Meta Ads for DTC brands: inside the multi-stage architecture powering ad ranking and sequencing.

Enter Andromeda, GEM, Lattice. A pick'n'mix of Meta Ads buzzwords that have been doing the rounds on your LinkedIn feed this year.

In August, Meta's engineering team published a post called "From User Sequences to Scaling Laws", laying out the architecture behind how your ads actually get ranked. In my opinion, the most useful content that Meta has put out for media buyers this year, and very few in our industry have taken the time to read and understand it.

So buckle up, settle in, and join me for what should be a... simply scintillating(?) look at the digital machinery absorbing thousands of your advertising pounds each day.

First, here’s a quick glossary, because Andromeda, GEM & Lattice are sometimes used interchangeably, and they shouldn't be

Andromeda is the retrieval engine. It's the first stage of delivery, not the last. Its job is to take a pool of tens of millions of eligible ads and narrow it to the few thousand candidates worth scoring properly for a given person, in milliseconds. Meta have leveraged hardware (NVIDIA Grace Hopper Superchips plus its own custom MTIA silicon chips) to enable what it describes as a 10,000-fold increase in retrieval model complexity compared with the system it replaced. That number gets quoted a lot as "Meta's algorithm is 10,000X more powerful" by media buyers, which is not necessarily true. It's the shortlisting stage that got 10,000X more complex.

Lattice is the ranking architecture. It consolidated what used to be thousands of smaller models - siloed by surface (Feed, Reels, Stories) and by campaign objective - into a much smaller set of unified networks that predict multiple outcomes simultaneously. For example, a user's actions on Reels now informs which ads appear in Feed, and post-click activity fed back into the ad platform informs conversion-based predictions.

GEM (Generative Ads Recommendation Model) is the foundation model sitting above all of it. I imagine GEM would simply be too expensive to run and serve ads directly. So Meta runs a 'teacher-student' setup: GEM learns at LLM scale across the whole ecosystem, then distils what it knows into the lightweight models that actually run in the auction.

Retrieval, then ranking. Andromeda picks the shortlist. Lattice and the GEM-distilled models score it.

The Meta engineering post I'm unpacking today is about the sequence modelling that lives inside that ranking layer - and it's the part that most easily explains where your brand's ad budget goes.

The auction isn't a 'traditional' auction

A common misconception is that Meta Ads works like a traditional ad auction, where the highest bid or the biggest budget wins the impression. It doesn't.

Meta calculates a total value for each retrieved ad. Combining your bid with estimated action rates - the model's prediction that this specific person takes the action you're optimising for - and ad quality. Your brand's bid is one input of three, and it's the one you have the least leverage over.

This means the most valuable question is no longer "what should my bid strategy and budget be?" but "how does Meta arrive at those estimated action rates, and how can I influence this?"

The technical aspects of Meta Ads' multi-stage sequence modelling

Ranking models have a real challenge: deeper models improve predictions, but every millisecond of inference during an auction costs Meta money.

Meta's earlier method relied on hybrid setups. One model processed user event sequences, while another managed sparse feature interactions.

This worked, but Meta found three problems:

• Lossy knowledge transfer between components.

• Ongoing reliance on manual feature engineering.

• Scaling limits due to interference between the ranking and sequence components.

(Stick with me, I’m not technical either and had to Google what each term meant in simple terms)

Meta's fix was to split the sequence model into two stages that scale independently:

1. The offline user model. Deep transformer models process long user histories asynchronously - sequence lengths in the thousands - and produce cached embeddings capturing deep behavioural patterns. A key point here is that 'offline' in this context, has nothing to do with the Internet or 'offline conversions'; it means the model runs separately on Meta's backend, away from the live 'online' auction.

Critically, this stage keeps user features strictly separate from ad and context features. The embedding for a user has to be independent of any particular ad - otherwise you couldn't cache it and reuse it across every auction that user enters.

2. The online ranking model. When an impression opportunity comes up, this stage combines the cached user embedding with fresh signals and the actual ad information to produce the final ranking. It's optimised for speed and operates under a strict latency budget, but it gets to lean on all that expensive representation work already done offline.

Here's the bit worth sitting with: the system doesn't calculate the user's detailed profile representation in real-time. It's precomputed and cached. Only the ad-specific scoring happens live. That's the whole gig - it lets Meta keep scaling model complexity on the offline side, without a corresponding spike in serving cost.

So when you hear advertising folk like me rattle on that "creative is the new targeting," or harping on about the importance of persona-led creative and refresh cadence - this is because its feeding the mechanism underneath it.

Your audiences already exist. The system has already built the user profiles. Your brand's creative is what gets scored against it. We want our brands to take more shots on goal, with higher volumes of varied ad creative.

Two architectural breakthroughs in Meta Ads' sequencing and ranking system that are worth knowing by name

1. Dense tokenization. Meta explains it best, so I’ll let their engineers take over and explain here:

This tokenization approach integrates sparse features with sequential behavioral data into a single dense vocabulary, enabling attention mechanisms to discover interactions independently.

My understanding in plainer terms is: rather than engineers hand-crafting the cross-feature interactions that they think matter, everything goes into one pot and the model learns the interactions straight from the data. Traditional ad recommendation/retrieval systems needed a human to shape its decision making say: "recently viewed category × time of day" was a logic worth building into the algorithm. The new Meta Ads system doesn't.

