| name | search-term-performance |
| description | Judge the MONEY each Google Ads search term makes or wastes — aggregate every row of the same term, compute cost / conversions / CPA / conv-rate (or ROAS for ecommerce) from the base counts, and classify each term Winning (converts at/under the account benchmark) / Watch (converts above it, or too few clicks to judge) / Waste (real spend, ~0 conversions, past the thin-data floor), ranked by spend, with a destination (promote to exact / leave / add negative) and the dollars at risk. Use this skill whenever the user works the search-terms report and asks which terms are wasting budget, which to promote to exact keywords, which to negativize for poor performance, "what's burning money", cost/CPA/ROAS per search term, or wants the spend-ranked keep/cut read. Performance ONLY — the matching verdict (is this term relevant to the keyword?) is the relevance skill; who should own a duplicated term is term-routing; new keyword/asset ideas are intent-discovery. |
Search Term Performance (the money axis)
Goal (job-to-be-done)
Per search term (aggregated across the keywords that triggered it), judge whether it is earning
its budget: Winning, Watch, or Waste — ranked by spend, so the biggest dollars move first. It reads
dollars, not semantics.
- Who: media buyer / PPC manager. When: recurring spend hygiene on the search-terms report.
- Decision it drives: which terms to promote to an exact keyword, which to add as negatives
for poor performance, and which to leave — in spend order.
- The discipline: irrelevant ≠ poor performance. A term can be perfectly relevant to its
keyword and still waste money, or be loosely matched and still convert. This skill does not
re-judge the match (that is
relevance); it judges the money.
Scope
- ✅ Per-term cost / clicks / conversions / value / CPA / conv-rate / ROAS from the base counts;
Winning / Watch / Waste against the account's own benchmark; a destination + the dollars at risk.
- ✅ Thin-data floor — a 0-conversion term is Waste only once it has had enough clicks that ~1
conversion would have been expected at the account conversion rate. Below the floor it is
unproven (Watch), never Waste.
- ❌ Relevance / matching →
relevance. ❌ Cannibalization / ownership →
term-routing. ❌ New keyword / asset ideas → intent-discovery.
❌ It does not auto-apply negatives or bid changes.
Components (read these references as needed)
Operate
Input: per row, a search term + its match type + campaign + cost / clicks /
conversions / value (one query — references/tools.md). Optional context:
target_cpa (the account tCPA, if known — overrides the median benchmark), roas_target (ecommerce
break-even), min_clicks_floor, and model (lead_gen/ecommerce — force the business model
when an account that values its leads would fool the value-present auto-detect; verified on a live lead-gen account).
Process (deterministic — process.py):
- Acquire the data (
scripts/query.json).
- Feed the raw
{columns, rows} to scripts/process.py. It aggregates the
base counts per term (cost/clicks/conv/value sum correctly across a term's keywords; native ratio
fields do not), computes CPA/conv-rate/ROAS, sets the benchmark (the account tCPA if given,
else the median CPA across converting terms), derives the thin-data floor (≈ clicks for one
expected conversion), classifies Winning / Watch / Waste, and sets the destination + dollars at
risk. Lead-gen vs ecommerce is auto-detected from whether conversion value is present.
- The model only writes the synthesis (and intervenes if a destination needs a human eye).
Emit the JSON in references/output.md: meta (account type, benchmark,
floor) + terms[] (sorted by spend; class + destination + dollars_at_risk) + rollup. A renderer
(porter-reporting) turns it into the spend-ranked read. Emit pure data — no emojis/tables.
⚠️ Cost scale: google_ads_cost_micros is returned already in currency units (Porter
pre-converts — verified live 2026-06-23 on a production account: it reads 166.01, not 166007841). Do NOT
divide by 1e6.
How it feeds the Search Terms dashboard
Performance is the money lane the dashboard's insights card needs. Its Waste terms — intersected
with the relevance verdict — split into two insight rows: a term that is leak + Waste belongs to
the Irrelevant negatives line; a term that is relevant + Waste is the "right-fit but wasteful"
line (review/fix — do not tag it Irrelevant). The waste_cost rollup is the recoverable-dollars
input. Performance is never a tag on a term row; it lives in the insights card.
Example (illustrative — FICTIONAL Acme Insurance, see scripts/example.json; NOT rules)
- Winning → promote:
term life insurance (PHRASE+BROAD, $380, 14 conv, CPA $27 ≤ $40) → broad-served
winner → promote to exact.
- Watch (converts, over benchmark):
cheap life insurance ($300, 1 conv, CPA $300) → leave, revisit.
- Waste:
life insurance jobs ($90, 45 clicks, 0 conv, past the 20-click floor) → add negative;
$90 at risk.
- Thin-data (NOT waste):
is life insurance worth it ($12, 5 clicks, 0 conv) → below the floor →
Watch (unproven), never called waste on 5 clicks.
