Developing
Vendor-reported, unaudited figures; the post says its scaling chart uses a lightweight model, not production GEM.
In 60 seconds
- Meta's Engineering blog (October 13, 2026) claims an 18.9% conversion lift and 12% click-through lift across ad surfaces and objectives.
- Over the past year GEM added 5.8% conversion uplift on Facebook, 4.1% on Instagram and 1.6% on off-site conversions "for targeted users".
- A single central model replaced separate per-surface, per-objective predictors, using a new Kunlun architecture.
- The post gives no baseline, measurement period or test design; its Facebook figure differs from the 15.7% Meta cited in July.
Impact Layer
- What changed
- Meta published its third long technical account of GEM, with new performance claims.
- Why it matters
- Ads ranking models decide which ads win auctions, so advertisers want to know whether reported gains reach their campaigns.
- Who is affected
- Meta advertisers, agencies and measurement teams.
- Martech stack area
- AI, Advertising
- Stack impact
- Do not assume an 18.9% lift in your account. Compare your own conversion and CPA trends over the same weeks and use holdouts to check.
- Business impact
- Moderate Company claims relayed by PPC Land. There is no baseline, period, A/B or offline method, and no statement of how the 5.8%, 4.1% and 1.6% figures relate to 18.9%.
- Related technologies
- What to watch
- Independent tests of Meta conversion lift and the next earnings call's GEM figures.
Meta published a technical post on its Engineering blog on October 13, 2026 about GEM, which it calls “the industry’s largest recommendation system foundation model”. Seven engineers are named as authors. PPC Land covered it the next day.
The claims
- An 18.9% conversion improvement and a 12% click-through improvement across ad surfaces and objectives.
- Over the past year, GEM added 5.8% conversion uplift on Facebook, 4.1% on Instagram and 1.6% on off-site conversions, “for targeted users”, a group the post does not define.
What changed technically
- Separate prediction systems for each surface and objective were replaced by one central model predicting click-through across surfaces and conversion rates across objectives.
- A new architecture, Kunlun, replaces Interformer; Meta says it gives “2x performance gains with the same computation scaling”.
- Training uses “several thousand” GPUs, and Meta says model FLOPs utilisation reached 20-25%, doubling over 12 months.
What is missing
The baseline, measurement period, whether results come from A/B tests or offline evaluation, and how the platform figures add up. The 5.8% Facebook figure also differs from the 15.7% Meta credited to GEM and user-understanding models on its July earnings call. The post’s scaling chart uses a “light-weight model”, not production GEM.
Our read
This is a description of how Meta says it builds ranking, not a test you can reproduce. Judge it by your own holdout results.
Sources
- PPC Land: Meta says its GEM ads model lifted conversions 18.9% across its apps (opens in a new tab)
Source for the post's claims, architecture, training figures, caveats and earlier GEM numbers.