest in vector distance to the user's interest representation become candidates.

Stage 2 — Ranking: The candidate set (typically thousands of ads) is then ranked using a more computationally expensive model that incorporates bid, predicted conversion rate, user signals, and contextual factors. This produces the final ranked list from which the winning ad is selected.

The retrieval stage is where Andromeda changed the game. Prior systems used keyword and category matching for retrieval — relatively crude filters. Andromeda's embedding-based retrieval understands semantic content. An ad for a running shoe can be retrieved as relevant for someone whose interest graph contains marathon training, trail running, and athletic gear — even if the ad copy never uses those exact words.

This matters for advertisers because it means your ad creative is being read and understood at a semantic level, not just matched against keyword lists.

Why This Changed the Advertiser Equation

Under the old system, the advertiser's primary lever was targeting: define an audience tightly enough that the estimated action rate was high, bid competitively, and keep quality score above threshold. Creative mattered for conversion, but the ranking mechanism itself was relatively agnostic to creative content.

Under Andromeda, creative content becomes a ranking input. Here's the concrete mechanism:

If your ad's semantic embedding is close to the user's interest embedding, it gets retrieved as a candidate. The closer the match, the higher the probability of retrieval — and retrieval is binary. An ad that isn't retrieved doesn't get ranked or shown, regardless of bid.

So even with a high bid and strong historical performance, an ad whose creative is semantically disconnected from the target audience's interest profile faces a retrieval disadvantage. The system is, in effect, penalizing creative that isn't substantively relevant to the person seeing it.

For media buying teams, this translates to a practical implication: creative quality is now upstream of bid efficiency. You can't bid your way past a semantic relevance gap.

What "Semantic Relevance" Means for Creative

Semantic relevance, in the context of Andromeda, refers to how well the content of your ad — its copy, offer framing, imagery context, and overall message — aligns with the inferred interests of the target user.

This isn't just about keywords. The embedding models understand context and relationships. An ad for a protein supplement that leads with "recovery after hard workouts" is semantically relevant to users whose interest graph includes fitness, training, and physical performance — even if those exact terms don't appear in your ad.

What this means in practice:

Be specific about the audience you're speaking to. Vague, generic ad copy reads as semantically low-signal. "Shop our new collection" is nearly content-free from an embedding perspective. "The trail shoe built for 50k elevation gains" encodes specific semantic meaning that maps to specific interest clusters.

The offer and the creative must match. An ad whose creative signals outdoor adventure but whose offer is a corporate software trial creates a semantic mismatch. Andromeda's retrieval will struggle to place it correctly regardless of how you've configured targeting.

Format signals matter too. Video ads, image composition, and even visual style encode semantic information. A UGC-style video sends different semantic signals than a polished brand spot — and different audiences respond better to each. Knowing which format resonates with which interest cluster is now part of the ranking optimization problem.

For teams thinking through ad performance problems, this reframes what counts as a creative failure. An ad that generates few impressions isn't just "not converting" — it may be failing at the retrieval stage. That's a different diagnostic than a conversion rate problem.

The Advantage+ Connection

Meta's Advantage+ products — Advantage+ Shopping Campaigns (ASC+), Advantage+ Audience, and Advantage+ Creative — are the advertiser-facing surface of the Andromeda infrastructure.

When you run an Advantage+ campaign, you're explicitly delegating retrieval and ranking decisions to the system. Meta's AI selects audiences, places your ad across placements, and may alter creative elements (background color, aspect ratio, text overlay) to optimize retrieval-level relevance.

This is why Advantage+ campaigns often outperform manually configured campaigns — not because Meta's AI is magic, but because it has direct access to the semantic matching infrastructure. It can test creative variations against different interest clusters and learn which semantic matches produce conversions, then reinforce those pairings.

