On 28 October 2026, Meta published details of an upgrade to Andromeda, the system that shortlists eligible ads before ranking and auction on Facebook and Instagram. Meta says its production retrieval pipeline can score more than one billion ad candidates per second, around 100 times the throughput of the 2024 system. For advertisers, the practical question is whether their ads give buyers distinct reasons to act and whether their reporting can tell a qualified enquiry from a form submission. Meta Engineering’s 28 October announcement.
What is Meta Andromeda, and where does it fit in ad delivery?
Andromeda is Meta’s AI retrieval engine. It scores eligible ads for an impression and sends a shortlist to later ranking and auction stages. Those stages help determine which ad appears. An ad can qualify for an impression and still miss the shortlist or lose the auction. Meta introduced Andromeda in December 2024 as part of its work on personalized ad retrieval and Advantage+ automation. Meta’s original Andromeda announcement.
- Eligibility. Campaign settings and platform rules determine which ads can be considered.
- Retrieval. Andromeda scores candidates and shortlists those it estimates are relevant.
- Ranking and auction. Later models and auction rules determine which ad is shown.
Do you need to turn it on? Andromeda operates within Meta’s delivery infrastructure. The October engineering announcement describes production changes. Advertisers can keep using their existing campaign settings and choose what to test based on their account evidence. Meta’s production architecture notes.
What changed in the 2026 Andromeda upgrade?
Meta redesigned the retrieval model and how it runs on GPUs. The model uses a small subset of its specialized components for each candidate. The system also groups candidate matching, scoring and shortlist selection on the same GPU, cutting the amount of data it needs to move between servers. Together, those changes let the production pipeline compare many more candidates within its time budget.
| Figure | What Meta measured |
|---|---|
| 100× | Retrieval throughput compared with the 2024 Andromeda system. |
| Over 1 billion per second | Ad candidates scored by the production retrieval pipeline. This measures scoring capacity. |
| About 300× fewer | Floating-point operations required at inference time by the specialized retrieval model versus the 2024 retrieval model. |
| +2.4% | Increase in Facebook conversions that Meta attributes to Andromeda since 2025. |
The first three figures describe technical performance. The conversion figure is a result Meta reports for Facebook over time. Use your account’s own results to set expectations for Facebook, Instagram and individual campaigns. The announcement date tells us when Meta published the details. It gives advertisers no single account-level rollout date. Meta Engineering’s benchmark definitions.
How do Andromeda, GEM and GenRecs work together?
They address different problems in Meta’s recommendation system. Andromeda retrieves candidate ads. GEM, Meta’s Generative Ads Model, learns from interactions and ad features at scale and transfers that learning to models used in delivery. GenRecs tokenization represents items through shared concepts drawn in part from text and images, alongside engagement signals. That helps models relate a new ad to similar items before it has much engagement history of its own. Meta Engineering on GEM · Meta Engineering on GenRecs.
What counts as creative diversification in Meta Ads?
A distinct creative gives someone a different reason to consider the same offer. Changing the headline or resizing the same apartment rendering can be useful for testing copy or placements. A genuine concept test compares different questions the buyer needs answered, with a visual and proof point suited to each one. Meta Blueprint teaches creative diversification across messages, visuals and formats. Meta Blueprint’s creative diversification training.
SEM Lion’s view. Group ads by the buyer question they answer before deciding how many new concepts to produce. If your team has one strong layout ad, try a price breakdown or an on-site progress update. That gives you a way to learn which information brings in qualified interest. Test as many ideas as your budget and expected lead volume can support. Keep campaign eligibility and audience controls in view, especially when the ads promote housing. Meta’s Special Ad Category guidance · SEM Lion’s guide to Meta Ads Creative Studio.
Here is a hypothetical example for a residential property campaign. Each ad answers a different buying question about the same development.
| Concept | Buyer question | Suggested creative |
|---|---|---|
| Affordability | What upfront costs should I plan for? | Short video with verified figures and clear assumptions. |
| Layout and space | How usable is the apartment day to day? | Walkthrough that shows room dimensions and storage. |
| Location | What is nearby and how do I get there? | Map or street footage showing verifiable routes and amenities. |
| Price transparency | What is included, and which fees are extra? | Carousel that separates the listed price from other verified costs. |
| Trust | What has actually been built so far? | Dated construction footage and a documented progress milestone. |
How will SEM Lion test the change?
We would keep a proven ad as the baseline, introduce a small set of distinct concepts, and check which ones produce qualified enquiries or viewings. A 14-day window gives the team a first review point. The decision depends on delivery volume and how long it takes sales to qualify the leads.
- Audit the creative. Label each active ad by buyer question, offer, proof and format. Group variations of the same idea together and pick two or three gaps worth testing first.
- Define the outcome before launch. Agree with sales on what counts as a qualified lead and how a booked or attended viewing is recorded. Keep the offer and conversion destination consistent where practical. If delivery volume supports it, use a split test to compare the new concept with the baseline. When ads share a budget and delivery varies, treat the comparison as directional.
- Check the tracking path. Follow a test enquiry from the website or Instant Form into the CRM. Check the Meta Pixel and Conversions API where applicable. For eligible Instant Form campaigns, Meta’s CRM integration can send later lead outcomes back to the platform. Align campaign reporting and CRM attribution before comparing results. Meta Blueprint on Conversions API for CRM.
- Read past the form submission. Compare spend, lead count, qualified leads, booked and attended viewings, and sales when enough time has passed. Calculate cost per qualified lead as attributed spend divided by qualified leads. A concept that barely delivered needs more evidence before a decision.
- Decide what to keep testing. Keep the ads that show sustained qualified results. Where results are thin, extend the observation window or reduce the number of concepts competing for the same budget.
Hypothetical example. Ad A spends €300 for 30 leads and three qualify. Ad B spends €300 for 15 leads and six qualify. Ad A has the cheaper form submission at €10 versus €20. Ad B has the lower cost per qualified lead at €50 versus €100. We would investigate the difference with sales and check how many of those qualified enquiries turn into viewings before scaling.
Did Andromeda cause a sudden change in campaign performance?
An account’s reporting can show when performance changed. To investigate why, compare the same weekdays and markets, then check what moved first. Rising CPM with stable click and conversion rates points to a different problem from stable CPM with fewer visitors completing the form. An increase in reported leads accompanied by a fall in qualified enquiries calls for a CRM and follow-up check. Auction pressure around holidays or major events is worth testing as a market-specific hypothesis alongside creative fatigue, delivery mix, site changes and tracking.
What should you ask an AI agent to audit?
Use read-only reporting access and share aggregate figures. Start with one market and a defined attribution window so the agent can compare like with like.
Review [Meta ad export] for [market and date range]. Group the ads by buyer question, offer, proof and format. Identify repeated concepts and suggest two distinct alternatives that use verified product facts. Cite the ad IDs or source rows. Label assumptions and give the team a testable hypothesis for each idea. Return a plan for review.
Using [ad-level spend and lead totals] and [aggregate CRM outcomes], compare cost per lead, cost per qualified lead and cost per booked viewing by concept. State the attribution window and qualification rule. Flag missing data, low delivery and any ads that are too recent for a fair comparison. Explain what extra evidence would change your recommendation.
If you have creative and lead data but the reports tell different stories, SEM Lion can help you trace what happened from impression to qualified enquiry. Explore agentic reporting for Google and Meta Ads · Request a SEM Lion marketing evaluation.