What is a Meta generative recommender?

It is the direction Meta's ads system is moving: a model that reasons jointly over ad content and a person's preferences to build the candidate set, rather than scoring each ad in isolation. The publicly documented predecessor is Andromeda, the retrieval engine Meta built with NVIDIA to widen how many ads it can consider before ranking. Generative retrieval is the successor idea. Instead of matching against a fixed index, the system generates and reasons about which ads fit a person, then ranks the survivors. For a buyer, the shift matters less as a feature and more as a change in where control lives.

What is the difference between retrieval and ranking in Meta ads?

Retrieval is the first pass. Out of every eligible ad, the system pulls a shortlist that could plausibly fit the person. Ranking is the second pass, which scores that shortlist and picks what actually serves. Historically retrieval was a narrow filter and ranking did the heavy lifting. A generative retrieval model widens and smartens the first pass, so more relevant candidates reach ranking in the first place. If retrieval never surfaces your ad, no ranking bid saves it. That is why creative variety and clean signal now feed the top of the funnel, not just the bid.

Is the Meta generative recommender the same as Advantage+?

No, and conflating them is a common mistake. Advantage+ is the campaign layer you configure: Advantage+ Shopping, Advantage+ audience, Advantage+ placements. The generative recommender sits underneath, in the retrieval and ranking engine that decides which ad meets which person. You do not toggle it. It is the plumbing every Meta campaign already runs on. Our Advantage+ campaigns guide covers the campaign layer. This post is about the engine beneath it, which is why the two read as separate stories.

Did Meta report specific performance lifts for the generative recommender?

Yes. On the Q2 2026 earnings call (29 July 2026), Meta's CFO, Susan Li, reported that early LLM retrieval pilots drove a 1% increase in app event conversions on Instagram. Combined with the GEM model for ads ranking and sequence learning, these advances generated an 8.3% increase in ad clicks and a 15.7% uplift in conversions on Facebook. These are platform-wide averages across Meta's full book of business; a platform-average lift does not transfer 1:1 to any single account, and your result will depend on account structure, signal quality, and vertical.

How should the generative recommender change my Meta campaign structure?

Three moves. First, consolidate. Fragmented ad sets starve the model of the volume it needs to learn, so fewer, broader ad sets usually beat many narrow ones. Second, raise creative volume and variety. The retrieval model reasons over ad content, so more distinct creative gives it more angles to match. Third, protect conversion signal. Clean, deduplicated, well-mapped conversion events are the fuel the model optimises against, and broken signal produces confident-looking waste. None of this is new advice. Generative retrieval raises the penalty for ignoring it.

What is Andromeda and why does it matter here?

Andromeda is Meta's retrieval engine, publicly documented with NVIDIA, built to expand how many ads the system can consider before ranking. Its job was scale: hold a far larger candidate pool without slowing the auction. The generative recommender is the next step in that lineage, moving from wider retrieval to smarter, generative retrieval that reasons about fit rather than only matching against an index. You do not need to track version names. The point is that the first stage of the auction, the part that decides whether your ad is even in the running, keeps getting more capable, and that is where account hygiene pays off.

Does this apply across Singapore, Malaysia, Australia, the US, and Canada?

The engine is global, so the mechanics are identical across all five markets. What differs is volume. US and Australian accounts often carry enough spend and conversions to give the model rich signal quickly. Smaller Singapore, Malaysia, and Canada accounts can take longer to exit the learning phase, which makes consolidation and signal hygiene more important, not less, because you have fewer events to spare. The playbook does not change by country. The patience required does.

Meta's ad auction went generative: what changes for you

Q: Should small advertisers change anything right now?

Mostly no. If you spend little and run one or two campaigns, the engine change is invisible and you should not restructure a working account to chase a headline. The teams that benefit most are those with enough volume to give the model signal and enough creative capacity to feed variety. A business in Singapore or Malaysia spending modestly is better served fixing conversion tracking than reorganising ad sets. Over-indexing on an architecture you cannot see or control is the more common error than under-reacting to it.

What Meta actually said

On 29 July 2026, Meta reported its second-quarter results. The headline financials were strong and, for once, the strategic tell sat in the same place as the money. Advertising revenue came in at $59.363 billion, up 27 percent year over year. Ad impressions across the Family of Apps rose 14 percent, and the average price per ad rose 12 percent. Both sides of the auction went up at once. That combination does not happen by accident.

