Amazon Review Intelligence Extractor

Amazon Review Intelligence Extractor — 11 Dimensions, 1B+ Reviews

Pre-analyzed consumer insights. Pain points, buying factors, user profiles, differentiation gaps.

Files

• Script: {skill_base_dir}/scripts/zoodata.py — run --help for params
• Reference: {skill_base_dir}/references/reference.md (field names & response structure)

Credential

Required: ZOODATA_API_KEY. Get free key at zoodata.ai/api-keys

Input (one of)

• Single ASIN: "Analyze reviews for B09V3KXJPB"
• Multi-ASIN: "Compare review pain points across these 5 competitor ASINs"
• Category-wide: keyword/category name → resolve via categories first (need ≥3-level deep path)

API Pitfalls (see zoodata skill for full list)

• reviews/analysis needs 50+ reviews. Fallback chain when sample is insufficient:

1. Lightweight: realtime/product ratingBreakdown — only star distribution, no themes
2. Full 11-dim insights: see "Insufficient Data Fallback" section below — use the

local toolkit (reviews-raw + review-tag-prompt + review-reduce-prompt + review-aggregate) to bypass /reviews/analysis entirely

• labelType is NOT an API request parameter — the API returns all 11 dimensions in one call. Filter by labelType client-side from the consumerInsights array.
• Category mode needs precise path (≥3 levels) — broad categories = diluted insights
• Field name is reviewRate (NOT the legacy reviewPercentage from API v1) for mention frequency
• ASIN-specific endpoints don't need --category; keyword-based ones do
• Category auto-detection: categoryPath is auto-detected from target ASIN. If category_source in output is inferred_from_search, confirm with user

On Missing Key

When ZOODATA_API_KEY is not set (verify via python {skill_base_dir}/scripts/zoodata.py check — exits 2 if no key in env or ~/.zoodata/config.json): follow the "On Missing Key" protocol in zoodata/SKILL.md — STOP before any call, link the user to https://zoodata.ai/en/api-keys, and DO NOT produce a "partial analysis from public knowledge" / "for reference only" fallback as a substitute.

On 401 Invalid Key

When zoodata.py returns code 401: follow the "On 401 Invalid Key" protocol in zoodata/SKILL.md — STOP further calls, tell the user the key was rejected and direct them to api-keys, do not fabricate missing data.

On 402 Credit Exhausted

When zoodata.py returns code 402: follow the "On 402 Credit Exhausted" protocol in zoodata/SKILL.md — STOP further calls, report partial findings already gathered, do not fabricate missing data.

11 Analysis Dimensions

painPoints · issues · positives · improvements · buyingFactors · keywords · userProfiles · scenarios · usageTimes · usageLocations · behaviors

Unique Logic

Analysis Modes

• Category mode: all reviews in category → market-level insights
• ASIN mode: specific products → competitive analysis
• Choose based on user intent. Category = broader, ASIN = deeper.

Pain Point Impact Ranking

Rank differentiation opportunities by: frequency × avg rating delta "Top pain point: durability — mentioned in 27/471 reviews (5.7%), avg rating 2.4 when mentioned"

| reviewRate | Frequency Level | Interpretation | |------------|----------------|---------------| | >10% | 🔴 Critical | Mentioned by 1 in 10 buyers — must address in product design 📊 | | 5-10% | 🟡 Significant | Common complaint — differentiator if solved 📊 | | 2-5% | 🟠 Notable | Worth mentioning in listing if you solve it 📊 | | 3.5 | Mild — noticed but not deal-breaker 🔍 |

Differentiation Priority = High frequency + Low avgRating = Biggest opportunity 🔍. If top 3 pain points all have reviewRate >5% and avgRating 5%) should appear in title or first bullet 💡.

Competitor Comparison

Align dimensions (pain points vs pain points) across products. If competitor review data unavailable, use brand-detail sampleProducts + note limitation.

• Your pain point rate competitor's: Risk — address in product iteration 💡
• Both high on same pain point: Category-wide issue — solving it is a strong differentiator 🔍

Composite Command

python3 {skill_base_dir}/scripts/zoodata.py review-deepdive --target-asin "" [--keyword ""] [--category ""]

Optional: --comp-asins "," for comparison. Runs: reviews × 11 dimensions + competitors + realtime + market context + price/trend. (`` are LITERAL — replace with actual values, no curly braces in commands.)

