Amazon Search Term Harvest Planner

---
name: amazon-search-term-harvest-planner
description: "Find Amazon Ads search terms ready for exact-match harvesting from auto, broad, phrase, or other discovery campaigns. Use when Codex needs standalone or Rocketcart MCP-aware search-term harvesting: classify terms by harvest readiness, resolve live IDs, choose safe destination campaigns/ad groups, preflight duplicates/current negatives/destination feasibility, decide whether source negatives are justified, execute only explicitly approved rows, and read back/monitor Sponsored Products or Sponsored Brands actions without unsafe negatives, duplicate routing, stale approvals, brand-defense cuts, own-ASIN-defense cuts, launch/rank mistakes, or low-sample overfitting."
---

# Amazon Search Term Harvest Planner

## Purpose

Turn Amazon Ads search term data into a safe, specific harvesting plan. Identify search terms that deserve exact-match isolation, product-target expansion, bid direction, destination routing, or watchlist treatment. Do not treat every converting query as harvest-ready, and do not add source negatives unless routing or waste evidence supports it. Use orders or conversions as the negative-safety signal; treat sales as revenue, not order count.

Optimize for clean traffic control, profitable growth, and learning. Preserve brand defense, own-ASIN defense, launch/rank-defense traffic, and strategically valuable discovery until the data proves a safer route.

This skill works in standalone mode from static exports and in Rocketcart MCP mode when a live Amazon Ads + product-intelligence layer is available. It may prepare approval-ready rows, but it must not create keywords, negatives, product targets, bids, budgets, or campaigns by default.

## Connection Modes

### Standalone Mode

Use this mode when the user provides pasted tables, CSVs, screenshots, or summaries.

- State that Rocketcart MCP was not used.
- Build the best possible harvest plan from static search-term, targeting, negative, product, and destination data.
- Lower confidence when live entity IDs, current negatives, current bids/states, product context, snapshots, or recent changes are unavailable.
- Do not present rows as executable without live preflight.

### Rocketcart MCP Mode

Use this mode when Rocketcart MCP capabilities are available or the user asks for live Rocketcart review, preflight, execution, readback, or monitoring.

Initial Rocketcart review is read-only. Do not use write capabilities until the user explicitly approves exact rows after live preflight.

- If profile is missing, use profile discovery first. If exactly one profile fits, state the assumption; if multiple profiles plausibly match, ask the user to choose.
- Read current campaign, ad group, product-ad, keyword/target, negative, budget, state, bid, and destination context where available.
- Read search-term, targeting, recent-change, snapshot/changelog, and entity-history context where available before trusting stale exports.
- Read product intelligence where available: ASIN/SKU mapping, inventory or availability, Featured Offer / Buy Box, price, reviews/rating, category rank/BSR movement, estimated demand, competitor signals, margin/readiness, and mixed-ASIN risk.
- Use live reads to resolve exact IDs, detect duplicate exacts/targets, detect current negative conflicts, verify destination delivery feasibility, and detect stale rows.
- Produce approval packets for exact candidate rows; execution remains separate.
- After any explicitly approved execution, read back affected entities and define 3/7/14-day monitoring.
- Read [Rocketcart search-term harvest mode](references/rocketcart-search-term-harvest-mode.md) when you need the Rocketcart-specific read, preflight, approval, execution, and readback sequence.

## Rocketcart Review Modes

Choose one primary mode when Rocketcart is involved:

| Mode | Use When | Output |
|---|---|---|
| `Live Harvest Review` | The user wants a current search-term harvest plan using live account state. | Live-resolved classifications and proposed rows, no writes. |
| `Preflight / Approval Readiness Review` | Candidate harvest, negative, target, bid, or delivery-fix rows already exist. | Row-by-row readiness, stale-row blockers, exact approval text. |
| `Execute Approved Rows` | The user explicitly approves exact rows for execution. | Execute only the approved subset after live preflight, then read back. |
| `Post-Change Readback / Monitoring Review` | Changes were executed and need verification or outcome review. | Readback status and 3/7/14-day monitoring verdicts. |

## Required Inputs

Gather, derive, or mark unavailable:

