Meta's ad ranking system changed in August. Not a tweak — a structural replacement.
Two engineering blog posts from Meta (August 3 and August 5) described the architecture in detail: the company has replaced its feature-scoring ranking model with Meta Generative Recommender, an LLM-based system that reasons about your ad and the user together, then predicts the best match. On the Q2 2026 earnings call, CFO Susan Li's description was direct: "a paradigm shift in how our ads system works." The combined improvement is 8.3% more ad clicks and 15.7% more conversions on Facebook, according to Engineering at Meta.
If you run Meta ads, this is not background information. It changes what your account actually needs to perform.
What Meta actually changed
Previous Meta ranking systems — including the retrieval improvements Andromeda introduced — worked by scoring individual features. The model evaluated creative elements separately, user signals separately, bid separately, then combined those scores. It was very good at what it was designed for. What it could not do well was understand what an ad actually means.
The Generative Recommender operates differently. The LLM reads the ad content and considers the user's context together, in a single reasoning step, rather than scoring features in parallel and multiplying them. Meta's engineering team describes this as reasoning about "the best ad for each person" rather than finding "the highest-scoring ad according to separate signal evaluations."
The August 3 blog described how GEM — Meta's Generative Ads Recommendation Model, the foundation behind this system — now trains at LLM scale on thousands of the latest-generation GPUs. The August 5 post introduced the multi-stage sequence architecture that makes this practical at delivery speed.
Why the performance gains are mechanically significant
These are not headline metrics someone picked to make a press release look good. They're production results from deploying an LLM-scale model to actual ad auctions.
The 8.3% increase in ad clicks means better ad-user matches, according to Engineering at Meta. The 15.7% conversion lift on Facebook is harder to fake: conversions happen off-platform, and the improvement requires that the better-matched users who clicked were also more likely to complete a purchase. That's the compounding effect of ranking by actual intent fit rather than by feature scores.
These numbers come from combining upgraded user models with GEM ranking, according to the August engineering posts. The system is not just reading ads better — it's reading users better too. Both sides of the match got an upgrade simultaneously.
What an LLM ranking system reads in your ad
This is where the operational shift is.
A feature-scoring system extracted signals from your creative — color, format, presence of text, CTA type — and scored them. What the ad actually said mattered mostly through click-through history: ads that got clicked got ranked higher over time.
An LLM reads your ad. The distinction sounds obvious but has real consequences for how you should write copy.
Vague ad copy — "Shop now. Amazing deals. Don't miss out." — gives the model minimal signal about what you sell, for whom, or in what context. The system has to infer intent from other signals, primarily behavioral. It's working with less.
Specific copy tells the model what the product is, who it's for, and why someone would want it. "Lightweight trail running shoe for flat feet. Made for runners who overpronate." That's semantic content the LLM can actually reason about when matching to a user whose behavioral signals suggest they're in-market for exactly that product category.
The ranking model is not reading your ad like a copyeditor. It's using content understanding to make a prediction about which users this ad is likely to resonate with. More informative content means a more accurate prediction. That's not a creative philosophy — it's a performance input.
The three layers your ad now runs through
The Meta Generative Recommender is not the only AI system shaping delivery. It sits in a stack:
Andromeda handles retrieval — deciding which ads are even considered for a given impression. The Entity ID and Creative Similarity Score logic from Andromeda is still active. Your ads need creative diversity to survive this stage.
The Generative Recommender handles ranking — deciding, from the surviving candidates, which ad gets served to which user. This is where the LLM reasoning happens.
Advantage+ Creative handles production — rewriting copy, applying image overlays, generating text variations. This system uses your original creative as source material. If the source is vague, the generated variations will be too.
These layers interact and the compounding is real: weak creative inputs degrade performance at every stage. Better inputs compound upward.
What to actually look at this week
Don't rebuild your campaigns. The practical question is whether your existing ad copy gives the ranking model enough to work with.
Pull the copy from your top five active ad sets. For each, ask two questions: does this describe what the product actually is, and does it say who it's for? If the body text is mostly CTA language without product description, that's the gap to close.
This does not mean longer ads. It means more specific ads. A 12-word headline that names the product and the use case is more useful to the ranking model than a 25-word headline built around energy language and vague urgency.
One more thing worth tracking: Meta's Search Terms equivalent is not as transparent as Google's, but you can review placement breakdowns and creative-level performance data to see whether the system is expanding or contracting delivery on specific creatives. Changes in the ranking engine often show up first as unexpected shifts in which creatives are getting impressions. Watch for it over the next few weeks.
If you want to see how your account's creative quality stacks up at the campaign level, the free audit at Gromerce benchmarks it against performance data from your category.
The accounts still writing copy as if a human decides who sees the ad are now being ranked by a system that reads what they wrote.
Sources: Engineering at Meta (August 3 and August 5, 2026), Meta Q2 2026 earnings call, August 2026