For advertisers using Advantage+, the implication is: give the system creative diversity. If you supply five ad variants that are semantically identical (same offer, same angle, same framing), Advantage+ has nothing to test across different interest clusters. Supply creative that signals different values — utility, aspiration, community, proof — and the system can find which semantic framing retrieves best for each user segment.

This is where studying what competitor ad research looks like becomes operationally important. You want to understand what semantic angles are currently resonating in your category, not just what ads look like.

How Andromeda Affects Audience Targeting

One of the more counterintuitive implications of Andromeda is that it changes the relative value of narrow audience targeting.

Before semantic retrieval, narrow targeting was a primary efficiency mechanism: define a tight audience with high expected action rates, and the bid-weighted ranking equation favored you. The smaller the audience, the higher the predicted relevance, the better the auction math.

Andromeda shifts this. If the semantic retrieval stage is doing meaningful audience-content matching, then broad broad targeting becomes viable in a way it wasn't before. The system can identify users within a broad audience who are semantically matched to your ad — doing the work that narrow targeting used to require.

This is consistent with what practitioners have observed since Advantage+ Audience launched: broad audience settings often outperform tightly constrained custom audiences, especially for accounts with sufficient conversion history for the system to learn from.

The caveat: this benefit scales with creative quality and diversity. A broad audience setting with three semantically identical ad variants gives Andromeda little to work with. A broad audience setting with six semantically distinct creative approaches gives the system vectors to match against many different interest clusters within that broad pool.

Under the old system, you did audience segmentation at targeting configuration time. Under Andromeda, semantic segmentation happens at retrieval time — driven by creative content, not audience definition.

What Andromeda Means for the Learning Phase

The learning phase — Meta's period of campaign-level data collection before delivery stabilizes — takes on a different character under Andromeda.

Previously, learning phase was primarily about the system building a conversion model: learning which users in your defined audience were most likely to take the desired action. The process was bid-and-conversion-rate driven.

Andromeda adds a retrieval learning component. The system needs to map your creative content into the embedding space and calibrate how well it retrieves against different interest clusters. This calibration takes signal — which means sufficient impressions, sufficient conversions, and sufficient creative variety to produce meaningful gradient updates in the retrieval model.

The practical implication for campaign setup:

Launching with only one or two creative variants lengthens effective learning because the system has fewer semantic vectors to test. Launching with three to five semantically distinct variants — different hooks, different offer framings, different emotional angles — gives the retrieval system more surface area to calibrate quickly.

Also: changing creative frequently during learning phase is more disruptive under Andromeda than it was before. You're resetting retrieval calibration, not just conversion modeling. Hold off on creative swaps until the learning phase completes unless performance is clearly failing.

For teams tracking creative test results, the ad timeline analysis view in AdLibrary shows how long competitor ads have been running — which is a useful proxy for how long it takes ads in your category to exit learning and stabilize in delivery. If category leaders are running ads for 60+ days without swapping creative, that's a signal that their campaigns have cleared learning and are in stable delivery — a benchmark for your own timeline expectations.

The Semantic Gap Problem: What Most Advertisers Are Getting Wrong

Here's the pattern we see repeatedly in accounts that underperform under Andromeda: advertisers are optimizing for the conversion funnel without optimizing for retrieval.

They build creative that converts well in testing but produces low impression volume. They blame bid competition or audience saturation. The actual problem is semantic mismatch at the retrieval stage.

The concrete failure mode: a brand builds ads that are visually polished, have strong CTAs, and test well with a warm audience. Those ads then underperform at scale with cold audiences. The warm audience has behavioral history with the brand, which boosts estimated action rate enough to compensate for retrieval-level mismatch. The cold audience has no such history — the ad has to earn its way into the auction via semantic relevance, and it can't.

The fix requires revisiting creative from the top: what semantic signals is this ad emitting? What user interest clusters does it match? Is the match tight and specific, or vague and generic?

A useful diagnostic: run your best-performing ad through a structured creative analysis. What is the core topic? Who is the implied audience? What problem or desire does it reference? Then ask whether those elements are prominent enough in the creative to encode clearly in an embedding.

