Meta Says Its GEM Model Lifted Conversions 18.9%. Compared to What?

Meta says its Generative Ads Model (GEM) has delivered an 18.9% conversion improvement and a 12% click-through rate improvement across its ad surfaces, according to an engineering post published October 13, 2026. The post gives no baseline, test period, or methodology, so the figure is a model benchmark, not a forecast for your account. The useful part is the mechanism: low-volume conversion ad sets seem to gain the most.

The numbers come from Meta's own engineering blog, written by seven engineers with a long list of contributors, and PPC Land went through it line by line the same day. Most of the post is about architecture: a new design called Kunlun, a post-training technique called LoopFM, attention kernels, GPU efficiency. That stuff is real engineering and I'm not going to pretend I can audit it. The two marketing numbers are a different story, because they are the part people will screenshot and paste into a QBR deck.

Anyway, let's look at what Meta actually said, and what it didn't.

Three sets of numbers that don't quite line up

The October post contains two sets of figures. The headline is "18.9% conversion improvement and 12% ad click-through rate (CTR) improvement." A few paragraphs later, Meta says GEM "contributed 5.8%, 4.1%, and 1.6% conversion" uplift on Facebook, Instagram, and off-site conversion "for targeted users." It never explains whether the second set sits inside the 18.9% or is a separate measurement. PPC Land flagged the same gap.

Then there's the earnings call. On Meta's Q2 2026 call in July, CFO Susan Li said its user understanding models, combined with GEM, generated an 8.3% increase in ad clicks and a 15.7% uplift in conversions on Facebook. So in July, Facebook conversions were up 15.7% (GEM plus another model). In October, GEM's Facebook contribution is 5.8% for targeted users. Those can both be true. They're probably measuring different things over different windows on different user populations. But nobody outside Meta can reconcile them, and Meta hasn't tried.

For context, the first disclosure was much smaller. Back in the Q2 2025 earnings cycle, Meta said GEM drove roughly 5% more conversions on Instagram and 3% on Facebook. The original November 2025 GEM post framed the model as the "central brain" for ads ranking. Going from 3-5% to 18.9% in about fifteen months is either a big jump in model quality or a change in how the number is counted. Honestly, it's probably some of both.

Put bluntly:

An 18.9% lift with no denominator is a press release figure, and it should go in your deck with that label attached.

Think of it like a car maker's fuel economy sticker. The engine improvements are real, the lab number is real, and nobody gets that mileage on their actual commute. Your commute is your account: your offer, your margins, your pixel quality, your creative.

The sentence in the post that actually matters to media buyers

Buried in the technical writing is one line I'd put above the headline number: "The generalizable representations learned on data-rich click tasks transfer into data-sparse conversion tasks."

In plain English, GEM learns from the huge volume of clicks and engagement across Facebook, Instagram, and other surfaces, then uses that to make better guesses about conversions where conversion data is thin. The post says GEM "jointly predicts clicks and conversions across all surfaces" and covers deep-funnel goals like off-site conversions and app installs.

If that holds up, the accounts that gain most probably aren't the ones already generating thousands of purchases a week. They seem more likely to be mid-size accounts with sparse conversion signal, where the old per-surface models were basically guessing. That's my read, not Meta's claim (the post never mentions small advertisers at all), so treat it as a hypothesis worth testing.

And to be fair, this isn't entirely new. Meta has been telling advertisers to feed it more signal for years. What's changed is that the model on the other end seems to get more out of each event, which makes fragmented account structures look even more wasteful than they already did.

Consolidation is boring, which is probably why it works

The practical move is to stop splitting thin conversion volume across a pile of ad sets. The common rule of thumb, as JetFuel's learning-phase guide lays it out, is that an ad set exits learning after about 50 optimization events within 7 days, counted per ad set and not per campaign. Their example: at a $40 target cost per purchase, an ad set needs roughly $2,000 of spend over the week to get there.

So here's a 10-minute check you can run this morning:

• Pull every active conversion ad set from the last 7 days and sort by optimization events.
• Flag any ad set under 50 events. Multiply your target CPA by 50 and compare it to the ad set's weekly spend. If the spend is less than half that number, it's probably too thin to give the model anything useful to work with.
• Merge flagged ad sets that share an optimization event and a broadly similar audience. Two ad sets at 25 events each becoming one at 50 is, in most of the write-ups I've read, the easiest win available in a typical account.
• Avoid big edits for the following week. JetFuel notes that budget changes of roughly 20% or more, plus audience, creative, or event changes, reset the 50-event count to zero.

Personally, I'd do this before touching anything else. It costs nothing, and if the cross-surface transfer works the way Meta describes, a consolidated ad set gives it a much cleaner problem to solve.

Run your own lift test before you believe anyone's percentage

The skeptic case deserves airtime too. PPC Land has covered growing advertiser skepticism about Meta's automation, including one incrementality test in which Advantage+ generated only 17% of the conversions Meta's attribution reported. That test isn't described in much detail, so I wouldn't hang a strategy on it. But it's a useful reminder that "conversions" in a Meta engineering post and "incremental conversions" in your P&L are different metric types. One is what the model predicted and got credit for. The other is what wouldn't have happened without the ad.

If you spend enough to run a Conversion Lift study or a geo holdout, now is a decent time. Run it on your consolidated structure, not the old fragmented one, so you're measuring the setup you actually plan to keep. A rough benchmark: if Meta-reported conversions come in at more than about 2x what the holdout shows, your attribution window is doing a lot of heavy lifting and your CPA targets are probably too generous.

If you can't afford a formal test, at least compare your blended conversion rate quarter over quarter using your backend data (Shopify, your CRM, whatever holds actual orders) instead of Ads Manager. If your backend numbers didn't move while Meta-reported conversions climbed, the 18.9% isn't showing up for you, whatever the chart says.

Why Meta keeps publishing these posts

This is Meta's third long-form GEM document in under a year, after the November 2025 post and an August 2026 post on training efficiency that claimed GEM now runs at 20-25% model FLOPs utilization while total training compute grew 4x in 12 months. That's a lot of public engineering detail for an ads product, and I don't think it's accidental. Meta is spending heavily on AI infrastructure, and engineering posts with big percentages help justify that spend to investors as much as to advertisers.

None of which makes the model bad. It probably is better than the old stack. The question is just whose number you plan around. We've written before about how Meta's own AI agent now grades Meta's own ads, and how automatic customer exclusions quietly changed what "new customer" means. The pattern is the same each time: the system gets smarter, and the reporting on how well it works comes from the same place.

My guess for the next disclosure

I'd bet that by the Q4 2026 earnings call, Meta quotes a GEM conversion figure above 20% and still doesn't publish a baseline. I'd also bet the accounts that see real gains are the ones that consolidated, not the ones that bought the most new tools.

I could be wrong on the first part. The second part I feel better about. Fewer, better-fed ad sets have outperformed fragmented ones for a long time now, and a model built to stretch thin signal across surfaces seems likely to make that gap wider, not smaller.

By Notice Me Senpai Editorial