- Marketplace, profile/account, currency, timezone, ad type, and date windows.
- Search term report with campaign, ad group, search term, match type/source targeting where available, impressions, clicks, spend, orders, sales, ACoS/ROAS, CPC, CTR, and CVR.
- Targeting or keyword report with existing keywords/targets, match type, bids, states, and campaign/ad group destinations.
- Campaign and ad group structure: targeting type, naming, budgets, states, portfolio, advertised ASINs, and strategic role.
- Product context: advertised ASIN, purchased ASIN where available, parent/child relationship, category, price, margin or target ACoS/CPA, inventory, Featured Offer / Buy Box, reviews, rating, delivery promise, and listing readiness.
- Traffic segmentation: branded, own-brand generic, category generic, competitor brand, competitor ASIN, own-ASIN defense, auto close/loose/substitute/complement, launch/rank-defense, and exploratory discovery where data supports it.
- Existing exact keywords, negatives, product targets, and campaign routing rules when available.
- Recent orders or conversions by normalized search term, keyword, target, ASIN-like term, and phrase family for negative decisions. Default to L30 when available; state any different window.
- Destination delivery feasibility: destination campaign/ad group state, budget status, negative conflicts, advertised ASIN fit, and whether the destination can receive traffic.
- In Rocketcart MCP mode: exact profile, live entity IDs, current live values, recent drift, snapshots/changelogs, entity history, product-ad ASIN/SKU mapping, and product-intelligence freshness where available.

## Missing Data Handling

- Missing margin or target ACoS/CPA: avoid firm profitability claims; use ACoS/ROAS/CPA only as proxies.
- Missing search term source mapping: do not recommend source negatives or precise routing.
- Missing destination campaign structure: propose a destination pattern, but mark the action `Needs Destination`.
- Missing existing exact keywords/targets: flag duplicate-risk before launching new exact terms.
- Missing current negative map: do not mark harvest or destination rows `APPROVAL_READY`.
- Missing exact profile, campaign, ad group, keyword, target, or destination IDs: keep write rows below `APPROVAL_READY`.
- Missing purchased-product data: avoid ASIN leakage conclusions.
- Missing recent order/conversion data or unmapped normalized traffic for a proposed negative: mark the negative row `NEEDS_DATA`, not `APPROVAL_READY`.
- Missing inventory or Featured Offer / Buy Box: do not recommend aggressive bid or budget scale.
- Missing comparison period: classify by current signal strength and confidence, not trend.
- Missing Rocketcart profile in live mode: list or request profile selection before live reads or execution.
- Missing live resolution or live preflight in Rocketcart mode: keep rows below `APPROVAL_READY`.
- Live state differs from a candidate row's current value: mark the row stale, block execution, and request refreshed approval.

## Write-Readiness Statuses

Every action row must use exactly one status:

- `PLANNING_ONLY`: strategic recommendation, destination pattern, or analysis that is not a live write candidate.
- `NEEDS_DATA`: required report, exact ID, destination, product context, economics, duplicate check, negative check, or strategic role is missing.
- `BLOCKED`: unsafe or non-executable due to duplicate conflict, negative conflict, broad negative blast radius, brand/own-ASIN/launch/rank protection, retail-readiness blocker, budget-starved destination, mixed ad-type contamination, or failed gate.
- `APPROVAL_REQUIRED`: row is specific enough to review, but still needs explicit human approval and live preflight before any write.
- `APPROVAL_READY`: human can approve the exact row after verifying live preflight. This status is rare and requires all readiness fields below.

No row may be `APPROVAL_READY` unless it includes exact profile/account, marketplace, ad type, source campaign ID, source ad group ID, destination campaign ID, destination ad group ID, keyword/target or negative ID where applicable, normalized search term, match type, current state, current value, proposed value, proposed action, duplicate checks, current negative checks, destination feasibility, approval text, preflight checks, readback checks, and 3/7/14-day monitoring criteria.

In Rocketcart MCP mode, `APPROVAL_READY` means the row is ready for explicit human approval after live preflight. It does not mean "execute now." Execution requires a separate user approval message naming the exact row or row IDs.

