Meta GEM ads model: 18.9% conversion claim, missing context

Meta said on 13 October 2026 that its Generative Ads Model (GEM) has delivered an 18.9% conversion improvement and a 12% ad click-through rate (CTR) improvement across its ad surfaces and objectives. According to a technical post on Engineering at Meta, the 2026 generation replaces separate prediction systems for Facebook, Instagram and Reels with one central foundation model. Meta did not publish the baseline, the measurement period or the test method behind the two headline figures, so advertisers should read them as Meta's own unaudited claims.

What Meta announced about GEM on 13 October 2026

The post, credited to seven Meta engineers (Huayu Li, Darren Liu, Xiaoyi Liu, Ellie Wen, Jackie Xu, Chunzhi Yang and Reva Srinivasan), describes GEM as "the industry's largest recommendation system (RecSys) foundation model". Meta introduced the first GEM in 2025. The 2026 generation is the subject of this update.

The main change is structural. Meta wrote that the new GEM moves away from per-domain models for Facebook, Instagram and Reels and toward a single central foundation model trained on data from all of Meta's ads platforms. In the old setup, each surface and each prediction objective had its own system. In the new one, one model predicts click-through rate across every surface and conversion rate across the major optimization goals at the same time. Meta calls it the first model to do both jointly.

Meta reported two sets of results:

• An 18.9% conversion improvement and a 12% ad CTR improvement "across Meta's ad surfaces and objectives". The post gives no baseline or time period for these.
• Conversion uplift "for targeted users over the past year" of 5.8% on Facebook, 4.1% on Instagram and 1.6% for off-site conversion. The post does not define "targeted users".

Meta's GEM figures at a glance

The table below lists the figures Meta published on 13 October 2026, with what each one is attached to in the post.

How one central model changes ad delivery

The part of the post that matters most for media buyers is how click data feeds conversion prediction. Meta wrote that "the generalizable representations learned on data-rich click tasks transfer into data-sparse conversion tasks." Clicks are plentiful. Purchases, leads and app installs are rare by comparison, so a model trained only on conversions for one surface has little to learn from. By training one model on everything, Meta says engagement patterns learned on one surface can inform conversion behavior such as purchases on another.

In plain terms: a person's behavior on Instagram Reels can now shape how Meta predicts whether that person will convert from a Facebook Feed ad, because both predictions come from the same model. Meta did not describe this in advertiser terms, and the post does not mention Advantage+ or any Ads Manager setting. The link to campaign setup is an inference, not something Meta stated.

Meta also said GEM uses a "domain expert framework": a shared expert learns patterns common to all surfaces, while domain-specific experts capture what is particular to each surface. So the model is central, but it still keeps some surface-specific behavior.

Kunlun, the new architecture

Meta named the new architecture Kunlun and said it delivers 2x performance over the previous design, Interformer, at the same computation scale, and enables 2.5x growth in model complexity. Kunlun combines attention, a summarization component and a mixture-of-experts module in each layer. One component, CompSkip, skips some sequence modules and cuts FLOPS by 43% with what Meta calls "negligible quality loss". Another, OneSeq, merges different types of user event sequences into one chronological stream.

Longer user histories

Meta said GEM doubled the engagement context length it uses in production and has shown it can scale that context to 25x. The post reports that prediction quality follows a power law with sequence length and that larger models benefit more from longer histories. Meta did not say how long the production history is in days, sessions or events, so the "2x" cannot be translated into anything an advertiser can check.

Why GEM does not serve ads directly

GEM is too large to score every ad request in real time. Instead, according to the 13 October post, its knowledge feeds "hundreds of vertical models" that serve ads live. Meta described three ways that knowledge moves downstream:

• Knowledge distillation, the older method, which Meta called "effective but limited in bandwidth".
• LoopFM, which compresses GEM's internal representations and combines them with the same user's earlier representations. Meta claims 2x transfer effectiveness over knowledge distillation at no extra foundation-model cost or latency.
• Solaris, which precomputes user-ad embeddings in background jobs for likely future requests. If an exact embedding is missing, Solaris falls back to an aggregate of the user's recent embeddings for the same ad groups, or to behaviorally similar users.

The Solaris fallback is worth a moment. When the system lacks a fresh signal for a person, it borrows from similar people.

Training scale

Meta said GEM trains on "several thousand" latest-generation GPUs. Total training compute grew 4x while end-to-end efficiency doubled to 20-25% model FLOPS utilization. Meta did not disclose training cost, energy use or parameter count.