Detecting when to fall back

After review-deepdive runs, programmatically inspect its JSON output for sparse review aggregation. The composite continues past review failures, so success is NOT proof of usable insights. Detection rule:

import json
deepdive = json.load(open("deepdive.json"))
reviews_section = deepdive.get("reviews", {})
# review-deepdive currently makes one /reviews/analysis call per labelType (target_painPoints,
# target_positives, etc.). All-fail means switch to fallback.
all_failed = all(
    sub.get("success") is False
    or not (sub.get("data") or {}).get("consumerInsights")
    for sub in reviews_section.values()
    if isinstance(sub, dict)
)
target_review_count = (deepdive.get("target_realtime", {}) or {}).get("data", {}).get("ratingCount", 0)
# Fallback triggers if ANY of these are true:
needs_fallback = (
    all_failed
    or target_review_count _/` containing `raw.json`,
`tagged.json`, `clusters.json`, `insights.json`. Example:
```bash
WORK=/tmp/review_B0XXXXXXXX_$(date +%s) && mkdir -p $WORK

Step 1 — Fetch raw reviews

python3 {skill_base_dir}/scripts/zoodata.py reviews-raw \
    --asin  [--marketplace US] [--max-pages 10] > $WORK/raw.json
# Cost: 1 credit/page, 10 reviews/page, hard cap 100 (10 pages).
# Stops automatically when nextCursor=null (small-volume ASINs may exhaust earlier).
# For cost control: --max-pages 5 = 50 reviews / 5 credits / ~30s.

Then save just the reviews array for downstream tooling:

python3 -c "import json,sys; d=json.load(open('$WORK/raw.json'))['data']['reviews']; json.dump(d,open('$WORK/reviews_array.json','w'),ensure_ascii=False)"

Step 2 — Map (per-review tagging)

review-tag-prompt renders the prompt but does NOT call any LLM — YOU (this skill's LLM) produce the JSON. The template is uniform per review, so render it ONCE to learn the schema, then mass-produce tags for all reviews in a single in-context pass.

# Render the prompt for ONE review to learn the schema (do this once per skill run)
python3 {skill_base_dir}/scripts/zoodata.py review-tag-prompt \
    --review "$(python3 -c 'import json,sys; print(json.dumps(json.load(open(sys.argv[1]))[0]))' $WORK/reviews_array.json)" \
    [--product-title "..."] [--product-category "..."]

The schema you must produce per review (12 fields, all required, empty arrays for empties):

{
  "sentiment": "positive|neutral|negative",
  "mentioned_scenarios": [], "mentioned_issues": [], "mentioned_positives": [],
  "mentioned_improvements": [], "mentioned_buying_factors": [],
  "mentioned_pain_points": [], "user_profiles": [],
  "mentioned_usage_times": [], "mentioned_usage_locations": [],
  "mentioned_behaviors": [], "keywords": []
}

After producing tags for all N reviews, save them as a JSON array preserving review order:

# Save your in-context output as the array (example structure)
cat > $WORK/tagged.json 150 unique candidates, split into
chunks of ~150, render the reduce prompt per chunk, and merge clusters across chunks
by case-insensitive canonical name match. Other dims rarely exceed 100 candidates.

### Step 4 — Aggregate (no LLM, pure local)

```bash
python3 {skill_base_dir}/scripts/zoodata.py review-aggregate \
    --reviews $WORK/raw.json \
    --tagged $WORK/tagged.json \
    --clusters $WORK/clusters.json > $WORK/insights.json
# Output structure matches /reviews/analysis:
#   { reviewCount, avgRating, sentimentDistribution, consumerInsights[], topKeywords[] }
# Each consumerInsight has: {element, labelType, count, reviewRate, avgRating}

Use the same Pain Point Impact Ranking and Differentiation Priority tables above — but apply the small-sample caveat when `reviewCount 15 percentage points, re-examine the Map tags before publishing. Common causes: LLM mis-classifying a 5★ "didn't love it but works" as positive; non-English reviews mis-tagged. Document residual mismatch in Data Provenance.