## Hard Blockers

Classify the row as `BLOCKED` or `NEEDS_DATA`, not `APPROVAL_READY`, when any of these apply:

- Existing exact keyword or product target already covers the normalized term, unless the action is delivery repair for an existing entity.
- Existing exact exists but is paused, archived, budget-starved, blocked by a negative, or in the wrong destination; classify as `Scale Existing Exact / Delivery Fix`, not duplicate harvest.
- Existing keyword/target report is missing; classify harvest rows as `NEEDS_DATA`.
- Current negative map is missing or destination has a negative that would block delivery.
- Destination campaign/ad group is missing, paused, budget-starved, wrong ASIN, wrong traffic type, or missing exact IDs.
- Search term is a brand-defense, own-ASIN-defense, launch/rank, or strategic discovery term and the proposed source negative would cut protected traffic.
- Competitor conquesting term has high ACoS but the source campaign's strategic role is conquesting or share defense; prefer `Bid Down / Keep Learning`, `Watchlist`, or controlled test.
- Proposed negative keyword, search-term negative, product-target negative, or ASIN/product-target negative has `>= 2` recent orders or conversions for the same normalized term, target, ASIN-like value, or covered phrase family. Default to L30 when available, aggregate across relevant campaigns/ad groups in live mode, and do not use sales revenue as an order count.
- One order, one large order, or thin samples create attractive ACoS; use `Controlled Test` or `Watchlist`, not `Harvest Ready`.
- Negative phrase would block relevant query families, own-brand variants, size/color variants, profitable discovery, protected strategic traffic, or any covered query/variant/target/ASIN-like value with `>= 2` recent orders or conversions. Use narrower negative exact or block the negative.
- Sponsored Products, Sponsored Brands, and Sponsored Display data are blended and cannot be separated for the action.
- Product readiness, inventory, Featured Offer / Buy Box, margin, or listing relevance could change the decision and is unavailable.
- Rocketcart live state shows a current value, entity state, negative, destination, product-readiness gate, or recent drift that differs from the candidate row.

## Evidence Thresholds

Use user-provided thresholds when available. Otherwise use these defaults, adjusted for product price, category, lifecycle stage, and account volume:

- `Harvest Ready`: at least 2-3 orders, relevant intent, ACoS/CPA within target economics, acceptable CVR, no retail-readiness blocker, duplicate checks passed, current negative checks passed, destination is feasible, and a clear exact destination exists.
- `Controlled Test`: 1-2 orders or promising CVR/CTR but not enough volume for a confident harvest; use lower bid, limited budget, routing review, or watchlist. Do not default a one-order term to a negative.
- `Scale Existing Exact`: the term already exists as exact and has enough profitable volume; recommend delivery, bid, budget, state, negative-conflict, or placement review instead of duplicate harvesting.
- `Product Target Candidate`: search term or purchased ASIN indicates an ASIN target should be tested, with relevance and economics checked.
- `Bid Down / Keep Learning`: relevant term with orders but above target economics; reduce bid only when current evidence and strategic role support it.
- `Negative Candidate`: irrelevant or structurally mismatched term, or spend exceeds 1.5-2.0x target CPA with zero orders and no strategic defense, launch, ranking, discovery, or recent-order guard reason.
- `Watchlist`: below threshold, missing data, unclear relevance, or strategic role not yet resolved.

Do not overfit tiny samples. If evidence is thin, use `Watchlist`, `Controlled Test`, or `Needs Data`.

Lifecycle adjustments:

- Launch/rank window: tolerate higher ACoS only when the rank objective, budget cap, stop-loss, inventory, and monitoring rules are explicit.
- Brand defense: prioritize coverage and defensive routing over short-window ACoS, but do not add source negatives without approved defense coverage.
- Mature profit campaign: require economics, duplicate checks, and stable CVR before harvesting or bid increases.
- Exploration/discovery: prefer controlled tests and watchlists over negatives until waste is isolated.

## Traffic-Type Routing Matrix

| Traffic Type | Preferred Destination | Source Negative Default | Key Gate |
|---|---|---|---|
| Brand defense | Brand exact or defense ad group | Blocked unless defense coverage is approved and verified | Preserve branded and own-brand demand. |
| Category generic | Nonbrand exact by ASIN/product group | Negative exact only after destination delivery is verified | Check relevance, margin, inventory, and duplicate exacts. |
| Competitor brand | Competitor/conquest exact campaign | Usually blocked; bid down or controlled test first | Confirm strategic role and expected ACoS tolerance. |
| Own-ASIN defense | Own-ASIN product-target or defense campaign | Blocked unless own-ASIN defense is intentionally rerouted | Requires ASIN map and purchased-product context. |
| Substitute/complement ASIN | Product-target test | Blocked until ASIN relationship is known | Avoid own-ASIN or irrelevant target mistakes. |
| Launch/rank | Rank-support exact or launch campaign | Blocked unless launch/rank coverage is preserved | Requires objective, stop-loss, inventory, and monitoring. |
| Discovery/exploratory | Controlled exact test or watchlist | Blocked unless waste is isolated | Do not cut learning from thin samples. |
| Irrelevant query family | No destination, negative review | Negative exact or narrow phrase only | Phrase negatives need blast-radius proof. |

