How GEM works
Choosing an advert happens in stages. A retrieval stage first narrows an enormous pool of eligible adverts down to a shortlist for one person. GEM is Meta’s name for what happens next: the model that scores that shortlist and puts it in order, predicting how likely each person is to take the action the advertiser asked for.
That prediction is not the final answer. It is combined with what the advertiser is willing to pay and with signals about the quality of the experience, and only then does the auction settle who appears and what they are charged. A predicted-response score and a bid pull in the same direction: an advert people are likely to act on can win against a higher bid, and a dull advert can lose while paying more.
Two cautions. The abbreviation is used for other things in other contexts, so confirm what someone means before agreeing with them. And this is Meta’s own engineering vocabulary — there is no GEM setting, no GEM report, and no published recipe for what it weighs.
Why it matters
It is the clearest explanation of why creative is a commercial lever rather than a design preference. If the system is ordering adverts by predicted response, then an advert that earns attention is competing on better terms than one that does not, and the difference shows up in what you pay for the same customer.
It also explains why the optimisation event you choose matters so much. The model predicts the action you named. Ask for link clicks and it will rank adverts by who is likely to click, which is a different population from those likely to buy. The instruction you give is the thing being predicted.
Where the term gets misused
Watch for anyone selling ranking-model expertise. Nobody outside Meta knows the weights, they change, and no advertiser can measure their own position in a ranking. A claim that cannot be tested is not a strategy.
The other misuse is fatalism. Blaming an unseen model is an easy way to skip the visible causes of poor results: adverts that were rejected, an event that stopped firing, a landing page that is slow on a mobile connection, or an audience so narrow that the shortlist was thin before ranking began. Nepal-based accounts hit that last one often, because a small country audience narrowed further by interests leaves very little to rank.
How to act on it
Optimise for the action that makes money, once you are producing enough of those actions for the system to learn from. If you are not, optimise for the nearest reliable step and move up later, rather than asking for a rare event and starving the model.
Then work on the inputs you control. Put real effort into the advert itself, retire adverts as response falls rather than waiting for them to collapse, and make sure conversions are actually reaching Meta through the pixel and Conversions API — a model cannot rank towards outcomes it never sees. Keep audiences wide enough that ranking has genuine choice, and judge the result on cost per customer rather than on any measure of how the platform ranked you, which you will never see anyway. The related idea worth understanding next is the estimated action rate.