Mark Zuckerberg framed the quarter around one sentence: "AI is accelerating our core business today, powering our next generation of products, and opening the door to entirely new enterprise opportunities." The core business is ads. The acceleration he means is not a new creative tool. It is the recommendation engine underneath every campaign, and the direction of travel is generative modelling moving into the earliest stage of the auction: retrieval.

This is the successor idea to Andromeda, the retrieval engine Meta built with NVIDIA to widen how many ads the system can consider. Andromeda made the candidate pool bigger. The Meta generative recommender makes the first pass smarter, reasoning over ad content and a person's preferences together instead of scoring each ad on its own. If you run Meta ads for a living, that architectural change matters more than any dashboard update, because it moves control one more step away from the levers you pull and toward the signal you feed.

On the reported lift figures

The performance figures in this post are reported from Meta's Q2 2026 earnings call (29 July 2026), attributed to CFO Susan Li: a 1% increase in app event conversions on Instagram from early LLM retrieval pilots, and an 8.3% increase in ad clicks plus 15.7% uplift in conversions on Facebook from those advances combined with the GEM ranking model and sequence learning. These are platform-wide averages. A platform-wide lift is not an account-level guarantee; your result will depend on your structure, signal quality, and vertical.

Retrieval versus ranking, in plain English

Every ad auction runs in two passes, and most marketers only ever think about the second one. Understanding the first is the whole point of this post.

Stage 1

Retrieval

Out of millions of eligible ads, the system pulls a shortlist that could plausibly fit this person. A wide, fast first cut.

Stage 2

Ranking

The shortlist gets scored on predicted action and bid, then ordered. This is the part buyers usually picture as "the auction".

Stage 3

Delivery

The winner serves. What you see in Ads Manager is the outcome of both prior passes, not just the bid you set.

For years, retrieval was the dumb part. It was a broad filter that grabbed a candidate set, and ranking did the intelligent work of scoring. Andromeda changed the scale of retrieval: it let Meta hold a much larger candidate pool without slowing the auction, so more ads got a look before ranking. The generative recommender changes the intelligence of retrieval. Rather than matching your ad against a fixed index, the model reasons about whether the ad fits the person, generating and evaluating candidates instead of only looking them up.

Here is why that reorders your priorities. If retrieval never surfaces your ad, no bid saves it in ranking. The ad simply is not in the shortlist. So the things that used to feel like ranking inputs, mainly creative relevance and clean conversion signal, now feed the first pass too. A generative retrieval model that reasons over ad content rewards accounts that give it varied, distinct creative to reason about, and it starves accounts that show it three near-identical images and a broken pixel. The lever moved. The auction did not get simpler. It got more sensitive to what you supply.

Why a headline lift does not transfer to your account

Meta's CFO, Susan Li, reported an 8.3% increase in ad clicks and a 15.7% uplift in Facebook conversions from these advances combined with the GEM ranking model and sequence learning, plus a 1% gain in Instagram app event conversions from early retrieval pilots (Q2 2026 earnings call, 29 July 2026). Those figures are genuine. They are also platform-wide averages, and treating them as a forecast for your account is how buyers get burned.

When a platform says a new model improved conversions by some tidy percentage, the number is doing a very specific job: it is proving the model works in aggregate, across the platform's own book of business, under the platform's own definition of success. Three reasons that platform average does not become your average.

Composition. A platform-wide lift is a weighted blend of millions of advertisers. The gain is concentrated where the model had the most room to improve, which is often the accounts that were structured worst before. If your account was already consolidated and well-signalled, there is less slack to recover, so your lift is smaller. If it was fragmented, your lift could be larger, but only after you fix the fragmentation.
Signal quality. A generative model optimises against the conversion events you send it. Feed it deduplicated, correctly mapped events and it has something honest to learn from. Feed it double-counted or mis-fired events and it will confidently optimise toward the wrong thing, producing spend that looks efficient in the dashboard and is not.

Then there is the base rate. A relative improvement sits on top of whatever your conversion rate already is. The same percentage move means a different absolute outcome for a business converting at 1 percent than for one converting at 6 percent. A tidy headline number hides that entirely.

So the correct operator posture is directional, not numeric. Assume the matching is getting better. Assume the accounts that gain most are the ones set up to give the model signal. Then go make yours one of those accounts, rather than waiting to see whether a borrowed percentage shows up in your own reporting. This is also where an agentic optimisation layer starts to earn its place: the more the auction runs on signal you cannot manually tune, the more value sits in monitoring and feeding it well.