## Harvest Safety Gates

Before recommending an exact harvest, check:

- Normalize the search term for case, punctuation, plural/singular variants, spacing, ASIN casing, and simple token order differences before duplicate checks.
- The search term is not already isolated as an exact keyword or product target unless the action is to fix routing, state, delivery, budget, or negative conflicts.
- The destination campaign/ad group is known or the output clearly proposes one.
- The destination matches traffic type and business goal: branded, category generic, competitor, own-ASIN defense, rank-growth, profit, discovery, or product targeting.
- Destination delivery is feasible: enabled state, budget headroom or budget plan, no current negative conflict, correct advertised ASIN/product group, and exact destination IDs when claiming approval readiness.
- The advertised ASIN is retail-ready enough for the traffic.
- The term is relevant to the destination ASIN or product group.
- The proposed match type, bid direction, budget stance, and negative strategy are explicit.

Before recommending a source negative, check:

- The term has been safely captured elsewhere or the source traffic is clearly wasteful.
- Recent order/conversion data is available for the normalized term, target, ASIN-like value, and covered phrase family. Default to L30 when available; state any different window.
- Orders/conversions are aggregated across relevant campaigns/ad groups when live data is available, not judged only from the source row.
- The proposed negative is blocked when the same normalized traffic has `>= 2` recent orders or conversions. If it has exactly 1 order/conversion, default to `Watchlist`, `Controlled Test`, `Bid Down / Keep Learning`, or routing review instead of a negative.
- Sales revenue is not used as an order count.
- Adding a negative will not cut brand defense, own-ASIN defense, launch/rank-defense, or profitable discovery.
- The negative scope is correct: campaign-level versus ad-group-level.
- The negative match type is justified: negative exact for routing control, negative phrase only for clearly irrelevant query families.
- Blast radius is reviewed: blocked tokens, variants, brand terms, ASINs, size/color variants, and profitable nearby queries are not harmed.

## Machine-Readable Action Row Schema

When producing action rows, include this schema in table form or JSON-like rows:

```text
row_id
mode
action_type
write_readiness
approval_status
execution_status
profile_or_account
rocketcart_profile
marketplace
ad_type
source_campaign_id
source_campaign_name
source_ad_group_id
source_ad_group_name
search_term_normalized
traffic_type
classification
destination_campaign_id
destination_campaign_name
destination_ad_group_id
destination_ad_group_name
keyword_or_target_id
negative_entity_id
match_type
current_state
current_value
proposed_value
proposed_action
source_negative_decision
duplicate_check
current_negative_check
recent_orders_window
recent_orders_check
negative_conversion_guard
destination_feasibility
live_resolution_status
live_preflight_status
product_readiness
reason
primary_risk
confidence
approval_text
preflight_checks
readback_checks
readback_status
monitoring_owner
monitoring_3d
monitoring_7d
monitoring_14d
```

Use `missing` for unavailable IDs or checks; do not invent them. If any required field for a live write is `missing`, the row cannot be `APPROVAL_READY`.

## Workflow

1. Establish data coverage and freshness.
   - State connection mode, Rocketcart review mode when relevant, windows, ad types, report freshness, attribution caveats, search-term grain, missing source/destination fields, and duplicate-risk.

2. In Rocketcart MCP mode, resolve live scope.
   - Confirm profile first.
   - Read live campaigns, ad groups, product ads, keyword/target coverage, negatives, budgets, states, recent changes, snapshots/changelogs, and product context where available.
   - If profile is ambiguous, live reads fail, or product context is unavailable for a product-sensitive action, stop at `NEEDS_DATA`, `BLOCKED`, or `APPROVAL_REQUIRED`.

3. Segment search terms.
   - Classify terms by traffic type, source campaign/ad group, advertised ASIN, purchased ASIN where available, and strategic role.