AdLibrary's AI ad enrichment tool does exactly this — surfaces the hook structure, audience framing, and emotional angle of any ad in its database. Running competitor ads through this analysis reveals which semantic approaches they're using, and comparing against your own creative shows where the gaps are.

Reading Competitor Creative as an Andromeda Signal

If Andromeda's retrieval system rewards semantic relevance, then the ads your competitors are running longest are the ones that have proven semantic match with your shared audience pool. That's a signal worth analyzing systematically.

The reasoning: a competitor's ad that runs for 90+ days in the same geographic market you're targeting has survived the retrieval stage repeatedly. Meta's auction keeps selecting it because it's semantically well-matched to users who convert. That's not just a creative success signal — it's an indication of which semantic angle resonates with your shared target audience.

The research workflow:

1. Search for your category's top 3-5 advertisers using AdLibrary's unified ad search.
2. Filter by date range (ads running 60+ days) using ad timeline analysis to surface long-running ads.
3. Filter by media type filters to isolate the formats — video vs. static — that are sustaining longest.
4. Use AI ad enrichment to surface the semantic framing of each long-running ad: what is the hook, what is the angle, what user desire or problem is addressed?
5. Map that landscape against your own creative's semantic framing. The gaps are your highest-probability creative hypotheses.

This isn't copying competitors. It's reading market signals about what the retrieval system has validated. You then build original creative that occupies the same semantic territory — or a differentiated adjacent position.

Save reference ads to your saved ads library as you research. Organized by semantic angle rather than advertiser, these collections become a semantic landscape map for your category.

How Andromeda Interacts with the Ad Auction

Andromeda doesn't replace the ad auction. It precedes it. Understanding how they interact clarifies the full picture.

After Andromeda's retrieval stage narrows the candidate set, the traditional auction mechanics still apply: total value = bid × estimated action rate × ad quality score. The winning ad is still the one with the highest total value.

But the retrieval stage has changed the effective playing field in two ways:

First, it narrows competition. Ads that are semantically mismatched don't enter the auction for semantically specific users. This means the auction you're competing in is smaller and more relevant — which should, in theory, produce better outcomes for well-matched ads at similar bid levels.

Second, it creates a semantic quality multiplier. The degree of semantic match influences estimated action rate (because a well-matched ad is more likely to produce the desired action), which in turn improves total value calculation. Strong semantic match at retrieval → better estimated action rate in ranking → more competitive total value at the same bid.

This is why advertiser accounts with strong semantic alignment between creative and audience tend to see lower effective CPMs over time. They're winning auctions at lower bid levels because the semantic quality multiplier is doing part of the work.

The CPM calculator can help you model what shifts in effective CPM mean for your total budget efficiency. Even a 15% improvement in effective CPM at the same CPA translates directly to more reach per euro spent.

Preparing Your Creative Strategy for Andromeda

Concrete steps for advertisers who want to optimize for Andromeda's retrieval stage:

Audit current creative for semantic specificity. Review your live ads and ask: what is the specific topic this ad is about? If the answer is vague — "our brand," "a great product," "a special offer" — the ad is semantically low-signal. Rewrite toward a specific audience problem or desire.

Diversify semantic angles, not just visual executions. Don't run five ads with the same offer framing in different colors. Run five ads with five different semantic angles: one focused on the result (outcome), one on the problem (pain), one on identity (who this is for), one on social proof (what others have experienced), one on utility (how it works). Each angle maps to a different part of the interest graph.

Align landing page content with ad semantic content. Andromeda doesn't just read the ad — Meta's systems evaluate landing page relevance as a component of quality score. An ad about trail running that lands on a generic homepage is semantically incomplete. Match the semantic angle of the ad to the landing page section it opens.

