Date        Source       Event
2026-03-14  git          Merged PR #2847: canonical URL restructure
2026-03-15  deploy       Deployed to prod
2026-03-16  GSC          First indexed-page drop detected
2026-03-17  GSC          Drop accelerates, /use/ cluster worst hit
2026-03-17  google_algo  Minor core update rolled out (unlikely cause — limited scope)
2026-03-19  retro        Flagged for postmortem

Phase 3: Root cause analysis

Use the "5 whys" discipline, but stop when you hit something you can act on:

Why did indexed pages drop?
  → Because /use/ URLs returned 404 for Google's crawler.
Why did they return 404?
  → Because we changed canonical from /use/X to /use/X/ without redirects.
Why did we not add redirects?
  → Because the seo-create skill doesn't flag canonical structure changes
    as needing redirect pairs.
Why doesn't it flag this?
  → Because we didn't previously have a lesson documenting the risk.
(Stop — this is the actionable layer.)

Differentiate root causes from contributing factors. A Google algo update in the same week might be noise; prove it's the cause before blaming it.

Phase 4: Fix the immediate damage

Postmortem doesn't just document — it also ships the fix if one is obvious:

• Restore redirects for the 227 de-indexed URLs
• Request re-indexing via GSC URL Inspection API
• Log a new experiment to measure recovery

• What was fixed
• When fix deployed
• Expected recovery timeline (days until re-indexation)

Phase 5: Write the lesson

Write {workspace}/reports/seo/memory/lessons/lesson-{YYYY-MM-DD}-{short_id}.md:

# Lesson: Canonical URL changes require paired 301 redirects

Incident
2026-03-15 → 2026-03-18. Indexed pages dropped from 324 to 97 (-70%).
Cause: changed canonical from /use/X to /use/X/ in PR #2847 without
adding 301 redirects from old URLs.

Root cause
Missed the redirect step. Canonical change alone is not enough —
Google treats old URLs as 404 and de-indexes.

Rule going forward
Any change that modifies URL structure (canonical, path, locale
prefix, trailing slash) MUST include paired 301 redirects deployed
in the same PR.

Concretely:

If changing href or canonical tag structure → audit all old URLs → add redirects
Pre-flight check: grep -r 'canonical' app/ before and after, diff must be paired with redirects
If using Next.js, check next.config.js redirects array before merging



Enforcement

Added pre-flight checklist item in seo-create skill: "URL structure change?"
Added to seo-experiment-log warnings: experiments of type=canonical or type=redirect
  prompt "is the redirect pair included?"


Recovery

Redirects deployed 2026-03-19
GSC re-indexing requested 2026-03-20
Measurement experiment: exp-2026-03-20-rec tracks recovery (measure 2026-04-20)


Related

Triggered by: postmortem-2026-03-19
Updates: seo-create/preflight-checklist.md
Updates: seo-experiment-log/SKILL.md warnings section

Phase 6: Update guardrails (THE KEY STEP)

A lesson that doesn't change behavior is a lesson that will re-teach itself. For each lesson:

1. Identify which skill should carry the guardrail
2. Edit that skill to explicitly reference the lesson
3. If there's a pre-flight checklist or warning section, add the specific check

Phase 7: Write postmortem summary

# Postmortem — {incident name}

Severity: high | medium | low
Affected: {what broke}
Detected: {date + source}
Resolved: {date + action}
Duration: {hours/days}
Data loss / impact: {numeric summary}

Timeline
{table from Phase 2}

Root cause
{concise from Phase 3}

Fix
{Phase 4 actions}

Lesson written
lessons/lesson-{date}-{id}.md

Guardrails updated

{skill 1 + section}
{skill 2 + section}


Follow-up tasks

{recovery measurement experiment}
{review related pages for same issue}
{anything else}


Could we have prevented this?
{Honest assessment. Was this predictable with existing knowledge?
If yes, why didn't we? If no, this is the class of incident we
couldn't avoid.}

Could we detect it faster next time?
{What signal would have caught this on day 1 instead of day 4?
Add detection — alert thresholds, dashboard views, etc.}

Quality bar

• [ ] Root cause is actionable, not a description ("lost topical authority" is not a root cause; "aggressive content rewrite removed cited sections" is)
• [ ] Lesson file exists with a concrete "rule going forward"
• [ ] At least one skill file was updated to enforce the lesson
• [ ] Immediate fix is deployed OR tracked as a P0 task if it's larger than a quick patch
• [ ] Recovery is being measured (not just "fixed, moving on")

What I refuse

• To write a postmortem without a "rule going forward" section. That makes the whole exercise performative.
• To attribute everything to "Google algo update" without evidence. That's how we learn nothing.
• To skip Phase 6 (updating guardrails). Without it, we WILL repeat the mistake.
• To write blame-laden prose. The goal is durable prevention, not accountability theater.

Integration

Reads: