google-ads-builder-search

Google Ads — Builder: Search

Design a launch-ready Search campaign. Unlike PMax, Search pushes cleanly via Google Ads Editor CSV — so the campaign-spec maps almost 1:1 to an import file.

Governing principle — Search is the CONTROL layer over PMax's DISCOVERY

On a PMax/Shopping-heavy account, PMax is the discovery engine: it serves broad and surfaces which queries actually convert. A non-brand Search campaign exists to take control of those proven winners — explicit keywords, bids, match types, and RSA messaging instead of a black box. So the keyword plan is a harvest of the account's whole search-term data (Search + PMax), not a catalog brainstorm. Concretely:

• Source from both: search_term_view (exact terms + cost, Search/Shopping only) AND campaign_search_term_insight per PMax (converting category themes + conversions — NOT per-term cost or exact verbatim, so it's directional; you still choose the exact keyword from the theme).
• Harvest into EXACT — that's the actual control lever. An exact-match keyword identical to a query outranks PMax for that query; phrase/broad may not. So proven converters go in as exact, not phrase.
• Harvest selectively. Pull queries where control pays off (volume / value / margin / needs distinct messaging — e.g. high-AOV equipment). Leave long-tail one-off converters in PMax; thousands of exacts is overhead with no control gain.
• Coordinate, don't double-pay. Once a query is an exact in Search, PMax de-prioritizes it; judge incrementality, not shifted credit.

Operating rules

• Read context (brand_terms, competitor_terms, guardrails, vertical, budget) + the plan output.
• Write to the working directory. Emit BOTH blueprint.xlsx (human) and campaign-spec.json (campaign_type: "search", with ad_groups[]).
• Verify every Final URL is live. Enforce guardrails (conversion goals at campaign level; ENABLED-only).

Model dispatch (run cheap, decide expensive) — see ${CLAUDE_PLUGIN_ROOT}/references/model-tier-dispatch.md

• Scout (haiku) — per-URL 200 checks; the char-count validation pass; keyword dedupe.
• Routine (sonnet) — pulling the dual-source search-term data (search_term_view + campaign_search_term_insight per active PMax, incl. the catch-all) + catalog/services + Keyword Planner volume. Dispatch as general-purpose; return raw converting terms, don't theme.
• Judge (main session) — STEP 2 ad-group theming, STEP 3 match-type strategy, STEP 4 RSA copy (via assets), STEP 5 brand/non-brand separation. The keyword data is cheap; the structure is judgment.