Meta AI Ad Delivery System Explained: How Andromeda Decides Who Sees Your Ads

How the Meta AI Ad Delivery System Actually Works

Meta's ad delivery infrastructure is built around a two-stage machine learning pipeline. The first stage is retrieval — Andromeda scans a candidate pool of potentially millions of ads and rapidly narrows the field to a shortlist of a few hundred that are plausible matches for a given user at a given moment. The second stage is ranking — a deeper, more computationally intensive model scores each shortlisted ad against a predicted probability of the advertiser's chosen outcome, factoring in bid, estimated action rate, and user experience signals.

What makes Andromeda distinct from the legacy delivery system is the scale of the embedding model underpinning retrieval. Rather than matching ads to users through rigid audience segments, Andromeda generates vector representations of both ads and users, then finds approximate nearest neighbours in that high-dimensional space. This means an ad can surface for a user who shares no explicit demographic overlap with your defined audience, but whose behavioural embedding is statistically adjacent to people who have converted for you before. This is why the Meta Andromeda vs legacy audience targeting shift feels so disorienting to advertisers who spent years engineering interest stacks and exclusion lists — the system has fundamentally changed how "matching" happens.

"Andromeda doesn't ask 'does this user fit this audience?' — it asks 'is this user statistically likely to take this action based on everything we know about people who already have?'"

The practical implication is that your inputs to the system — conversion events, creative quality signals, landing page behaviour — are now more determinative of who sees your ads than any targeting parameter you set manually. The system uses your optimisation target as its north star, then works backwards to find the users most likely to deliver it.

Understand the Signals Andromeda Weighs in Every Auction

To work effectively with Andromeda, you need a clear map of the signal types it ingests. These fall into four broad categories, each with different levels of advertiser control.

The most critical insight here is that conversion signals and creative signals are both high-control levers. Many advertisers fixate on conversion signal quality (rightfully so) while underinvesting in the creative signals that determine whether Andromeda's retrieval stage even surfaces your ad in the first place. An ad with poor early engagement metrics will be demoted from the candidate pool before the ranking model ever scores its bid competitiveness. For a deeper dive into how these signals interact strategically, the Meta Andromeda ad targeting strategy guide covers the creative-led performance framework in full.

Structure Your Campaigns to Feed the Learning Phase Faster

Andromeda's ranking model is personalised — it improves its predictions for your specific account as it accumulates conversion data. The learning phase is the window during which the model is actively calibrating. Structuring campaigns to exit this phase quickly and cleanly is one of the highest-leverage operational decisions you can make.

Follow these specific actions to accelerate learning phase completion:

• Consolidate ad sets aggressively. Each ad set runs its own learning phase independently. Splitting audiences across twelve ad sets dilutes conversion volume per unit and extends the time each one spends in learning. Industry practitioners commonly report that accounts with fewer, broader ad sets exit learning phase significantly faster than fragmented structures.
• Choose the lowest-funnel event you can reliably hit at volume. If your campaign generates fewer than 50 purchase events per week per ad set, optimise for a higher-volume micro-conversion like Add to Cart or Initiate Checkout, then graduate to Purchase once volume supports it.
• Avoid edits during active learning. Changing bids, budgets above roughly 20%, creative sets, or audiences resets the learning phase. Batch changes together and time them to low-traffic periods.
• Use Campaign Budget Optimisation (CBO) to let the system allocate across ad sets. CBO allows Andromeda to concentrate spend on whichever ad set is currently finding the most efficient delivery, compounding learning faster than manual budget splits.
• Seed new campaigns with Advantage+ Audience rather than narrow custom audiences. This gives the retrieval model a broader initial candidate space to explore, reducing the time needed to identify high-probability converters.

A campaign that exits learning cleanly typically shows more predictable CPAs, tighter delivery windows, and less volatility during scaling — the system has built a reliable model of who to show your ads to, and it stops experimenting as aggressively.

Optimise Creative to Influence Retrieval and Ranking Scores

Creative is now the primary targeting mechanism inside Andromeda. The retrieval model generates an embedding for each ad based partly on the semantic and visual content of the creative itself. An ad that communicates a specific problem, product category, or emotional register will be retrieved more consistently for users whose profiles align with that space. Vague or generic creative produces a diffuse embedding that makes Andromeda's job harder.

Take these actions to strengthen your creative's signal value:

• Lead with a specific, recognisable problem or outcome in the first three seconds. Andromeda's early engagement signals — specifically the three-second video view rate and the swipe-away rate — are proxy indicators the retrieval model uses to assess relevance match quality. A hook that is immediately specific to the right viewer keeps the right people watching and bounces the wrong ones quickly, which actually improves your signals.
• Test creative variables in isolation. Change one element at a time — hook, format, offer framing, CTA — so you can identify which variable drove any change in CTR or conversion rate. Andromeda will automatically allocate more impressions to better-performing variants within an ad set.
• Maintain creative diversity across format types. Static images, short-form video, carousels, and collection ads produce different embedding signatures. Running multiple formats increases the surface area across which Andromeda can find retrieval matches.
• Refresh creative before fatigue metrics degrade. When frequency rises above 3–4 and CTR begins declining, the creative's engagement signals start turning negative. Replacing fatigued creative before this point preserves your overall ad set health score.
• Align landing page content with ad creative semantics. Andromeda factors post-click quality signals — including landing page load speed and time-on-page indicators fed back through the Pixel — into its quality ranking scores. Mismatches between ad promise and landing page delivery create post-click signal noise.

Measure and Iterate Without Destabilising Delivery

One of the less-discussed consequences of Andromeda's architecture is that standard A/B testing instincts — frequent experiments, rapid iteration, constant optimisation — can actively destabilise delivery if applied without care. Every significant structural change restarts learning. The challenge is building a measurement cadence that generates enough signal to improve without constantly resetting the model.

These actions create a sustainable measurement loop:

• Implement Conversions API (CAPI) with direct server-to-server integration. Browser-side Pixel alone loses a significant share of conversion events to ad blockers and browser privacy restrictions. CAPI closes this gap, giving Andromeda a more complete conversion signal to learn from. Many practitioners report that adding CAPI alongside Pixel can materially reduce CPA by improving the quality of training data.
• Set a consistent attribution window and hold it stable. Switching between 1-day click and 7-day click attribution mid-flight changes the apparent conversion volume the system sees, which can trigger unnecessary learning phase resets. For a full breakdown of attribution model implications, see the Meta Andromeda measurement attribution guide.
• Use Meta's built-in A/B test tool for creative experiments rather than duplicating ad sets manually. The built-in tool controls for audience overlap and ensures clean holdout groups, which produces more statistically reliable results without disrupting your primary delivery campaigns.
• Review performance at the ad set level on a weekly cadence, not daily. Daily variance in Andromeda's delivery is high enough that single-day metrics are routinely misleading. Weekly aggregates smooth out exploration noise and give you a more accurate read on underlying efficiency.
• Track view-through conversions alongside click-based conversions to capture the full influence of upper-funnel impression exposure, particularly for brand-aware purchase decisions with longer consideration cycles.

Common Mistakes That Undermine Andromeda's Performance

Even advertisers who understand Andromeda conceptually consistently make a handful of structural errors that limit what the system can deliver for them.

• Over-segmenting audiences into micro ad sets. Splitting by age bracket, gender, device type, and interest simultaneously creates a fragmented signal environment where no single ad set accumulates enough conversion data to learn efficiently. The system works best with broader, consolidated structures.
• Optimising for a conversion event with insufficient weekly volume. Choosing Purchase as your optimisation event when your campaign generates only eight purchases per week gives Andromeda too little training data to build a reliable prediction model. The result is erratic delivery and high CPA volatility.
• Making frequent small budget adjustments. Increasing or decreasing budgets by more than 20% in a short window triggers renewed exploration behaviour. Incremental scaling — around 15–20% increases held for at least three to five days before the next increment — preserves delivery stability.
• Neglecting creative refresh cycles. Running the same creative for extended periods while frequency climbs is one of the fastest ways to generate negative engagement signals that suppress your ad's retrieval scores. Creative fatigue is a system signal, not just a user perception issue.
• Relying on exclusion audiences to do the work that creative should do. In the Andromeda era, exclusion lists have diminished targeting precision relative to their legacy-system utility. Investing that effort into sharper creative that self-selects the right audience produces more durable results.
• Ignoring CAPI event match quality scores. A low event match quality score means Andromeda is receiving conversion signals it can't reliably match to user profiles. Improving the match quality — by passing more customer data parameters like email, phone, and external ID — directly improves the training data quality.

Expected Results and Timeline

Aligning your campaign structure, creative strategy, and measurement setup with how Andromeda actually operates is not an overnight fix. Here is a realistic timeline for what to expect at each stage:

• Days 1–7 (Learning Phase): Expect elevated CPAs and inconsistent delivery volume. The system is actively exploring. Resist the urge to edit. Focus on ensuring CAPI is firing cleanly and that your optimisation event is accumulating volume.
• Days 7–21 (Early Stabilisation): Delivery patterns begin to regularise. You will typically see CPA trending downward and CTR stabilising as the model identifies its highest-probability audience clusters. This is the window to begin noting creative performance differentials.
• Days 21–45 (Scaling Window): If your primary ad set has exited learning and is performing within target, this is when to begin incremental budget scaling. Introduce new creative variants through the built-in test tool rather than by adding new ad sets.
• Day 45+ (Optimised Delivery): Well-structured campaigns operating in this window typically show predictable CPAs, efficient spend pacing, and clear creative performance hierarchies. The main maintenance tasks are creative refresh before fatigue and monitoring CAPI signal quality.

"The accounts that scale most efficiently on Meta in 2026 are those that treat Andromeda as a partner system — feeding it clean signals, stable structures, and strong creative — rather than trying to engineer around it with audience gymnastics."

Industry observations from practitioners managing large Meta budgets consistently show that accounts making the structural shift toward fewer, broader ad sets with high-quality conversion signals achieve more stable and scalable performance than accounts running highly segmented legacy structures. The timeline to realise these gains is typically four to eight weeks from when the structural changes are fully implemented.

Frequently Asked Questions