Deep contractMeta Ads
How to read a Meta lift study without fooling yourself
What the conversion lift or brand lift study actually proved, at what confidence, and what it licenses you to decide, in plain language.
Contract v1.0.0last verified reviewed
IntentTurn a lift study into a decision-grade readout: effect, confidence, caveats, and the decisions it does and does not support.
- 4
- Steps
- 3
- Tools
- 3
- Rules
- 3
- Failure modes
When to use it
Use it when
- A lift study just completed.
- Someone quotes a lift number in a budget argument and the source needs checking.
- Planning the next test: reading past ones first.
Do not use it when
- Mid-study: interim peeking inflates false positives; wait for completion.
- As a substitute for platform attribution reconciliation: lift answers incrementality, not reporting truth.
What it needs first
Required
Preconditions
- At least one study exists on the ticked account.
Inputs
study- The study to read, or the latest completed one.Default: latest completed.
Procedure
In order. Every tool named is one the gateway ships, and links to its reference.
List studies and pick the target.
Pull the results: cells, reach, conversions, lift, confidence.
Run the structured interpretation over the results.
Translate into the decision language: what this licenses, what it does not.
Decision rules
The observable condition, what it lets you conclude, and what takes the conclusion back.
When
The confidence interval on lift crosses zero
Conclude
The study is inconclusive, not negative: report power, not failure, and size what a conclusive rerun needs.
When
Lift is significant but the absolute incremental volume is small
Conclude
Statistically real, commercially minor: say both, because budget arguments need the second.
When
The test cells were contaminated (audience changes mid-study, overlapping campaigns)
Conclude
Quote the contamination and refuse a causal readout.
Evidence every conclusion must carry
- Every claim carries the cell sizes, the effect, and the interval.
- The readout distinguishes platform-attributed conversions from incremental ones explicitly.
Where the agent stops
- Budget decisions taken on the readout are human; the skill states what the evidence licenses.
- Designing the next study is proposed, never launched: study setup is not exposed as an agent capability.
How it goes wrong quietly
The cases where the analysis is wrong and still looks right. Read them before trusting a number.
- Underpowered studies read as 'no effect' when they measured nothing; power first, verdict second.
- Lift on the platform's own conversion event inherits that event's tracking gaps; a tracking-audit finding invalidates a lift readout.
- One geography or season does not generalise; the readout names its scope.
What the answer contains
- 01Verdict
- Effect, interval, power, in one paragraph a CMO can quote.
- 02Licensed decisions
- What this evidence supports doing.
- 03Not licensed
- The extrapolations it does not support, said before someone makes them.