Meta’s ad delivery now depends on a model trained to predict what a person will do next, more than on the audience you describe. That model is GEM, Meta’s Generative Ads Recommendation Model, and Meta says it is already lifting ad conversions. For a home services company, it changes which inputs actually move results: the signals you send back, the range of creative you supply, and the volume of conversion data the system can learn from.
We covered the retrieval side of this shift in How Meta Andromeda Is Changing Paid Social for Home Service Businesses. This piece picks up where that one stops: what GEM does, what “predictive intent” means in practice, and what an operator should change.
What GEM is, according to Meta
Meta’s engineering team published details on GEM in November 2025. Meta calls it “the largest foundation model for recommendation systems in the industry,” trained across thousands of GPUs with techniques borrowed from large language models.
A few details matter for advertisers:
- GEM learns from ad content and from engagement with both ads and organic posts. Meta lists “advertiser goals, creative formats, measurement signals, and user behaviors” among its training data.
- It reads long behavior histories, sequences Meta describes as “up to thousands of events.”
- It does not serve ads directly. It acts as a teacher, passing what it learns to hundreds of smaller models that make delivery decisions.
Meta reported that GEM drove a 5% increase in ad conversions on Instagram and a 3% increase on Facebook Feed in Q2 2025. Those are Meta’s platform-wide figures, not a forecast for any single account.
What “predictive intent” actually means
Predictive intent is a working label, not a Meta product name. It describes the goal of models like GEM: estimating how likely a specific person is to take a specific action, based on what they have done, before they search for anything.
Search advertising waits for someone to type “roof leak repair.” Meta’s systems try to find the homeowner whose recent behavior resembles the pattern that came before past roof repair inquiries. The model is making a forecast, and the forecast depends on what it has learned from advertisers like you.
That is the operator’s lever. Meta names measurement signals and advertiser goals as GEM training data. Your account supplies both.
Input one: the signals you send back
A prediction model learns what success looks like from the conversion events you report. Optimize for a form submission, and the system gets very good at finding people who submit forms. For a home services business, that is not the same as finding people who book a job.
- Send conversions server-side through the Conversions API as well as the Pixel, so fewer events are lost to browser restrictions.
- Optimize toward the event closest to revenue that still happens often enough to learn from. A qualified call or booked appointment tells the model more than a page view.
- Pass outcomes back from your CRM when an inquiry turns into a sold job, so the model sees which people became customers.
- Keep phone calls in the picture. Many homeowners call instead of filling out a form, and call tracking that reports back to Meta closes that gap.
Bad signals are worse than few signals. If events fire twice or count spam submissions, the model learns the wrong pattern and finds more of it.
Input two: creative range
A model that predicts individual behavior needs options to match to individual people. Retrieval narrows the candidate ads, and GEM-trained models rank what remains. One ad in rotation gives the whole system almost nothing to choose from.
Meta lists “creative formats” among GEM’s training inputs and names “creative representation” as one of the ad attributes its models read. Our reading: the system pays close attention to what your ads contain. Distinct concepts, such as an emergency repair ad, a planned replacement ad, a customer story, and a crew walkthrough, each give the model a different way to connect with a different homeowner.
Volume alone does not help. Twelve near-copies of one image add little. Range does.
Input three: conversion volume
Prediction improves with examples. An account that reports a handful of conversions a month gives the model very little to learn from, however clean the tracking.
For smaller home services accounts, that argues for consolidation. Fewer campaigns and ad sets mean each one collects more conversion events. It also argues for starting with an optimization event that happens often enough, then graduating to a deeper one, such as a booked appointment instead of an inquiry, once volume supports it. That is the Signal-First approach: validate with leading indicators before you scale media investment.
What stays the same
GEM changes how Meta finds people. It does not change what makes a homeowner pick up the phone. Your ad still has to name a real problem. Your landing page still has to load fast and make contact easy. Someone still has to answer when the phone rings. A better prediction model sends you more of the right people, and it cannot fix what happens after they arrive.
A quick audit for your account
Before changing anything else, answer these:
- What conversion event is each campaign optimizing for, and how close is it to a booked job?
- Is the Conversions API active, and do event counts match what your CRM shows?
- How many genuinely distinct creative concepts are live right now?
- How many conversion events does each ad set collect per week?
- Are phone calls from ads being counted at all?
The answers show you which input is starving the model. Fix that one first.
Frequently asked questions
What is Meta’s GEM model?
GEM is Meta’s Generative Ads Recommendation Model, a large foundation model Meta described publicly in November 2025. It learns from ad content, engagement, advertiser goals, and measurement signals, then passes that learning to the smaller models that decide which ads people see.
How is GEM different from Meta Andromeda?
Andromeda handles retrieval, narrowing tens of millions of possible ads to a few thousand candidates. GEM is a foundation model that trains the models downstream. Both reward advertisers who supply distinct creative and accurate conversion data.
What does predictive intent mean for home service leads?
It means Meta estimates who is likely to act before they search. The quality of that estimate depends on the conversion events you report, so optimize toward qualified calls and booked appointments, not only form submissions.
Do I still need detailed targeting on Meta?
Set your real service area and necessary exclusions. Past that, narrow targeting usually limits the model more than it helps. Your creative and conversion data do more to steer delivery than interest selections.
How much creative does a home services account need for GEM and Andromeda?
Enough distinct concepts to cover the main reasons customers call you: emergencies, planned replacements, maintenance, and proof from past customers. Range matters more than count, and near-identical versions add little.
Need a clearer view of what is actually driving growth? ajile MEDIA helps home service businesses connect marketing activity to qualified opportunities, booked calls, search visibility, and revenue influence. Book a strategy call to see where your marketing is producing signals, and where it is only producing noise.
