PERFORMANCE MARKETING 15 min read · Updated 25 August 2026

Andromeda and Lattice: how Meta’s ad algorithm actually works now

Meta rebuilt its ad system around two things: a retrieval engine called Andromeda and a unified ranking model called Lattice. What each does, and what it means for how you should actually structure campaigns.

Careers Ninza performance marketing faculty

Careers Ninza · Kolkata, India

Most advertisers still structure Meta campaigns the way they did in 2020: tight audiences, separate ad sets per interest, manual bid control, one or two creatives per set. That approach now actively fights the system it runs on.

Meta rebuilt the machinery underneath. Two components matter: Andromeda, which retrieves candidate ads, and Lattice, which ranks and predicts. Understanding what each does explains why old habits stopped working.

What Andromeda does

Andromeda is the retrieval layer — the stage that decides which handful of ads, out of millions eligible, even get considered for a given impression. It runs on Meta’s own MTIA silicon and it dramatically expanded how many candidates can be evaluated per auction.

Previously, retrieval was a bottleneck. Coarse filters narrowed the pool early, which meant a well-matched ad could be discarded before ranking ever saw it. Andromeda widened that funnel by orders of magnitude.

The practical consequence: the system can now find the right person for your ad far more reliably than you can describe that person in a targeting interface. Manually narrowing an audience mostly removes people Andromeda would have found for you.

Why this broke interest targeting

When retrieval was narrow, telling Meta "show this to people interested in digital marketing, aged 25–34, in Pune" genuinely helped. It reduced a search space the system struggled with.

Now that constraint is a handicap. You are restricting a retrieval engine that would have surfaced buyers your interest list never contained. This is the mechanical reason broad targeting outperforms layered interests in most accounts — not fashion, architecture.

What Lattice does

Lattice is the ranking and prediction layer. Its significance is consolidation: Meta replaced a sprawl of separate, narrowly-trained models with a unified architecture that learns across objectives, placements and formats simultaneously.

Before Lattice, a model predicting purchase probability on Instagram Reels learned largely in isolation from one predicting lead submissions on Facebook Feed. Signal did not transfer well. Lattice shares learning across all of it.

AspectBeforeWith Lattice

Model structureMany narrow models per objective and surfaceUnified model learning across all of them

Signal sharingLimited transfer between placementsLearning transfers across placements and objectives

Cold startSlow — each model relearned from scratchFaster — shared representations carry over

Creative evaluationLargely surface-specificUnderstands creative across formats together

Why this changed the learning phase

Because signal transfers, campaigns exit learning faster and tolerate consolidation better. Fifty conversions per ad set per week remains the guidance, but a consolidated campaign reaches that threshold far quicker than eight fragmented ad sets each starving for data.

This is the mechanical argument for consolidation. Splitting budget across many ad sets does not give you control — it gives every ad set too little data to learn from.

What this means for how you build campaigns

1. Consolidate ruthlessly

—One campaign per objective, not per audience segment

—Few ad sets, generously funded, rather than many on small budgets

—Advantage+ or broad targeting as the default; narrow only with a specific reason you can articulate

—Let Advantage+ placements run everywhere rather than hand-picking surfaces

2. Creative volume and diversity are now the lever

When targeting and bidding are largely automated, creative is what remains under your control — and Lattice evaluates creative more capably than any earlier system. It is the highest-leverage input you have.

—Diversity over iteration. Five genuinely different concepts beat twenty variations of one headline. The system needs distinct signals, not near-duplicates.

—Format range. Static, video, carousel, Reels-native vertical. Lattice reads across formats, so give it range.

—Hook in the first two seconds. Still the single biggest determinant of video performance.

—Refresh on fatigue, not on a calendar. Watch frequency and CTR decay rather than replacing creative every fortnight by habit.

3. Stop micromanaging bids

Manual bidding and aggressive bid caps now constrain a system with better information than you have. Use the lowest-cost or cost-cap strategies and let the auction work. Bid caps have a place in genuinely constrained economics — not as a default.

4. Fix your signal quality instead

This is where advertisers should spend the attention they used to spend on targeting. Lattice is only as good as the conversion data reaching it.

—Conversions API server-side, not Pixel alone — browser-only tracking loses substantial signal

—Deduplicate events properly between Pixel and CAPI

—Send high-quality match parameters: hashed email, phone, external ID

—Optimise for the event that actually matters commercially, not the one that fires most often

—Feed offline conversions back where the sale closes on a call or in person — critical for education, real estate and B2B in India

An account with broad targeting and clean server-side signal will beat a meticulously segmented account with Pixel-only tracking almost every time. The leverage moved from targeting to measurement.

