Segment rules (tunable) + gates

Document every threshold in the final report. Use references/threshold-elicitation.md before mining when values are missing.

---

Brand vs generic (required split)

Problem: Branded queries often look like “gold” because intent is navigational; mixing them with generic misleads “scale” recommendations.

Brand markers (required input)

• Before segmenting, obtain brand_markers: a list of substrings or patterns the user identifies as their brand (e.g. ["solid gold", "solidgold"]).
• If the user has not provided markers and has not opted out, ask before producing Gold / Rising lists. If they opt out, state split not applied and warn in §9.
• Best practice: collect brand markers and thresholds before heavy data pulls (e.g. before large SQL or before CreateReport in the MCP skill) so segmentation is planned up front and outputs always target two gold tables (generic + branded).

Classification

• ASIN-format terms (see below) are not branded/generic — they are a separate stream; exclude from branded/generic keyword tables by default.
• For remaining text queries: branded if any marker matches with token boundaries (avoid "sony" matching "baloney" — use word-boundary or curated whole-word lists). Misspellings can be added to markers for defensive rising-star handling.
• Generic = not branded under the above rule.

Output implications

• Generic gold → scale, exact harvest, bid tests.
• Branded gold → defend, monitor conquest, budget protection — not “new discovery.”
• Rising stars: same split; branded misspell = defensive, not generic scale.

---

ASIN-format “search terms” (product targeting stream)

SP reports surface ASIN strings (from product targeting / PAT campaigns) in the same column as real customer queries. In MCP v1 CSV that column is always searchTerm.value (dotted header in file). This is a fact about how SP reports work — partition before any keyword-segment logic.

Detection

• Primary pattern (common modern ASINs): regex ^B0[A-Z0-9]{8}$ (10 characters; verify against your catalog).
• Legacy / edge: ASINs can be other 10-character alphanumerics; if the user’s catalog shows exceptions, extend the detector and document it.

Default behavior

• Exclude ASIN-format rows from keyword Gold / Rising / Negative tables.
• In portfolio summary, report ASIN stream volume (rows, spend, clicks) separately.
• Why: Keyword negatives ≠ negative product targets; PAT campaigns (conquest/complement) often have different efficiency expectations — conflating them breaks recommendations.

Optional section

• If the user wants PAT analysis: rank ASIN-format rows by spend/efficiency in references/report-template.md §7, with PAT-appropriate actions only.

---

Data sufficiency (hard gate)

Thin windows (e.g. 1 day) and sparse clicks make confident negation and “scale gold” unsafe.

After aggregating by search term

1. Window length — window_days from file or user (if 1 day, say so explicitly).
2. Median clicks per term — median of clicks over terms with clicks > 0 (or over all terms; state which).

Gates (baseline)

Window-aware refinement (when window_days is known)

Use whenever the analysis window is known. For the MCP path, window_days derives directly from reports[0].periods[0].datePeriod (endDate - startDate + 1) — that's authoritative. The other sources listed here apply only if you ever process pre-existing exports.

Use median clicks per term after keyword-stream aggregation (same definition as above).

These rows refine the baseline table — apply the stricter outcome when both a baseline rule and a window-aware rule apply.

Confidence badges (per row or segment)

• High — clicks (and spend if neg) comfortably above segment minimums.
• Marginal — barely cleared thresholds.
• Degraded mode — sufficiency gate failed for strict actions; tables are exploratory only.

Copy for thin data

When gates fail: “This window is too short / sparse for strict negative or scale recommendations. Extend to 14–30 days and re-run.”

---

Gold terms (keyword stream, generic vs branded tables)

Intent: See report template — generic = discover; branded = defend.

• Sales > 0
• ACoS ≤ user gold ceiling (elicited)
• Clicks ≥ user minimum for gold (default 3 — raise when median clicks are low)
• Sort: typically sales desc for generic; user may override

---

Rising stars

• Spend band and sales > 0 per user
• Split generic (scale test) vs branded (defense / capture variants)
• Sort: CVR if orders+clicks; else sales

---

Negative candidates (keyword stream only)

• Spend > user floor; sales == 0 (or policy for “bad ACoS”)
• Strict recommendations only if sufficiency + min clicks for negatives met
• Sort: spend descending

---

Interest-only mode (cost missing)

When cost / spend is missing:

• No ACoS, no CPC, no spend-based negatives, no “historical waste” sum.
• Allowed: CTR/CVR-style mining if clicks, impressions, and sales OR orders exist — label interest signals, not “gold.”
• Checklist per segment:

---

Attribution alignment

See references/attribution-window.md. Mismatched 7d/14d breaks gold and negative efficiency logic.

---

Exclusions (legacy)

• Low impressions + zero sales: prefer hold vs negate
• New campaigns in learning: mention volatility