amazon-sp-sqp-asin-ngram

• CHANGELOG.mddoc · 4,846 bytes
• LICENSE.txtdoc · 443 bytes
• SKILL.mdcore · 7,727 bytes
• SKILL.mdcore · 7,727 bytes
• SKILL.mdcore · 7,727 bytes
• evals/evals.jsonasset · 81,691 bytes
• evals/fixtures/seller-language.jsonasset · 15,319 bytes
• evals/fixtures/theme-decisions.jsonasset · 4,364 bytes
• references/amazon-sources.mddoc · 3,705 bytes
• references/asin-rules.mddoc · 2,675 bytes
• references/decision-records.mddoc · 9,423 bytes
• references/opportunity-score-and-qa.mddoc · 4,495 bytes
• references/relevance-and-intent.mddoc · 5,911 bytes
• references/sqp-metrics.mddoc · 5,682 bytes
• references/theme-analysis.mddoc · 7,133 bytes
• scripts/.gitkeepruns · 0 bytes
• scripts/rollup_sqp.pyruns · 12,964 bytes
• tests/test_rollup_sqp.pyruns · 9,344 bytes

# Amazon SQP / ASIN theme decisions

Find recurring query themes and return specific listing decisions in seller language: **what to change → why → what to watch**. Usually a few short paragraphs are enough. For no change, explain why the existing content should stay. For missing evidence, ask for the one fact that blocks the choice and explain what it will decide. Follow the [seller-answer contract](references/decision-records.md); the main answer must stand alone without a technical appendix.

Amazon definitions ground the metrics; ESCI informs full-query relevance and COSMO informs evidence-filtered intent hypotheses. Our gates and prioritization are analyst conventions, not Amazon ranking rules.

## Workflow

For Ads-only search-term, bid or negative-keyword work, use `amazon-ads-sp-search-term-explorer` when available. Otherwise explain the handoff and request suitable SQP data; do not apply SQP rules to ad-only totals.

1. **Establish scope.** Read [SQP metrics](references/sqp-metrics.md). Obtain marketplace from report request metadata or ask. Use complete WEEK/MONTH/QUARTER periods. Default to `traffic_scope: sqp_reported_scope`. Preserve reported rates with the denominator-conflict note; analytical ratios use raw counts. Carry `searchQueryScore` unchanged: neither it nor impression share establishes organic position.
2. **Get product and listing evidence.** Read [relevance and intent](references/relevance-and-intent.md). Retrieve available read-only catalog/listing evidence with correct identity/marketplace context, or use supplied facts. Reuse it per ASIN; record source, date and field coverage. Catalog discrepancies need seller/user confirmation before a factual correction. Assess each full query as exact, substitute, complement, irrelevant or unknown; do not infer capability from an intent hypothesis.
3. **Aggregate compatible themes.** Read [theme analysis](references/theme-analysis.md). Use `scripts/rollup_sqp.py` for reproducible contiguous 1/2/3-grams, summed counts, disjoint remainder comparisons and Wilson intervals. Preserve incompatible intent and unknown/excluded members. Record pooled defaults in the supporting analysis: 200 market query impressions, 20 ASIN clicks, 3 queries and 5 ASIN purchases in each cohort; all configurable. Keep concentration and membership checks in supporting records; explain uncertainty to the seller only through its effect on the decision. Individual thin rows may form an adequate theme.
4. **Choose the decision.** Apply [decision records](references/decision-records.md) and [ASIN rules](references/asin-rules.md). A specific verified content gap plus a localized comparison can justify one exact listing test; a seller-confirmed contradiction can justify a correction even with thin SQP. A metric gap alone cannot. Retrieve accessible decisive evidence; otherwise name one blocker. Keep no-change decisions brief.