Use Advantage+ Creative thoughtfully. Advantage+ Creative modifications (background removal, text overlays, image enhancement) are designed to improve retrieval-level relevance for individual users. But they work best when the underlying creative is semantically strong. A weak brief produces weak modifications regardless of the AI enhancement layer.

Monitor delivery curve, not just conversion rate. A campaign that performs well in week 1 but plateaus in week 3 may be hitting audience saturation within a semantic cluster. That's a creative refresh signal — not a bid signal. Refreshing with a new semantic angle re-enters you into adjacent clusters rather than competing for the same already-saturated match.

For a broader view of how algorithmic systems like Andromeda fit into the 2026 paid media landscape, see algorithmic convergence across Meta, Google, and TikTok — which covers how all three platforms are moving toward embedding-based retrieval and what that means for cross-platform creative strategy.

The AdLibrary API and Andromeda Research at Scale

For teams running systematic competitor ad research — tracking dozens of advertisers across categories to map the semantic landscape — manual research in AdLibrary's interface works for regular sessions but doesn't scale to continuous monitoring.

Meta's own Ad Library API provides basic programmatic access to public ad data, which is adequate for single-platform research with minimal field requirements.

When systematic research requires richer creative metadata, multi-platform coverage (TikTok, YouTube, LinkedIn in the same query), and automated enrichment pipelines, Meta's free API stops being enough. AdLibrary's API access — available on the Business plan at €329/mo — provides programmatic access to the same ad intelligence database, with richer fields and no business verification friction.

Meta's free API is fine for one platform and basic use. The moment you add TikTok, YouTube, or LinkedIn data into the same research query — mapping semantic angles across platforms to understand which are platform-specific and which are cross-platform signals — you need something else.

For teams building Andromeda-aware creative research workflows in code, Claude Code + adlibrary API covers how to combine programmatic ad retrieval with LLM analysis to systematically map competitive semantic landscapes.

Frequently Asked Questions

What is Meta Andromeda?

Andromeda is Meta's AI-native ad ranking and retrieval system, announced in 2023 and scaled through 2024-2026. It uses semantic embedding models to match ads to users based on interest relevance — not just bid price and historical CTR. The system evaluates ad creative content, landing page context, and user interest graphs simultaneously to determine which ad gets shown and at what effective cost.

How does Andromeda differ from Meta's previous ad ranking system?

Meta's previous ranking relied primarily on a combination of bid price, estimated action rate (predicted CTR and conversion rate), and ad quality score. Andromeda adds a semantic retrieval layer: it uses large embedding models to assess how well an ad's content, copy, and creative align with a user's inferred interest graph. This means relevance — not just bid — determines whether your ad gets into the candidate pool in the first place.

Does Andromeda affect how much advertisers pay per click?

Yes, indirectly. Andromeda's semantic relevance scoring affects whether your ad enters the auction at all, and with what quality multiplier. Ads that are semantically well-matched to audience interests receive better quality scores, which improves effective CPM even at the same nominal bid. Ads that are semantically mismatched — even with high bids — face higher effective costs or reduced delivery.

What does Andromeda mean for creative strategy?

It means creative content is now a ranking signal, not just a conversion signal. An ad that clearly communicates its topic, audience fit, and offer in a way that aligns with user interest patterns will receive a semantic relevance boost. This makes creative differentiation — and studying what creative approaches competitors are actually running — more strategically important than it was under pure CTR-based ranking.

How can advertisers research what creative approaches are working under Andromeda?

The most direct method is competitive creative research: studying which ads from competitors in your category are running longest and at highest volume — which serves as a proxy for semantic relevance and auction performance. Tools like AdLibrary's ad timeline analysis and AI ad enrichment let you systematically identify which creative formats, hooks, and messaging angles are sustaining delivery in your niche, giving you a data-informed starting point for your own Andromeda-optimized creative.

The Bottom Line

Andromeda isn't a black box that advertisers can't influence. It's a system with clear inputs — semantic content of your creative, alignment with user interest graphs, landing page relevance — that respond to deliberate creative decisions.