4. Score harvest candidates.
   - Evaluate orders, spend, sales, ACoS/CPA, ROAS, CVR, CPC, relevance, margin fit, retail readiness, lifecycle stage, destination clarity, duplicate risk, current negative conflicts, recent-order negative guard, source-negative blast radius, and incrementality caveats.

5. Decide the route.
   - Assign each meaningful term to one primary outcome: `Harvest Ready`, `Controlled Test`, `Scale Existing Exact`, `Product Target Candidate`, `Bid Down / Keep Learning`, `Negative Candidate`, `Watchlist`, or `Needs Data`.

6. Build action rows.
   - Use the machine-readable schema and write-readiness statuses.
   - No row can be `APPROVAL_READY` unless all required exact IDs, current/proposed values, duplicate checks, negative checks, recent-order guard, destination feasibility, approval text, preflight, readback, and monitoring fields are complete.

7. In Rocketcart MCP mode, separate approval from execution.
   - In live review and preflight modes, produce proposed rows only.
   - In execute mode, execute only exact row IDs explicitly approved by the user after live preflight confirms current values still match.
   - If live preflight differs from the approved row, block execution and produce a refreshed row for renewed approval.
   - After any execution, read back affected entities before reporting success.

8. Define monitoring.
   - 3-day pass: destination entity is enabled, receives impressions/clicks, and is not blocked by negatives, state, or budget.
   - 7-day pass: spend, clicks, orders, CPC, and CVR are within expected range; source traffic did not collapse unexpectedly.
   - 14-day pass: ACoS/CPA, ROAS, route quality, duplicate traffic, and query drift are acceptable.
   - Failure criteria should name the rollback, bid-down, negative-review, or watchlist response.

## Output Format

Return these sections unless the user asks for a shorter version:

1. **Data Coverage And Harvest Gate**: windows, sources, missing fields, duplicate-risk, destination clarity, retail-readiness caveats, and whether harvest actions are action-safe.
2. **Rocketcart Live Context**: mode, profile, live reads used, snapshots/changelogs checked, product context checked, stale-state findings, and live limitations. In standalone mode, say Rocketcart MCP was not used.
3. **Executive Summary**: top harvest-ready terms, blocked terms, negative-risk warnings, and budget/bid posture.
4. **Search Term Classification Table**: Search Term | Source Campaign / Ad Group | Traffic Type | Orders | Spend | Sales | ACoS/CPA | Relevance | Destination | Classification | Confidence.
5. **Harvest Action Rows**: include the machine-readable schema fields, especially mode, write readiness, approval status, execution status, exact IDs, current/proposed values, duplicate check, current negative check, recent-order negative guard, live resolution status, live preflight status, destination feasibility, approval text, preflight, readback, and monitoring fields.
6. **Negative And Routing Decisions**: explain which source negatives are safe, blocked, or need more data.
7. **Blocked / Watchlist Terms**: terms below threshold, missing data, duplicate-risk, retail-readiness blocked, or strategically sensitive.
8. **Execution Gate**: state which rows are planning-only, blocked, approval-required, approval-ready, executed, readback-pending, or readback-confirmed.
9. **Monitoring Plan**: 3-day, 7-day, and 14-day checks with success/failure criteria.
10. **Missing Data / Next Pulls**: existing exact keywords, negatives, target report, purchased-product report, margin, inventory, Featured Offer / Buy Box, destination structure, live IDs, live negatives, snapshots, and product context.

## Live Execution

This skill proposes actions by default. If connected to Rocketcart MCP or another live execution layer, do not create keywords, negatives, targets, bid changes, budgets, or campaigns without explicit approval, live preflight, exact entity IDs, current/proposed values, readback, and monitoring criteria.

When execution is explicitly requested and approved:

- Execute only rows whose `write_readiness` is `APPROVAL_READY`, whose `approval_status` is explicitly approved, whose live preflight still matches the approved current values, and whose recent-order negative guard still passes.
- Never execute `PLANNING_ONLY`, `NEEDS_DATA`, `BLOCKED`, or merely `APPROVAL_REQUIRED` rows.
- Never execute rows selected by vague language such as "all recommendations"; require exact row IDs or exact entity/action approval.
- Execute only the approved subset; leave unrelated rows untouched.
- Immediately read back created or changed keywords, targets, negatives, bids, budgets, product-ad states, or campaign states.
- If readback cannot verify the change, report `readback_pending` or `readback_failed` rather than claiming success.