Sort by measure_at ascending — oldest first, so the retro always catches stale ones.

If an experiment is > 7 days past its measure_at with no measurement, flag it as a discipline failure in the retro summary. This creates back-pressure against letting experiments rot.

Phase 2: Measure each experiment

For each due experiment:

1. Re-read its expected.metric and expected.baseline.
2. Pull the current value from the appropriate data source:
- ctr_14d / clicks_30d / position_30d → GSC - traffic_90d / conversions_90d → GA4 - citation_rate_60d → latest aeo-probe result for the page - lcp_ms / inp_ms → PageSpeed / CrUX
3. Compute actual vs expected:
- delta_vs_expected = (actual - expected.target) / (expected.target - expected.baseline) - >= 1.0 → hit target or better - 0.3 – 1.0 → partial win (directionally right) - -0.3 – 0.3 → inconclusive (noise range) - < -0.3 → loss (moved wrong direction or got worse)

Phase 3: Score the verdict

Rules for assigning verdict:

• Confounding events: If a Google algo update happened in the window, note it in result.notes and still record the verdict. Don't use confounds as an excuse to avoid measurement.
• Seasonality: For pages with known seasonality (e.g., tax software in April), compare year-over-year where possible.

Phase 4: Write result back to experiment YAML

Update the experiment file in place:

status: measured   # was: open

result:
  measured_at: 2026-04-30T16:12:00Z
  actual:
    metric: ctr_14d
    value: 0.034
  verdict: win
  delta_vs_expected: 1.32   # exceeded target by 32%
  notes: >
    Clean lift. No Google algo events in window. Position held steady
    at 3.1 ± 0.2, so the CTR gain is attributable to the title change,
    not to a ranking shift.

Do NOT delete or move the experiment. Append result fields. Append only.

Phase 5: Write retro summary

Write {workspace}/reports/seo/retros/retro-{YYYY-MM-DD}.md:

# SEO Retro — Week of {start_date} – {end_date}

Experiments measured this week: {N}

Wins ({K})

[exp-...-a7k] /alternatives/cursor — title +8 Tools → CTR +62% (target +20%)
[exp-...-b9q] /compare/x-vs-y — schema FAQPage → impressions +14%


Losses ({M})

[exp-...-p2n] /blog/X — content refresh → traffic -8% (expected +15%)
  - Note: coincided with Google March core update


Inconclusive ({L})

[exp-...-t4m] /use/gpt-5 — new page → 30d too early for ranking signal
  - Keeping open, remeasure at 60d mark


Patterns emerging
_{from seo-learn scan — only noted, not promoted yet unless ≥3 wins}_

Title numbers continue to win (4/5 recent experiments)
Schema markup on /compare/ pages consistently lifts impressions


Discipline check

Experiments overdue > 7d: 0 ✅
Changes shipped without exp log this week: 0 ✅
  (check: git log --since="1 week ago" -- 'content/' 'app//page.tsx'
   then diff against count of experiments with deployed_at in range)


Recommended actions for next week (≤ 3)

Apply add-numbers-to-title.md playbook to top 5 remaining P4-10 pages without numbers
Investigate /blog/X regression — was the refresh too aggressive?
Run AEO probe refresh — citation rate measured 4 weeks ago, need new baseline


Next retro: {YYYY-MM-DD}

Phase 6: Trigger follow-ups

• If any loss has delta_vs_expected < -0.5 → trigger seo-postmortem on that experiment
• If any pattern shows ≥ 3 wins in a category → flag for seo-learn (runs monthly, but note the candidate)
• If discipline check fails (any experiments overdue OR changes without logs) → note it in owner's weekly report

Quality bar

Retro is done only when:

• [ ] Every open experiment with measure_at <= today has a verdict assigned
• [ ] Retro markdown exists at retros/retro-{date}.md
• [ ] Each win/loss/inconclusive entry cites its experiment ID (linkable)
• [ ] Discipline check ran
• [ ] Recommended actions ≤ 3 (more = unfocused)
• [ ] If a loss had Δ < -0.5, postmortem is triggered (not just noted)

What I refuse

• To mark an experiment "win" because the owner wants a win. The measurement is the measurement.
• To retroactively change expected.target to make a loss look like a win.
• To delete experiments that came up short. Losses are the training data for lessons.
• To skip a week because "nothing interesting happened". The discipline check IS the interesting thing.
• To bundle "probably a win" verdicts. If I'm not sure, it's inconclusive.

Integration

Called by:

• seo-rhythm-weekly (every Friday)
• Ad-hoc when owner says "run the retro" or "measure last week's experiments"