Issue #2 of AI This Week. In issue #1 I said I'd bolted a self-running SEO system onto this site — a month has passed, so as promised, here are the real numbers.
The system in one paragraph (if you skipped issue #1)
Every Monday morning, two things run with no human involved:
- Data collection — a GitHub Action pulls query/position/click data from the Search Console API and commits it to the repo
- Analysis and action — a cloud Claude agent reads that data and commits its own fixes: title rewrites, internal-link reinforcement, even new posts (rules: no invented facts, no deploy if validation fails)
The first month's report card
| Metric | Start (2nd week of Aug) | One month later |
|---|---|---|
| Weekly impressions | 0 | fluctuating 40–100 |
| 28-day impressions | 0 | 238–343 |
| Cumulative clicks | 0 | 8 |
| Flagship post's position | 47 | 5.5 (page 1) |
Small numbers — but on a one-month-old domain, $0 in ads, and a few minutes of human time per week.
Three things that worked
1. Rewriting titles against real query data. The install guide was originally titled "Getting Started with Claude Code," but all ~50 incoming queries said install. Renaming it to match moved it from the 40s to around position 5. Highest-impact action of the month.
2. Adding a comparison table. The ChatGPT vs Claude post was stuck at 32; adding a decision-criteria table moved it to 16 within two weeks. Search engines and AI answer engines both love tables.
3. Concrete numbers win the clicks. The first click — and the next ones — went to "actual cost breakdown"-style posts. Vague posts get impressions; specific posts get clicks.
What didn't work / lessons
The new-site honeymoon is real. Week three hit 100 impressions; week four dropped to 40. At first it looked like a mistake on our end — it's actually Google's normal process of test-showing a new site broadly, then narrowing to where it truly ranks. If impressions drop while average position jumps (ours went from the 30–60 range to 3–9), that's not a bad signal.
Never let the publishing rhythm stall. I'd given the automation a rule — "only write new posts when data justifies it" — and in the early weeks data never justified it, so the site published nothing for a month. The freshness signal to crawlers ("this site changes weekly") got cut off by the system's own caution. Fixed this week: biweekly publishing continues even without data triggers.
My two Monday robots collided. The data-collection bot and the analysis bot both commit to the repo in the same window, and one week a push got rejected. Three lines of retry logic fixed it — automation teaches you more in operation than in construction.
Next experiments
- How fast impressions recover once the publishing rhythm restarts
- First AI answer-engine citation (ChatGPT/Perplexity) — testing the hypothesis that citations follow rankings by a few weeks
- Whether free kit skill downloads correlate with search arrivals
The whole process stays public. Get issues by newsletter.