# How to read a Meta lift study without fooling yourself

> Tier: deep. Full contract: quantified decision rules and evidence requirements.

What the conversion lift or brand lift study actually proved, at what confidence, and what it licenses you to decide, in plain language.

**Intent.** Turn a lift study into a decision-grade readout: effect, confidence, caveats, and the decisions it does and does not support.

## When to use

- 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.

## When not to use

- Mid-study: interim peeking inflates false positives; wait for completion.
- As a substitute for platform attribution reconciliation: lift answers incrementality, not reporting truth.

## Sources

Required: Meta Ads

## 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

1. List studies and pick the target. — tools: `meta_get_ad_studies`
2. Pull the results: cells, reach, conversions, lift, confidence. — tools: `meta_get_study_results`
3. Run the structured interpretation over the results. — tools: `meta_interpret_experiment_results`
4. Translate into the decision language: what this licenses, what it does not.

## Decision rules

- **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 requirements

- Every claim carries the cell sizes, the effect, and the interval.
- The readout distinguishes platform-attributed conversions from incremental ones explicitly.

## Human approval boundary

- 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.

## Failure modes

- 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.

## Output

- **Verdict**: Effect, interval, power, in one paragraph a CMO can quote.
- **Licensed decisions**: What this evidence supports doing.
- **Not licensed**: The extrapolations it does not support, said before someone makes them.

## Related

- Skill: [cross-platform-performance](https://www.getmcpads.com/skills/cross-platform-performance.md)
- Skill: [tracking-audit](https://www.getmcpads.com/skills/tracking-audit.md)
- Guide: [cross-platform-report](https://www.getmcpads.com/guides/cross-platform-report.md)