review-mining

name

review-mining

description

Deep review extraction for an Amazon ASIN. Pulls multi-page reviews (1★/3★/5★ + recent verified) via SellerApp n8n MCP, then decomposes them into purchase drivers, objections, surprises, recurring questions, comparison mentions, and photo references — each tagged by frequency. Use when running `review-mining {ASIN}`, when an existing brief needs deeper review evidence, or when surfacing objections for a CVR-leak fix.

review-mining — Claim-Level Review Decomposition

Reviews are the single most important CRO signal source. This skill pulls a wide-and-deep review sample, decomposes it using the 8-layer framework from ~/.claude/skills/cro/review-analysis.md, and outputs structured insights ready to feed /cro-content-plan, objection-killer, or any image/copy planning skill.

Invocation

review-mining {ASIN}
review-mining {ASIN} {marketplace}
review-mining {ASIN} --deep   # 5 pages per star bucket instead of 3

Output

Writes to /tmp/cro-research/{ASIN}-reviews-{date}.md and opens.

Tools Used

n8n MCP wrapper (workflow 9RmjDT107uXtrImf). Tool prefix mcp__f32016b6-7c77-45e2-b4a2-70195c5f2d2d__*.

Pull	Tool	Params
Verified positive	Get_Product_Reviews	rating=5, reviewer_type=verified_purchase, pages 1-3
Negative	Get_Product_Reviews	rating=1, pages 1-3
Critical-but-buying	Get_Product_Reviews	rating=3, pages 1-2
Nuanced	Get_Product_Reviews	rating=4, page 1
Recency baseline	Get_Product_Reviews	sort=recent, page 1
With media	Get_Product_Reviews	media_type=multimedia, page 1 (photo refs)

Phase 1 — Pull

Parallel batch where possible. Default = 3 pages per star bucket = ~30 reviews per bucket. --deep flag = 5 pages = ~50 per bucket.

Note: each review page costs 10 SellerApp tokens. Default mode = ~120 tokens per ASIN.

Phase 2 — Decompose (8-layer framework from cro/review-analysis.md)

For every review, extract:

1. Aspect — what part of the product (taste, durability, sizing, packaging, instructions, etc.)
2. Sentiment — positive / negative / neutral (with negation handling)
3. Driver vs Barrier — does this push them to buy, or stop them?
4. Modifier — qualifier ("loud at first" / "small for the price")
5. Conditional — when does it apply? ("if you have small hands" / "for the price")
6. Repurchase signal — explicit ("would buy again") / implicit (gifting, replacing)
7. Competitor mention — did they compare to anyone?
8. Photo evidence — did they post a photo? What does it show (demographic / use context)?

Phase 3 — Categorize Output

Aggregate decomposed claims into 6 buckets per 02-visual-content/listing-images.md rules:

A. Purchase Drivers (from 5★ + 4★)

Top 5-10 themes ranked by frequency. Each entry:

• Theme
• Frequency: N/{total} reviews = X%
• Tag: 🔴 >30% mandatory visual / 🟡 10-30% supporting / ⚪ <10% nice-to-have
• Quote (one short example, fewer than 15 words, in quotes)

B. Objections (from 1-3★)

Top 5-10 themes. Same structure. Tag: 🔴 >20% of negatives preempt with image / 🟡 <20% address in A+ FAQ.

C. Positive Surprises (from 5★)

Things customers love that nobody pitched. Untapped selling points → candidates for slot 7 or A+ hero.

D. Recurring Questions / Confusions

Cross-cuts all star levels. "Wait, does it…?" / "I wasn't sure if…" → these are visual-FAQ candidates.

E. Competitor Mentions

Compared to whom, on what dimension, who won. Builds the differentiation column for competitor-sweep.

F. Photo Review Evidence

• Demographic distribution (age/gender/race/setting visible in customer photos)
• Use context (where they use it, when, with what)
• Defects shown (if any)
• This is the ground truth for "Right Demographic" rule — beats stock photo intuition.

Phase 4 — Write Output File

# Review Mining — {Title}

**ASIN:** {ASIN} | **Date:** {date} | **Sample size:** {N reviews across {breakdown}}

## Data Sources

| Bucket | Reviews pulled | Token cost |
|---|---|---|
| 5★ verified | {N} | 30 |
| 4★ | {N} | 10 |
| 3★ | {N} | 20 |
| 1★ | {N} | 30 |
| Recent | {N} | 10 |
| Media | {N} | 10 |
| **Total** | **{N}** | **~120** |

## A. Purchase Drivers
{table with theme, frequency, tag, quote}

## B. Objections
{table}

## C. Positive Surprises
{list}

## D. Recurring Questions
{list, each tagged: → image / → A+ FAQ / → bullet}

## E. Competitor Mentions
{table: competitor, dimension, who won, frequency}

## F. Photo Review Evidence
{demographic summary + use-context summary + 3-5 representative photo URLs}

## Slot Mapping (handoff to content planning)

| Driver / Objection | Slot recommendation | Image type |
|---|---|---|
| {Top driver} | Slot 2 | Benefit photography |
| {Top objection} | Slot 3 | Objection-handling visual |
| ... | ... | ... |

## Sentiment Trajectory

{Pull review_date from each — flag if recent reviews trend more negative; that's a regression signal}

Reference Files

• ~/.claude/skills/cro/review-analysis.md — full 8-layer framework + validation rules
• Vault: CRO-Knowledge-Base/01-research/review-mining.md
• Vault: CRO-Knowledge-Base/MASTER-CRO-REFERENCE.md §2 (Review Mining is signal #1 of 4)

Quality Bar

• Sample includes all of: 1★, 3★, 5★, recent, media (no bucket missing)
• Every theme in A & B has a frequency count, not just a list
• Frequency tags applied (🔴 / 🟡 / ⚪)
• Quotes fewer than 15 words, in quotation marks (copyright)
• Slot mapping table is filled — this is the handoff to image planning
• Sentiment trajectory checked (regressed?)

Common Mistakes (from review-analysis.md validation rules)

• ❌ Treating "loud at first but you get used to it" as pure negative — it's conditional
• ❌ Counting "doesn't have X" feature requests as objections to existing features
• ❌ Ignoring 5★ reviews because "they're all positive" — they have the purchase drivers
• ❌ Missing comparison mentions ("better than {Brand}") — these are gold for differentiation
• ❌ Not weighting verified purchases higher than non-verified

Auto-Triggers

Runs (without prefix) when:

• User asks "what do reviews say about {ASIN}" / "mine reviews for X"
• A pipeline (/cro-content-plan, objection-killer, cvr-leak-fix) needs the data
• User uploads a review export and asks for analysis (skill switches to file-input mode)