The advertisers who will benefit most from Andromeda are those who treat creative content as a ranking variable, not just a conversion variable. That means being specific, being semantically coherent, and building creative diversity that gives the retrieval system real variation to test.

The advertisers who will struggle are those still optimizing as if bid and CTR are the only levers — ignoring retrieval-stage match in favor of conversion-funnel metrics. That approach worked in 2020. It's increasingly insufficient in 2026.

For the creative intelligence layer — knowing which semantic angles competitors are validating in your market, which formats are sustaining delivery longest, and which hooks are earning the ad runs that signal retrieval success — AdLibrary's Pro plan at €179/mo gives you 300 credits/month for systematic competitor research without rationing. That's the research budget that makes Andromeda-aware creative strategy executable, not just theoretical.

For teams building programmatic competitive monitoring — tracking semantic landscape shifts at scale, across platforms — the Business plan at €329/mo adds API access for pulling that intelligence into your own workflows. See how to use that in competitor ad research and automate competitor ad monitoring.

Common Misunderstandings About Andromeda

Several misconceptions have spread in the practitioner community since Andromeda's announcement. Clearing them up prevents strategy mistakes.

Misconception 1: Andromeda means targeting doesn't matter.

Targeting still matters, but its role has shifted. Andromeda doesn't eliminate audience targeting — it changes what targeting is for. Targeting now constrains the retrieval pool (which users the system considers) rather than doing the semantic matching work. Setting broad targeting with a semantically strong creative is viable because Andromeda's retrieval handles the matching within that broad pool. But very narrow targeting that conflicts with your creative's semantic signals creates constraints without adding relevance.

The practical rule: let your creative define the semantic target; use targeting to constrain the geographic, demographic, or behavioral universe within which that semantic matching happens.

Misconception 2: Andromeda rewards verbose, keyword-heavy ad copy.

Embedding models don't work like keyword matching. They understand meaning, not term frequency. Stuffing your ad copy with terms you think are relevant to your target audience doesn't improve retrieval performance — and may actively degrade it by making the copy unnatural and conversion-unfriendly.

Andromeda rewards semantic clarity, not keyword density. One tightly written ad that clearly communicates a specific problem and a specific offer outperforms a keyword-stuffed ad in retrieval and in conversion. See ad relevance diagnostics for how Meta surfaces signals about creative quality.

Misconception 3: Andromeda penalizes broad creative concepts.

Broad isn't the same as vague. A brand awareness ad can have broad creative — it doesn't have to name a specific feature or narrow use case. But it still needs to clearly communicate: who this is for and what emotional territory it occupies. "Broad" and "semantically clear" are not contradictions. "Vague" and "generic" are the actual retrieval penalties.

Misconception 4: You can't diagnose Andromeda-related delivery problems.

You can't see the embedding distance between your ad and a user's interest graph directly. But you can see symptoms. If your campaign's impression volume is low relative to audience size, and your bid is competitive, and your landing page isn't obviously mismatched — suspect retrieval. Test a new creative variant with a sharper semantic angle and watch whether impression volume changes within the first 48-72 hours. That's a retrieval-stage signal.

For structured creative testing at the retrieval stage, use separate ad sets per creative variant (don't consolidate into one ad set for Andromeda retrieval testing). You want clean signal on which creative is getting retrieved, not blended performance from multiple variants competing for the same ad slot.

How Andromeda Changes the Competitive Landscape

The broader strategic implication of Andromeda is a shift in what constitutes a sustainable competitive advantage in Meta advertising.

Before Andromeda, scale advantages mattered most: larger budgets could bid higher, access larger audiences, and outspend smaller competitors. Creative quality mattered for conversion, but well-funded generic ads could still win at scale.

Andromeda introduces a competency-based advantage that partially offsets scale: if your creative is semantically better-matched to your audience than a competitor's, you win auctions even at lower bid levels. The competition isn't just a budget race — it's a relevance race.

