What happens if cannibalization is left uncorrected?

Outcomes advertising did not create are still counted as paid results, so ROAS reads higher than reality.

What does it take to confirm cannibalization?

Observational signals only establish correlation. A holdout or an on/off test is what lets you state it causally.

How do you check whether ads cannibalized organic?

Estimate how many paid installs would have arrived organically, then correct ROAS.

4 Analysis · 4-3

Organic vs Paid Cannibalization Analysis

Some installs attributed to paid ads would have arrived organically anyway. Without measuring and correcting for this, ROAS is overestimated.

Methods · GeoX / HoldoutRun quarterly

Summary

Cannibalization is paid advertising eating into organic installs. The two standard ways to measure true incremental installs are (1) Geo Holdout experiments and (2) Causal Impact modeling. Large ad channels need an incrementality measurement once a quarter.

§1 Cannibalization measurement model

The share of observed paid installs that is truly incremental: iROAS = incremental_revenue / cost. The wider the gap between regular ROAS and iROAS, the higher the cannibalization share.

Metric↕	Definition↕	How measured↕
Apparent ROAS	ROAS from network attribution	Regular Adjust reports
Incremental ROAS	Only revenue that wouldn't exist without the ad	GeoX or Holdout
Carbon Tax	Cannibalization share	1 - (iROAS / ROAS)

§2 GeoX (Geographic Experiment)

Split countries/regions into two groups (Treatment / Control), matched on population and revenue share.
Treatment keeps ads running as usual; Control turns the network/campaign OFF.
Run 2–4 weeks, then measure the install/revenue gap between groups.
incremental = treatment - control_baseline

GeoX design example (Meta campaign)

{
  "experiment_name": "meta_aap_geox_2026_q2",
  "duration_days": 28,
  "treatment_geo": ["KR-Seoul", "KR-Busan", "KR-Incheon"],
  "control_geo":   ["KR-Daegu", "KR-Daejeon", "KR-Gwangju"],
  "metric_primary":   "d7_revenue_per_user",
  "metric_secondary": "install_count",
  "min_detectable_lift": 0.05,
  "alpha": 0.05,
  "power": 0.80
}

§3 Holdout (Audience / User Holdout)

Where GeoX turns ads off by region, Holdout excludes a share of users from ad exposure. It's the fallback when GeoX is impossible in single-region operations (e.g. Korea only). Meta, Google, and TikTok all offer campaign-level holdouts.

3.1 Holdout options by network

Network↕	Feature name↕	Holdout share↕	Analysis unit↕
Meta	Conversion Lift / Brand Lift	10% / 20% / 30%	User level (FB ID)
Google	Conversion Lift Study	10% / 50%	Google account level
TikTok	Lift Studies	10%–50% flexible	Advertising ID level
AppLovin	Incrementality Test	Campaign-level OFF	Geo + device ID

3.2 Experiment design — power & duration

A holdout group that's too small can't reach statistical significance. In general:

Minimum holdout size: at ~1,000 installs/day, a holdout share of 30%+ is recommended
Duration: at least 4 weeks (to remove day-of-week effects) — observe through D14 revenue
Significance: α = 0.05, power = 0.80, MDE (Minimum Detectable Effect) = 5%
Use each network's built-in power calculators (Meta Conversion Lift, Google Lift Study)

Holdout design example (Meta Conversion Lift)

{
  "study_name": "meta_aap_holdout_2026_q2",
  "objective": "incremental_purchase",
  "test_groups": [
    { "name": "test",    "split": 0.70, "exposure": "ads_on"  },
    { "name": "control", "split": 0.30, "exposure": "ads_off" }
  ],
  "duration_days": 28,
  "primary_metric":   "d7_revenue_per_user",
  "secondary_metric": "purchase_count",
  "minimum_detectable_lift": 0.05,
  "alpha": 0.05,
  "expected_p_value_threshold": 0.05
}

3.3 GeoX vs Holdout — how to choose

Condition↕	Recommended↕	Why↕
Single country/region operation	Holdout	GeoX can't split regions
Multi-country + brand awareness campaigns	GeoX	User-level exposure leaks
Large campaigns, 5000+ installs/day	Holdout	User-level power is sufficient
Small campaigns under 1000 installs/day	GeoX or Causal Impact	Holdout lacks power
Brands with heavy TV/OOH offline media	GeoX	Offline exposure contaminates user level
Goal is per-network incremental comparison	Holdout	Simultaneous per-network experiments possible

§4 Result adjustment (attribution adjustment)

Compute iROAS / ROAS per network (the Lift Factor)
Multiply every network's ROAS by its Lift Factor before comparing (raw ROAS isn't comparable across networks)
Typical industry Lift Factors: Meta UA ≈ 0.7–0.8, Google UAC ≈ 0.6–0.75, ASA Brand ≈ 0.3–0.5 (brand terms cannibalize heavily)
Re-measure Lift Factors every 6 months — traffic quality and market conditions shift

Per-network Lift Factor application

{
  "as_of": "2026-05-01",
  "media_lift_factors": {
    "Meta_UA":     { "apparent_roas": 1.85, "lift": 0.75, "iroas": 1.39 },
    "Google_UAC":  { "apparent_roas": 2.10, "lift": 0.68, "iroas": 1.43 },
    "TikTok":      { "apparent_roas": 1.60, "lift": 0.82, "iroas": 1.31 },
    "ASA_Brand":   { "apparent_roas": 4.20, "lift": 0.35, "iroas": 1.47 },
    "ASA_Generic": { "apparent_roas": 1.95, "lift": 0.78, "iroas": 1.52 }
  },
  "decision": "Scale ASA_Generic first; hold ASA_Brand at current share"
}

§5 Troubleshooting

Organic in the Control group is also affected

Brand awareness campaigns leak into other regions (TV, out-of-home). GeoX only isolates digital channels. Mass brand campaigns need a separate holdout window (nationwide OFF).

Results are at noise level

Below roughly 5,000 installs/week or $50,000+/week revenue, significance is hard to reach. For small campaigns use Causal Impact (time-series modeling) instead of GeoX.

Revenue in the holdout group is normal

Don't misread this — turning ads off doesn't zero out revenue. Organic, returning users, and other networks keep baseline revenue flowing. The true effect is the gap: incremental = test_revenue − control_revenue.

How to feed lift results back into bidding

Reallocate budget by iROAS. Never fully kill a low-Lift-Factor network (e.g. ASA Brand at 0.35) — brand keywords also block competitor intrusion, a value beyond ROI. Adjust the share instead.

Loading…

Related resources