What to stop doing

—Stop building interest-stacked ad sets. You are narrowing a retrieval engine that works better unconstrained.

—Stop splitting by age and gender without a genuine creative or compliance reason.

—Stop duplicating winning ad sets to "scale". Raise the budget on what exists.

—Stop judging performance after two days. Consolidated campaigns need the learning phase to complete.

—Stop running one creative per ad set. You are denying the ranking model the comparison it needs.

—Stop optimising for link clicks when you want purchases or leads.

What actually still needs your judgement

Automation did not remove the marketer. It moved the job.

—The offer. No algorithm rescues a proposition nobody wants. This remains entirely yours.

—Creative concepts. The system tests and ranks; it does not invent the idea.

—Landing page and funnel. Meta optimises the click, not what happens after it.

—Economics. Deciding what a customer is worth, and therefore what you can afford to pay, is a business decision.

—Exclusions. Existing customers, recent purchasers, unqualified geographies. Genuine exclusions still matter.

A practical structure for Indian advertisers

1One prospecting campaign. Advantage+ or broad, single objective, optimised for your real conversion event. Six to ten distinct creatives across formats.

2One retargeting campaign. Site visitors, engagers and video viewers, excluding recent purchasers. Different creative angle from prospecting, not the same ads.

3Server-side CAPI from day one. With India’s device and browser mix, Pixel-only tracking loses a substantial share of signal.

4Offline conversion upload if your sale closes on a phone call. In education and services this single step often reshapes what the algorithm optimises towards.

5Weekly review, not daily. Judge on blended CAC and contribution, and give changes a full week to settle.

The honest caveat

Meta publishes engineering detail selectively, and the names, architecture and behaviour of these systems change. Treat the specifics as directional rather than permanent, and hold the principle instead: the system now retrieves and ranks better than manual targeting can, so your leverage lies in offer, creative, signal quality and economics.

That principle has survived every algorithm change of the last five years, and it will survive the next.

Meta Ads Mastery is 20 hours in a live ad account — structure, creative testing, CAPI setup and reading the numbers honestly.

SEE THE COURSE

Frequently asked questions

What is Meta Andromeda?+

Andromeda is Meta’s ad retrieval engine — the layer that selects which candidate ads are considered for each impression before ranking happens. Running on Meta’s custom MTIA hardware, it evaluates vastly more candidates than the previous system, which is why broad targeting now outperforms narrow interest stacking.

What is Meta Lattice?+

Lattice is Meta’s unified ad ranking and prediction architecture. It replaced many narrow, separately-trained models with one that learns across objectives, placements and formats simultaneously, so signal transfers between them and campaigns exit the learning phase faster.

Should I use broad targeting or interest targeting on Meta in 2026?+

Broad targeting or Advantage+ as the default. Because Andromeda retrieves from a far wider candidate pool, manually narrowing an audience mostly removes buyers the system would have found. Narrow only when you have a specific creative or compliance reason you can articulate.

How many creatives should I run per ad set?+

Six to ten genuinely different concepts across formats — static, video, carousel, vertical Reels — rather than many variations of one idea. Lattice evaluates creative across formats together, so diversity gives it useful signal while near-duplicates do not.

Why did my Meta campaigns get worse after I split them into more ad sets?+

Splitting budget fragments your conversion data, so each ad set gets too little to learn from and none exits the learning phase cleanly. Consolidation into fewer, better-funded ad sets is how you work with the current system rather than against it.

Is the Conversions API necessary or is the Pixel enough?+

For serious advertising, server-side Conversions API is necessary. Browser-based Pixel tracking loses a substantial share of events to blockers, privacy settings and browser restrictions. Since Lattice is only as good as the data reaching it, signal quality is now higher-leverage than targeting.

Should I still use manual bidding or bid caps?+

Generally no. Manual bidding and aggressive bid caps constrain a system with better auction information than you have. Use lowest-cost or cost-cap strategies. Bid caps have a place in genuinely tight unit economics, not as a default setting.

How long should I wait before judging a Meta campaign?+

A full week minimum, and let the learning phase complete first — roughly fifty conversions per ad set. Judging after two days on a consolidated campaign means judging incomplete learning, and reacting to it usually makes things worse.

We teach this, live

Every article here comes from something we teach. Sit in on a free masterclass and judge the mentors yourself.

FREE MASTERCLASS CALL +91 90888 39993