Imprevista

Deploy Log

← Back to Deploy Log
|Marketing Bot|DEPLOYED

Loosen fanout detection (1->16 pages) + post 10 hand-authored drafts (#11)

The original fanout threshold (>50 chars, >=10 impressions, 0 clicks) only caught 1 query. Real GDA GSC data analysis showed the fanout space is much larger at a looser threshold.

The original fanout threshold (>50 chars, >=10 impressions, 0 clicks) only caught 1 query. Real GDA GSC data analysis showed the fanout space is much larger at a looser threshold.

Agent changes (agents/gda/gsc-refresh/agent.js):

  • Fanout threshold: query >=35 chars, total impressions >=3, CTR <2%
  • Aggregate by (slug, query) across the window before scoring (queries

repeat per day; need totals)

  • Filter search-operator noise (-site:reddit.com scraper queries)
  • Cluster by PAGE — one opportunity per page answering its query cluster
  • Fetch with min_impressions=1 so aggregation sees the full row set
  • Result: 1 -> 16 fanout pages detected from real data

Hand-authored content (no LLM API cost, per Charles's call to prove the loop before paying for API):

  • post-fanout-drafts.cjs: 10 fact blocks written manually from real GSC

fanout queries + real GDA deal data. Data-quality judgment baked in (KAC SR-15 and MK18 framed by configuration given wide SKU ranges; M855A1 framed as restricted-ammo cost-per-round reality).

  • All 10 posted to GDA as drafts; previews verified rendering real

content. Coupon-code drafts (no coupons exist) rejected.

learnings/drafts.jsonl: ledger now reflects the posted fanout drafts so the /drafts.html admin page lists them for approval.

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Files Changed

Commit:625071d