Meta GEM model: 18.9% more conversions, according to Meta

What is Meta’s GEM model?

GEM, for Generative Ads Model, is Meta’s ad recommendation system. The change the company announces is architectural. It moves from isolated per-domain optimisations to a single central model, learning from data across all its platforms at once.

Put plainly: what the system learns from your behaviour on Reels influences the ad you see on Facebook. Those lessons used to live further apart.

For an advertiser, the practical effect is that the boundary between platforms blurs. What works on one feeds the others, which makes attributing a result to a specific place harder.

The figures Meta claims

In October 2026 Meta published these improvements, attributed to GEM.

• 18.9% more conversions.
• 12% more clicks.
• 5.8% more conversions on Facebook.
• 4.1% more on Instagram.
• 1.6% more on off-site conversions.

The first reading is good. The second raises questions the post itself does not answer.

The gap between the headline figure and the per-platform ones is worth noticing. A general 18.9% alongside 5.8% on Facebook and 4.1% on Instagram only adds up if the rest comes from somewhere Meta does not detail.

What the figure does not say

According to PPC Land’s analysis, Meta’s post leaves out three things. What that 18.9% is measured against, over what period it accrued, and whether it comes from control-group tests or an internal evaluation.

Without those three, a percentage stays a claim rather than a result. And it is worth remembering whose claim: these are vendor figures about the vendor’s own system, with no independent audit.

There is also an inconsistency worth noting. In its second-quarter 2026 results, Meta spoke of a 15.7% conversion improvement on Facebook. It now reports 5.8% for the same platform. Both can be right while measuring different things. Meta does not explain the difference.

The figure that changes the conversation

PPC Land reports a contrast worth more than all the percentages above. In an incrementality test, Advantage+ produced only 17% of the conversions Meta’s attribution reported.

That is the distance between what the platform claims and what the business actually gains. There is no fraud in it. Platform attribution counts everything it touches, including sales that would have happened anyway. A control-group test counts only the ones that would not.

That difference can be measured, and we explain how in our guide to incrementality testing.

While that gap exists, any improvement Meta declares on its own attribution is an improvement on a number that was already inflated.

So what can you do about it?

Little, and that is worth saying. None of the mechanisms Meta describes has a setting in Ads Manager. The changes arrive through Advantage+ automatically, with no options to configure.

That leaves advertisers one useful lever, the usual one: measure independently. If Meta decides more things and publishes less detail about how, the only way to know what the spend adds is checking it against real sales. That requires the data to arrive properly, which is the job of the Conversions API, and measurement that does not depend on the vendor’s own dashboard.

Why Meta publishes this now

There is a financial context explaining the tone. Meta has announced infrastructure investment of between 130 and 145 billion dollars for 2026, and it needs to show a return on that spend.

That does not invalidate the figures, but it explains why they arrive without methodology. A number justifying investment to the market has a different audience than a number helping an advertiser decide.

The practical recommendation is to treat these announcements as what they are. Information about where the product is heading, rather than about what will happen in your account.

Frequently asked questions