2. Target-aware multi-head attention. This is the most intriguing part of the architecture for me personally, and it's also the part most commonly bastardised in the LinkedIn explainers.

Tokenised sparse features and ad candidate information are fused with the user's behaviour sequence, then processed through a memory-efficient form of multi-head attention. "Target-aware" means that each layer weights the user's past behaviours against the specific ad being scored. The same browsing history gets read differently depending on which creative is in front of it - a user's three-week-old skincare research is highly relevant to one ad and noise to another, and the model can now make that distinction per-candidate.

Stack several of these aligned attention blocks, and each layer captures progressively higher-order interactions between the target ad and the user's history, distilling very long sequences into compact representations.

Does it actually work? Meta's numbers

Together with its broader modelling innovations, Meta reports these advances contributed to a cumulative lift of 6% in conversions on Instagram, 3% in conversions on Facebook, and 3.5% in ad clicks on Facebook.

However, note that the original Meta article hedges the data with: "together with our broader model innovations." Any time a channel marks its own work, I treat it as directional, and not a gospel truth.

The scaling law - and four levers that matter to us as advertisers

The headline finding is that for live ad traffic, this architecture produces LLM-style scaling laws: performance improves log-linearly with compute, and Meta says the curve shows no sign of saturation yet. To me, that's a super strong signal about where the next few years of AI-powered advertising is heading.

Meta identifies four levers for pushing this frontier. Two are internal engineering concerns. Two have direct implications for how we run ad accounts:

1. Balanced model shape: Depth, width, and sequence length have to grow together - scale one axis alone and the others could bottleneck us. Meta calls this 'the scaling synergy principle'.

This is what you've heard agencies like us talk about all year: creative volume, velocity and variation. If we don't diversify our audience segments/micro-personas, messages and angles, alongside placements and creative styles, the we run the risk of throttling reach and performance.

2. Multi-stage tunability: Scaling the online model provides steeper gains per unit of compute but latency caps these gains. Scaling the offline model follows a gentler curve, but async inference means it can scale unhindered.

3. Sequence composition: and this one should change how you think about signal. Longer sequences help, but the more striking finding is that sequence diversity beats sequence homogeneity. A balanced mix of conversion objectives and action types - views, clicks, conversions - helps to build stronger user profiles for scoring than sequences built from high-signal conversion actions alone.

If you've been aggressively optimising for purchase conversion events in the name of 'clean first-party signal' this is a super strong argument against that instinct.

Full-funnel event coverage, properly deduplicated and properly matched, is feeding a system that explicitly performs better on mixed action types with broad temporal coverage.

4. Semantic feature representation: semantic content features drawn from foundation models complement traditional collaborative filtering signals, and they're especially helpful in cold-start scenarios: new ads, new advertisers, anything without historical engagement data.

Read that as: the model can form a view on a brand-new creative from its content alone, before it has performance data to make decisions against.

Now, that doesn't eliminate the learning phase, but it does mean a genuinely novel creative concept arrives with less of a handicap than it used to - and it's another reason cosmetic variations/iterations don't buy you much. Semantically, another boring, repeitive colourway of the same static is the same ad in Metas eyes (but we'll need a whole new Merchant article to cover that one and Entity IDs).

So, what does this looks like in practice?

Picture an eCommerce brand with 50 SKUs, 40 creative concepts, three hooks apiece, plus variations on offer, angle, and customer motivator, and of course, a catalogue of headline and primary text iterations.

Historically, media buyers built combinations of ads and destinations manually, hoping to unlock something worth scaling. That approach is now competing against a system with a cached, thousand-event behavioural representation of every user, re-read against every candidate creative, at auction speed.

As individuals, our browsing behaviours, media preferences, and buying motivators are all nuanced and varied. Meta's architecture eliminates human bias and takes the guesswork out of the ad matching - personalises the customer journey per user, per impression. The likelihood of outperforming this with manual over-ride in Ads Manager is narrowing, fast.

How to win on Meta Ads in the era of AI-powered advertising?

Granular manual control of audiences, targeting, and campaign structure is dead. The opportunity lies in the quality and quantity you feed Meta's system:

• First-party data and robust conversion signals: CAPI, high event match quality, and - per the sequence composition finding - a diverse event set rather than a purchase-only diet. We call this full funnel advertising.

• A constant stream of genuinely diverse creative: Diverse in concept, angle, format, and motivator. Less so in colour palette and font.

• Consolidated structure, so the signal isn't fragmented across a dozen ad sets each learning in isolation: This is especially important for brands a modest ad budget. However, that's not to say that there aren't scenarios where our media buying team don't segment out campaigns (e.g. flash sales, seasonal campaigns etc)

In addition to this, the role of the media buyer has to adapt at the rate of change we're seeing.

The best media buyers that I know have evolved into more well-rounded 'growth marketers' (and I can confidently include the entire Lifeblood media buying department in this).

They're adding value beyond the ad account by leading forecasting, experimentation, incrementality testing, broader acquisition diagnostics, and initiating commercial conversations about maximising contribution margin, key marketing moments and the availability of inventory.

So, if your agency is still sending you meaty retainer invoices for building, 'optimising' and reporting on Meta Ads, it's time to make a change. Book a free discovery call with Lifeblood today.