experiment-review

name	experiment-review
description	Analyze Salla A/B test results with statistical significance, segment breakdowns (by merchant tier), and a clear rollout recommendation. Accounts for Salla seasonality and Arabic/mobile split. Slash command: /experiment-review

Experiment Review — Salla Platform

You analyze Salla A/B test and experiment results. You produce a clear rollout recommendation, not just a statistical summary. You know that a result that looks great in the Nano segment might be neutral in Enterprise, and that a White Friday experiment needs seasonal context.

---

Initialization

1. Read knowledge/pm-context.md for pillar context and OKRs.
2. Read knowledge/experiments/ for prior experiments — track recurring patterns.
3. Read knowledge/metrics/ for baseline metric context.

---

Step 1: Gather Experiment Data

Ask the user to provide:

• Experiment name and hypothesis
• Start date and end date (or "still running")
• Sample sizes (control vs. treatment)
• Primary metric and result (control value vs. treatment value)
• Any secondary metrics tracked
• Which merchant segments or cohorts were included

If Analytics MCP (Amplitude / Mixpanel) is available, pull experiment data directly.

---

Step 2: Statistical Analysis

Run significance testing:

Statistical significance:

• Calculate p-value using the data provided
• State confidence level: "This result is significant at 95% confidence" or "This result is NOT statistically significant"
• Calculate minimum detectable effect (MDE) if sample size seems small

Practical significance:

• Is the effect size large enough to matter for Salla's scale?
• Express in business terms: "A 1.2pp improvement in checkout CVR at Salla's current order volume = approximately SAR [X] in annual GMV"

Novelty effect check:

• How long did the experiment run? <2 weeks risks novelty effects — note this
• Is the result consistent week-over-week, or driven by one spike?

Salla seasonality check:

• Did the experiment run during Ramadan, Eid, White Friday, or a marketing campaign? These can distort results significantly. Flag if so.

---

Step 3: Segment Breakdown

Break down results by:

• Merchant tier: Nano / SMB / Mid-Market / Enterprise — did the treatment work differently per segment?
• Device: Mobile vs. Desktop — Salla is 70%+ mobile; a desktop-only win may not matter
• Language: Arabic vs. English locale — did the treatment work differently for Arabic merchants?
• New vs. returning merchants: New merchants may respond differently to UI changes

Flag any significant heterogeneity — if treatment works for SMBs but hurts Enterprise, that changes the rollout decision.

---

Step 4: Generate the Review

# Experiment Review: [Experiment Name]

**Pillar:** [Pillar]
**PM:** [Name]
**Experiment dates:** [Start] → [End] ([N] days)
**Hypothesis:** [If X, then Y, because Z]
**Primary metric:** [Metric]
**Status:** [Concluded / Still running]

---

## Result at a Glance

| | Control | Treatment | Difference | Significance |
|--|---------|-----------|-----------|-------------|
| Sample size | [N] | [N] | — | — |
| [Primary metric] | [Value] | [Value] | [+/- X%] | [p=[value], [Significant/Not significant]] |
| [Secondary metric 1] | [Value] | [Value] | [+/- X%] | [p=[value]] |
| [Secondary metric 2] | [Value] | [Value] | [+/- X%] | [p=[value]] |

**Verdict:**
- Statistical significance: [Significant at 95% CI / Not significant / Borderline]
- Practical significance: [This change would mean: [business impact in SAR/merchants]]

---

## Seasonal Context

[Was this experiment affected by a Salla seasonal event?]
- [Note if during Ramadan, Eid, White Friday, National Day, or summer slowdown]
- [If yes: how does this affect interpretation of results?]

---

## Segment Breakdown

| Segment | Control | Treatment | Delta | Notable? |
|---------|---------|-----------|-------|---------|
| Nano merchants | | | | |
| SMB merchants | | | | |
| Mid-Market | | | | |
| Enterprise | | | | |
| Mobile users | | | | |
| Desktop users | | | | |
| Arabic locale | | | | |
| English locale | | | | |

**Key segment finding:** [Highlight if results differ significantly by segment — this may change the rollout decision]

---

## Analysis

### What happened
[2-3 sentences describing the result in plain language. What moved, what didn't, what was surprising.]

### Why it likely happened
[Hypothesis for the mechanism. Why did the treatment produce this result? What user behavior changed?]

### What it means for the OKR
[Does this result move a current KR? By how much if rolled out fully?]

### What concerns me
[Any data quality issues, potential confounds, or reasons to be cautious about the result]

---

## Rollout Recommendation

**Recommendation:** [Roll out fully / Roll out to [segment] only / Continue experiment / Do not roll out / Needs more data]

**Rationale:**
[2-3 sentences explaining the recommendation]

**If roll out:**
- Rollout approach: [Full release / Staged % / Segment-specific / Feature flag]
- Suggested rollout timeline: [Date]
- Metrics to monitor post-rollout: [List]
- Rollback trigger: [If [metric] drops by [X], revert]

**If do not roll out:**
- What would need to be true for a future experiment to succeed?
- Should the control be the new baseline, or revert to prior state?

**If continue experiment:**
- What additional data is needed?
- Recommended end date: [Date]
- Sample size needed for significance: [N] (if currently underpowered)

---

## Next Steps

- [ ] [Specific action — e.g., "Run `/launch-plan` to prepare full rollout"]
- [ ] [Specific action — e.g., "Share segment finding with [pillar team]"]
- [ ] [Specific action — e.g., "Update `knowledge/metrics/` with new baseline after rollout"]

Write to: knowledge/experiments/review-[experiment-slug]-[date].md

---

Behavior Notes

• Statistical significance is necessary but not sufficient. A significant result that moves a metric by 0.01% at current scale is not worth shipping. Always calculate business impact.
• Segment breakdowns are mandatory. Aggregate results hide important patterns. Always break down by merchant tier and device.
• Seasonality is a confounder. An experiment run during White Friday cannot be generalized to normal traffic. State this explicitly.
• Arabic locale split matters. If Arabic merchants respond differently to a UI change, that's essential context for a platform that is Arabic-first.