Clean up search terms and negatives

PPC specialist / Google Ads operator / agency paid-search manager · First run in ~20 min after install; subsequent daily runs take <10 min of human time (approve/reject the proposal list).

GTMEdit flowon X◆ new (unverified) 83

step 1connector

Connect to the Google Ads account & export yesterday's search terms

⚡ bring your own tool

step 2ready

De-risk probe: validate query pull schema before full run

optimization-loop-smoke53 alts

step 3partial

Intent-classify every query at scale

keyword-research15 alts

step 4ready

Run the full optimization loop: surface negatives + asset hygiene

optimization-loop

step 5ready

Pull account-level Google Ads data for spend-impact context

ads-google3 alts

step 6connector

Human approval gate: review and approve/reject negative proposals

⚡ bring your own tool

step 7connector

Upload approved negatives via Google Ads Editor import

⚡ bring your own tool

Did this recipe work for you?

What you get

▸approved_negatives_csv — upload-ready negative keyword list segmented by campaign and match type
▸daily_cleanup_log — timestamped record of what was added, rejected, and escalated
▸spend_impact_summary — wasted spend recovered headline (impressions/clicks/cost on junk terms blocked)
▸ads_audit_report — broader account health snapshot (quality score, asset hygiene) produced as a by-product of the optimization-loop run

tools: Google Ads Editor: free. Google Ads API: free quota covers typical daily pulls. Two Minute Reports MCP: ~$20-49/mo if used. No other paid tools required.

tokens: ~$1-4 per daily run depending on query volume (intent-classification of 1k-10k rows is the main token cost)

≈ $0-50/mo in tooling + ~$30-120/mo in tokens at steady daily cadence — well under the cost of one hour of manual PPC work

What stays yours

· Review and approve/reject the negative proposals each day — this gate is intentional and should not be automated away; false negatives (blocking good terms) cost real revenue
· Tune the intent-classification prompt after the first 3-5 runs as you see which query patterns the agent miscategorizes for your specific account and vertical
· Escalate ambiguous terms (competitor brand queries, generic category terms) that require business-context judgment the agent cannot make autonomously

The problem

When my search campaigns accumulate tens of thousands of queries, I want an agent to scan every query daily, intent-filter the waste, and propose negatives for one-click approval, so I stop hand-reviewing 30,000+ terms and stop bleeding spend on junk between manual reviews.

Cybernetic layers

Memorypartial

The optimization-loop workflow tracks changes 'since last run' — it has lightweight run-state memory baked in. Full cross-session negative history (what was added, when, which campaign) still needs a persistent store such as a Google Sheet or local CSV log. Recommend appending each day's approved_negatives_csv to a master negatives ledger file.

Background-schedulingbyo

The recipe is batch-shaped and runs in one session. To make it truly daily-unattended, wire it to a Claude Code cron (CronCreate) or a system cron that fires the agent each morning before business hours. The human approval gate still requires a human touch-point, so full unattended automation is not appropriate here.

Harnessready

Claude Code is the runtime. Install the skill and workflow artifacts into ~/.claude/skills/ and invoke by name. The optimization-loop workflow is a .workflow.js file — drop it into ~/.claude/workflows/.

Click any step to see what fills it — and the ranked alternatives you can swap in.