What the Meta generative recommender changes for your campaign structure

None of the following is new advice. Consolidation, creative volume, and signal hygiene have been the automated-auction playbook for years. The Meta generative recommender does one thing: it raises the penalty for ignoring them, because the engine you are feeding got hungrier and better at using what you send. Three moves, in the order most accounts need them.

Consolidate the ad sets. Fragmented structures split your conversions across too many learning phases, and a model that never leaves learning never gets the signal density it needs, so fewer and broader ad sets concentrate events and let the system learn who converts. Those dozen narrow audiences carved up for manual control? That control is now mostly theatre. Collapse them and let the retrieval model do the segmentation you used to do by hand. Our Advantage+ campaigns guide covers the consolidated structure in depth.
Raise creative variety, not just creative count. The retrieval model reasons over ad content, so distinct angles, formats, and messages give it more ways to match a person. Ten near-identical variants are one angle wearing ten costumes. Three genuinely different concepts beat that. This is where the generative creative tools connect: the point of Advantage+ creative AI is to make varied creative cheap enough to supply at the volume retrieval now rewards. Supply feeds the engine; the engine does the selection.
Protect the conversion signal. Audit the events you send. Deduplicate them. Map them to the actions that actually matter to the business, and kill the ones that fire on the wrong page or double-count a purchase. Clean signal is the fuel the model optimises against, and it is the single input a small team can most improve. A model reasoning generatively over bad signal is not smarter waste. It is faster waste.

The through-line: the levers you used to pull are being replaced by the signal you feed. An account built for manual control, many ad sets, tightly gated audiences, one hero creative, is built for an auction that no longer exists. The rebuild is not exotic. It is consolidation, supply, and hygiene, done deliberately. See our Meta ads mastery guide for the full account walk-through, and our paid social service if you want it done as a system rather than a checklist.

The manual-lever era versus the generative-retrieval era

The clearest way to see the shift is to line up what you used to control against what actually moves the outcome now. The left column is not wrong history. It is a set of habits that used to pay off and increasingly do not.

What you controlled in the manual-lever era versus what moves results in the generative-retrieval era.
Control point	Manual-lever era	Generative-retrieval era
Audience	You built and gated tight interest and lookalike audiences by hand.	The retrieval model finds the audience from signal. Your job is to give it clean conversions to learn from.
Account structure	Many narrow ad sets for granular control and reporting.	Fewer, broader ad sets so conversions concentrate and the model exits learning.
Creative	One or two proven hero ads, iterated slowly.	Varied, distinct creative supplied at volume, so retrieval has more angles to match.
Bidding	Manual bids and hand-tuned budgets per ad set.	Value-based goals and budget at the campaign level, with the system allocating.
Conversion signal	Nice to have. Optimisation still ran on clicks and reach.	The primary fuel. Deduplicated, well-mapped events decide what the model learns.
Where your effort pays	Inside the auction, tuning knobs mid-flight.	Before the auction, on supply and signal quality that feed retrieval.

Read the right column as a job description. The work did not disappear. It moved earlier, from tuning a live auction to preparing the inputs an increasingly capable engine consumes. Buyers who miss that keep pulling levers that are no longer connected to anything.

The five-market read and who should wait

The engine is global. The mechanics are identical whether you buy from Singapore, Kuala Lumpur, Sydney, New York, or Toronto. What differs is not the model. It is how much signal your account can hand it. Geography is a distraction here, and most account reviews we sit in spend far too long on it.

Volume decides the payoff. A US or Australian account carrying substantial monthly spend and a steady stream of conversions gives the retrieval model rich signal quickly, exits learning fast, and sees the benefit of consolidation soonest. A smaller account in Singapore, Malaysia, or Canada, running the same playbook on a fraction of the events, takes longer to leave learning and needs the hygiene discipline more, not less, because it has fewer conversions to waste on a fragmented structure or a broken pixel. The advice does not change across the five markets. The patience required does.

Now the honest part about who should not act. If you spend little, run one or two campaigns, and the account is quietly working, do nothing structural because of this news. The engine change is invisible at low volume, and restructuring a functioning account to chase an architecture you cannot see or control is the more common mistake than under-reacting to it. The teams this genuinely rewards have two things: enough volume to give the model signal, and enough creative capacity to feed variety. If you have neither yet, the highest-return move is not reorganising ad sets. It is fixing conversion tracking and building a creative supply you can sustain. Do that, and you are ready for the generative-retrieval era whenever your volume arrives.