For smaller advertisers competing against funded brands, this is an opening. A €5,000/month advertiser with precise semantic understanding of their audience can outbid a €50,000/month advertiser running semantically generic creative — at least for the specific audience segment where the smaller advertiser has semantic depth.

This is why competitor ad research matters more now than it did under pure auction dynamics. Understanding what your competitors' creative signals to Andromeda's retrieval system tells you where the semantic territory is claimed and where the gaps exist. Gaps are adjacencies where your creative can establish semantic relevance without directly competing against incumbents in the auction.

For a data-driven view of how competitive dynamics play out across Meta ad categories, campaign benchmarking provides context on what typical CPMs, CTRs, and creative cadences look like in specific verticals — essential context for interpreting your own Andromeda-era performance numbers.

Andromeda in Context: The Broader Meta AI Stack

Andromeda doesn't operate in isolation. It's part of a broader Meta AI infrastructure that includes:

• Advantage+ Creative: AI-driven creative adaptation at serving time. Modifies image backgrounds, text placement, aspect ratios based on placement and user context.
• Advantage+ Audience: AI-driven audience expansion beyond manual targeting, using the same embedding infrastructure.
• Advantage+ Shopping (ASC+): The fully automated campaign format that delegates budget, audience, and creative optimization to Meta's systems end-to-end.
• Automated Rules and Bid Strategies: Budget pacing, bid caps, and spend automation that interacts with the auction stage after retrieval.

Andromeda is the retrieval layer. Everything else in this stack operates on the candidate pool that Andromeda produces. Which means optimizing for Andromeda is not optional if you want the rest of the stack to work well — it's the entry gate.

For a comprehensive view of how these AI systems interact across Meta, Google, and TikTok — and what the convergence toward embedding-based retrieval means for multi-platform strategy — algorithmic convergence in paid media is the companion read to this article.

If you're also working through how to calibrate learning phase expectations for Andromeda-aware campaigns, meta ad performance inconsistency covers the diagnostic framework for separating learning-phase noise from genuine performance signals.

Three Immediate Actions for Your Account

If you've read this far and want to apply something this week:

Action 1: Semantic audit of top-spending ads. Pull your top 5 ads by spend for the last 30 days. Write one sentence for each: what is the specific audience this ad is speaking to, and what is the specific thing it's saying to them? If you can't write that sentence clearly, the ad's semantic signal is probably weak. Prioritize creative refreshes for the vaguest ads first.

Action 2: Competitive semantic mapping. Spend 30 minutes in AdLibrary's unified ad search searching for your top 3 competitors. Filter for ads that have been running 45+ days using ad timeline filters. For each long-running ad, note the core semantic angle: what problem, desire, or identity is it speaking to? This is your category's validated semantic territory map.

Action 3: Launch one semantically differentiated variant. Based on your audit and competitive research, identify one semantic angle you're not currently using but that appears in your competitive landscape (or an adjacent category). Write one ad that leads hard with that angle. Run it in a separate ad set with broad audience targeting and your standard bid. Track impression volume in the first 48 hours relative to your other ad sets. That's your Andromeda retrieval comparison.

For the creative strategist workflow view of this process, AdLibrary's research tools support all three steps — from search to timeline filtering to AI enrichment for semantic framing analysis. The Ad Budget Planner can help you size the test budget for that differentiated variant relative to your current account spend.

The ROAS calculator gives you a quick benchmark: if the semantically differentiated creative produces even a 10% improvement in effective CPM — through better retrieval performance — what does that mean for your overall return at current budget levels? Often it's larger than it first appears.

Andromeda has been live long enough that its effects are no longer theoretical. The accounts adapting their creative strategy to semantic retrieval are seeing it. The accounts still running generic creative and blaming bid competition haven't connected the cause. Understanding the mechanism puts you on the right side of that diagnostic.