Language (required)

Output language MUST match the user's input language. If the user asks in Chinese, the entire report is in Chinese. If in English, output in English. Exception: API field names (e.g. monthlySalesFloor, categoryPath), endpoint names, technical terms (e.g. ASIN, BSR, CR10, FBA, credits) remain in English.

Disclaimer (required, at the top of every report)

> Data is based on ZooData API sampling as of [date]. Monthly sales (monthlySalesFloor) are lower-bound estimates. This analysis is for reference only and should not be the sole basis for business decisions. Validate with additional sources before acting.

Confidence Labels (required, tag EVERY conclusion)

• 📊 Data-backed — direct API data (e.g. "painPoint 'durability' mentioned by 27% of reviewers 📊")
• 🔍 Inferred — logical reasoning from data (e.g. "durability is the #1 differentiation opportunity 🔍")
• 💡 Directional — suggestions, predictions, strategy (e.g. "highlight durability in bullet point #1 💡")

Rules: Strategy recommendations and listing copy suggestions are NEVER 📊. User criteria override AI judgment.

Aggregate-label rule (applies to ALL report output, not just fallback): NEVER attach 📊 to ANY element that aggregates or groups underlying content when ANY piece of that content is 🔍 or 💡. "Aggregate/grouping elements" include:

• Section headers at EVERY level (#, ##, ###, ####) — including top-level summary sections like "Overall Score", "Verdict", "Executive Summary"
• Summary/score lines anywhere in the report (e.g. ## Overall Score — 27/100 · Grade F 📊 is WRONG if any Basis row inside is 🔍)
• Table column headers in comparison tables (e.g. **Target ASIN** 📊 as a column label is WRONG if any cell in that column contains 🔍)
• Table row headers or row-aggregation labels (when the row aggregates multiple cells of mixed confidence)
• Any other visual grouping label — bullet-list group titles, callout box titles, etc.

A group-level 📊 implies the whole block/column/row is data-backed, which smuggles inferred/directional content into the 📊 tier via visual grouping. Either (a) omit the group-level label entirely (preferred when content mixes tiers), or (b) use the LOWEST confidence present inside (🔍 if any underlying content is 🔍; 💡 if any is 💡). This is a universal output-quality rule — it applies regardless of which fallback path (if any) was triggered.

Emoji reservation rule (closely related): The three confidence symbols 📊 🔍 💡 are RESERVED for confidence labeling. NEVER use them as decorative prefixes on section headers, table headers, or any aggregate element — even when you also include a correct confidence suffix on the same line. Example:

• ❌ WRONG: ## 📊 Overall Score — 27/100 · Grade F 🔍 (the leading 📊 reads as a data-backed claim even though the trailing 🔍 is correct)
• ✅ RIGHT: ## Overall Score — 27/100 · Grade F 🔍 (no decorative emoji, just the proper confidence suffix)
• ✅ RIGHT: ## 🎯 Overall Score — 27/100 · Grade F 🔍 (use non-reserved decorative icons like 🎯 🧭 📋 📝 📂 🏁 🚨 🏆 🔔 when a visual prefix is desired)

Decorative emoji ≠ confidence label — but from a reader's perspective, a leading 📊/🔍/💡 is indistinguishable from a confidence claim. Reserve these three symbols EXCLUSIVELY for confidence annotation to avoid ambiguity.

Data Provenance (required)

Include a table at the end of every report:

| Data | Endpoint | Key Params | Notes | |------|----------|------------|-------| | (e.g. Market Overview) | markets/search | categoryPath, topN=10 | 📊 Top N sampling, sales are lower-bound | | ... | ... | ... | ... |

Extract endpoint and params from _query in JSON output. Add notes: sampling method, T+1 delay, realtime vs DB, minimum review threshold, etc.

API Usage (required)

| Endpoint | Calls | Credits | |----------|-------|---------| | (each endpoint used) | N | N | | Total | N | N |

Extract from meta.creditsConsumed per response. End with Credits remaining: N.

API Budget: ~20-30 credits

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: SerendipityOneInc
• Source: SerendipityOneInc/ZooData-Skills
• License: MIT
• Homepage: https://apiclaw.io/

Install and usage instructions live in the source repository linked above.