← Work · Case 03
ProductionAutomation & monitoring · n8n workflows

Search monitoring that diagnoses before it alerts

n8n collects, an AI step diagnoses, a human decides.

StatusProduction
Typen8n workflows
RoleConcept, design and build, with AI agents

01 · Problem

Problem

Search Console data across several properties is too much to read by hand and too noisy to trust blindly. Raw alerts either get ignored or trigger panic changes.

02 · Approach

Approach

Scheduled n8n workflows pull the data on daily, weekly and monthly runs. Each signal gets an AI diagnosis with its evidence attached, lands as a task in Notion and waits there. A person decides what, if anything, changes.

03 · Architecture

Architecture

Search ConsoleSOURCEcollect · n8nSTEP 01comparehistorySTEP 02AI diagnosisSTEP 03task · NotionSTEP 04human decisionHUMAN GATEchange or nochangeOUTPUTSOURCE → PROPOSAL │ HUMAN GATE │ INSPECTABLE OUTPUTSearch ConsoleSOURCEcollect · n8nSTEP 01compare historySTEP 02AI diagnosisSTEP 03task · NotionSTEP 04human decisionHUMAN GATEchange or no changeOUTPUT
Sanitized architecture · roles and decisions, never real data.

04 · Guarantees

Guarantees

  1. 01No workflow changes a live site.
  2. 02Every task carries the data it was based on.
  3. 03Unverified diagnoses are labelled as unverified.
  4. 04One run per schedule: no duplicate executions.

05 · Stack & decisions

Stack & decisions

n8nVisual workflows that are easy to audit and to roll back.
Google Search ConsoleThe source of truth for search performance.
Google SheetsA history anyone can open and check.
NotionTasks with a status, an owner and the evidence attached.
ClaudeThe diagnosis step, and only that step.

06 · Proof

Proof

Search monitoring · principle and hard rulesrules · authentic, sanitized
  1. principledetect → compare → verify → human decision
  2. the system never changes a live site.
  3. the human decides; the agent diagnoses.
  4. hard rules, each one paid for
  5. 1no bursts of manual runs: one recovery run, only if today has none
  6. 2never re-run weekly or monthly jobs blind: appends would duplicate
  7. 3deploy by procedure: read live → versioned backup → minimal change
  8. → local simulation → write → check status → read back → drift check
  9. 4data from a crashed run is unreliable
  10. 5fact ≠ hypothesis ≠ verified root cause ≠ unknown
  11. 6output is a file and a task. never a live change.

Excerpt from the real system, translated and sanitized: no names, identifiers or data.

Next case · 04Evidence-first agentic audits