How Lattice works
Meta’s advertising system used to run on a collection of separate models, each trained for a narrow job: one surface, one objective, one kind of prediction. Lattice is Meta’s name for the architecture that consolidates that collection into a shared foundation, so that a smaller number of larger models serve many objectives and many surfaces at once.
The point of consolidation is that what the system learns in one place becomes useful in another. A pattern learned from a great deal of activity on one surface can inform predictions on a surface with far less of its own data, instead of every model starting from scratch inside its own silo.
This is a description of how Meta builds its systems, published by Meta’s own engineering teams. It is not a product, not a campaign setting and not something that appears anywhere in Ads Manager. You will meet the word in technical write-ups and in conference talks, rarely in an account.
Why it matters
It is context for a shift advertisers have felt without being told the reason. Over recent years Meta has removed granular controls, merged campaign types and pushed advertisers towards broader targeting and fewer, larger campaigns. A shared model that learns across surfaces and objectives is the engineering behind that direction: the system gets better when it sees a whole account’s behaviour than when it sees a dozen partitioned slices.
It also sets expectations about change. When the models underneath are shared rather than isolated, an update does not stay in one corner. Performance can move across several campaigns at once for reasons that have nothing to do with anything you did that week, which is an argument for judging results over longer stretches than a daily check.
Where the term gets misused
Any claim to have optimised an account for Lattice should be treated as marketing language. Nobody outside Meta can measure how a shared architecture treats one advertiser, and the details change. Ask what evidence would show the work had succeeded, and there is usually none.
The second misuse is fatalistic. Model consolidation does not mean your account structure stopped mattering, that budgets no longer need care, or that broken conversion tracking will be quietly compensated for. A shared model still learns from the data you send it, so bad or missing data spreads its damage further, not less far.
How to act on it
Consolidate where consolidating is honest. Fewer, better-funded campaigns give the system more to learn from than many thin ones, which is the practical version of this idea for a small account. Keep separate only what must be separate: different countries, different languages, different offers, and any rule your business genuinely requires, as you would in a manual campaign.
Then protect the data. Make sure conversions are actually reaching Meta and that the events describe real business outcomes rather than easy clicks. Review over weeks rather than days, so you are reading your own performance rather than the noise of a system that is being updated continuously underneath you. If the account is large enough for these decisions to carry real money, structure is worth deciding deliberately as part of Meta ads consulting rather than by habit.