What Meta did not disclose

The post is a first-party engineering blog. It is written by the people who built the system, about their own system, and nobody outside Meta has audited the numbers. Several gaps stand out:

• No baseline for the 18.9% and 12% figures. Improvement over what model, and over what period, is not stated.
• No method. The post does not say whether the figures come from live A/B tests or offline evaluation.
• No definition of "targeted users" for the 5.8%, 4.1% and 1.6% figures, and no explanation of whether those sit inside the 18.9% or on top of it.
• No breakdown by country, industry or campaign objective. Nothing in the post refers to Thailand or Southeast Asia.

The trade publication PPC Land, reviewing the post, also pointed out that Meta's earnings calls have attributed different figures to GEM. PPC Land reported that in July 2026 Meta cited a 15.7% uplift in Facebook conversions credited to GEM and user understanding models together, against the 5.8% Facebook figure in the new post, which the post does not reconcile. PPC Land also noted that the 1.6% off-site figure matches a lift Meta earlier credited to a different model. These are PPC Land's observations; the Meta post itself does not address them.

None of this means the gains are false. It means an advertiser cannot assume an 18.9% conversion lift will show up in their own account. Meta's numbers are measured by Meta, against a baseline Meta has not shared.

What advertisers running Meta ads should check

The checklist below is analysis, not guidance from Meta. The post describes no Ads Manager setting for GEM. What an advertiser controls is the signal the model learns from and the room it is given to deliver.

1. Conversion signal quality

Meta's post says click-rich data helps the model predict data-sparse conversions. The model still needs real conversion events to learn from. Check that the Meta Pixel and Conversions API both send the events you optimize for, that purchase or lead events are deduplicated, and that event match quality in Events Manager is not low. A model that generalizes well from clicks will still optimize toward the wrong outcome if the conversion event fires on a thank-you page that people reach without buying.

2. Placement restrictions

If one model now learns across Facebook, Instagram and Reels, campaigns locked to a single placement give it less room. That does not mean every campaign should run on every placement. Brand-safety rules, creative formats and past results in your own account are good reasons to restrict. But if a campaign was restricted by habit rather than evidence, a controlled test against Advantage+ placements is a fair thing to run now.

3. Creative built for each surface

Broad placements only help if the ad works where it lands. A static square image shown in Reels is unlikely to perform well no matter how good the prediction model is. Supplying vertical video for Reels and Stories alongside feed formats gives the system usable options on each surface.

4. Measure lift, not only platform-reported results

Meta's 12% CTR claim is measured by Meta. Clicks and attributed conversions inside Ads Manager do not prove incremental sales. Compare Meta-reported conversions against your CRM, Shopify, LINE OA or GA4 data, and use a holdout or conversion lift study where budget allows. If Meta-reported conversions rise while backend sales stay flat, the model may be finding people who would have bought anyway.

5. Do not change everything at once

The GEM changes happened on Meta's side over the past year. If your results moved during that time, it is tempting to credit or blame the model. Without a controlled comparison, you cannot separate GEM from seasonality, creative fatigue, auction pressure or your own edits.

What this means for Thai marketers

Meta's post does not mention Thailand, and it gives no market-level results. GEM is described as a central model trained on data from all of Meta's ads platforms, which suggests Thai campaigns run through the same system, but whether Thai advertisers see gains near the headline figures is unknown. The practical work for Thai advertisers is the same as above: clean conversion events through the Conversions API, creative for each placement, and a backend check of what Meta reports. Teams that want help setting this up can look at a managed paid social program, and the GA4 setup service covers the backend measurement side.

FAQ

What is Meta's GEM?

GEM is Meta's Generative Ads Model, which Meta describes as the industry's largest recommendation system foundation model. Meta introduced it in 2025, and the 2026 version described on 13 October 2026 uses one central model to predict clicks and conversions across Facebook, Instagram, Reels and other surfaces.

Did GEM really raise conversions by 18.9%?

Meta says so, but the figure is not independently verified. Meta did not state the baseline, the time period or whether the result came from live A/B tests or offline evaluation, so it should be treated as a vendor claim.

Can I turn GEM on or off in Ads Manager?

The post describes no advertiser setting for GEM. It runs inside Meta's delivery system, so the levers available to advertisers are conversion data quality, placement choices, creative and budget.

Does GEM affect ads in Thailand?

Meta did not say anything specific about Thailand. The post describes GEM as a central model trained on data from all of Meta's ads platforms, which suggests it applies broadly, but Meta published no country-level results.

Should I switch every campaign to Advantage+ placements?

Not automatically; the Meta post does not mention Advantage+ at all. It is reasonable to test broader placements against restricted ones where restrictions were not based on evidence, and to judge the result on backend sales rather than platform-reported conversions alone.

Meta's 13 October post explains how its ads model now learns across surfaces, even if the headline numbers come without the context needed to verify them. The useful response is to give the model clean conversion data and to measure results outside Meta's own reporting. If you want a second pair of eyes on your Meta setup, Relevant Audience's Facebook ads team can review it with you.