Three Meta ads systems are often discussed together in 2026, but their status differs. GEM and ARM have published engineering descriptions and production use; Hierarchical Interest Representation (HIR) is still described as research as of September 2026. This guide draws practical tests from Meta's publications and credits the original industry commentary by Ziming without presenting its proposed rules as official policy.
What do GEM, ARM and HIR do?
| System | Meta's description | Practical limit |
|---|---|---|
| GEM | An ads foundation model that transfers learning to downstream recommenders | Cross-surface learning does not imply advertisers share private conversion data |
| ARM | An online ranking system that allocates computation by request | No evidence that a specific campaign structure automatically earns more compute |
| HIR | Research into hierarchical representations of people, advertisers and products | Not a confirmed, universally deployed targeting feature |
Meta's reported ARM gains concern selected target users, not guaranteed results for every account. Its HIR article describes an intended integration with existing systems, so test any related tactic before scaling it.
Test creative concepts before executions
Write down the category, use case, purchase barrier and proof behind each concept. Test two genuinely different promises, such as saving time versus reducing learning effort, before testing a new actor or video opening. This four-part brief is an editorial planning framework, not a Meta targeting taxonomy.
Keep optimization event, attribution window and key audience boundaries comparable. Track CTR, conversion rate, CPA and a mature business outcome. The number of concepts should reflect budget and event volume rather than a universal quota.
Preserve useful structure and trustworthy signals
Excessive fragmentation can make each ad set's signals sparse, but necessary country, language, product, budget and compliance boundaries should remain. Check Pixel or SDK events, CAPI mapping, deduplication, value parameters and attribution settings before explaining performance as an algorithm effect. Label standard and incremental attribution separately; their conversion and ROAS figures should not be pooled into one creative ranking.
For app campaigns, use the Meta App launch checklist to validate events, then the event quality review to inspect downstream outcomes.
Diagnose fatigue with a test
Rising frequency and falling CTR alone do not prove audience exhaustion. Check auctions, budget, placement, landing page and measurement first. Under comparable conditions, try a new execution of the same concept. If that fails, test a different need or promise. Treat the result as evidence from a test, not a diagnosis from one chart.
Six checks for media teams
- Define the need and objection addressed by each ad.
- Remove splits that add no measurement value.
- Validate events, deduplication and values.
- Compare concepts before visual executions.
- Use controlled tests to investigate creative fatigue.
- Keep ad promises, product data and landing pages aligned.
These are editorial recommendations, not six official Meta algorithm rules.
Sources
- Meta Engineering: GEM
- Meta Engineering: ARM
- Meta Engineering: HIR research
- Industry commentary: Ziming, Meta ads algorithm analysis (2026), reposted by Xiaowei. This guide is independently rewritten.