search-term-keyword-relevance

Search Term ↔ Keyword Relevance

Goal (job-to-be-done)

Per (search term × triggering keyword) pair, judge whether the keyword deserved to trigger the term. Google builds those matchings to maximize its spend, not the advertiser's commercial intent. The unit of analysis is the term↔keyword pair; the output is a relevance verdict that yields the action (negative, tighten keyword, move, review).

• Who: media buyer / PPC manager. When: recurring, on the search terms report.
• Decision it drives: what to negativize, which keyword to tighten, what to move, what to review by hand — without killing good traffic.
• The differentiator: not "better NLP than Google" — business context. The AI reads the account signals (products, geos, brand, competitors); it does not imagine the business.

Scope

• ✅ Semantic relevance, term↔keyword only. Minimal input: keyword + search term.
• ❌ Performance (conversions, CPA, ROAS, landing) → complementary performance skill. A relevant term that doesn't convert is still relevant.
• ❌ Account audit (bids, budgets, building campaigns/PMax).

Components (read these references as needed)

• Tools / data plan: references/tools.md — the exact 2-field query.
• Framework / rubric: references/framework.md — the brain: verdicts, traps, the "fix the keyword vs negativize the term" rule, PMax mode.
• Output schema: references/output.md — the JSON this skill emits.

Operate

Input: per row, a search term + its triggering keyword (match type splits the output into 3 tables; it is not a data column). Plus business context (required): products & geos served, what is NOT sold, own brand name(s), competitor names + bidding policy. Derive the context from account signals first; confirm the doubtful; ask only what can't be inferred.

Process: run one pass per match type (EXACT, PHRASE, BROAD), grouped by keyword. Apply the rubric in references/framework.md. Never use conversions / CPA / ROAS / cost to classify — relevance only.

Emit the JSON in references/output.md:

• synthesis — three strings: headline (where relevance leaks most + the single action), diagnosis (the keyword most off-target terms leak from + a recurring drift/leak n-gram across keywords), action (the one fix to take now, where / what / why).
• groups[] — one per match type; inside, keywords with their classified terms (verdict + action). Flag a keyword too_loose when most of its terms are loose/leak.
• ngrams[] — phrase/exact negative candidates, each with a blast-radius note.

A renderer (porter-reporting, or a chat view) turns that JSON into the human table. Do not bake emojis/layout into the analysis output.

Example (illustrative — from real accounts, NOT rules)

• Misplaced: keyword best dental insurance plan triggered health insurance → relevant to the business (sells health) but to another product → move_to_other_campaign, not negative.
• Leak factory: a broad keyword with dozens of terms irrelevant to it → tighten_keyword, don't chase each term (negativizing each = mopping with the faucet open).
• Performance trap: a keyword whose terms are relevant but don't convert → justified; the low performance belongs to the other skill.