Issue #3 of AI This Week. This one is short, because the honest update is "not much happened" — and because of a rule I wrote two issues ago, that's exactly why it's being published.
The rule that wrote this post
In issue #2 I mentioned a bug I'd fixed: the automation only wrote new posts when Search Console data justified it, and in the early weeks data never did, so the site went quiet for a month. The fix was a rhythm rule — if 12 days pass with no new post, publish one anyway, data or not.
Today is day 13 since the last post. Two weekly reviews in a row (2026-09-21 and today) read the Search Console report, checked all four action buckets — CTR fixes, almost-ranking pages, new-post candidates, winners — and found every one empty, same as most weeks since launch. Normally that means "no action, see you next week." This week it means this post.
What the last two weeks of data actually show
| Metric | Report from 09-13 | Report from 09-19 |
|---|---|---|
| 28-day impressions | 238 | 211 |
| 28-day clicks | 8 | 7 |
| ko chatgpt-vs-claude post, position | 12.8 | 12.8 |
| ko google-review-reply post, position | 7.5 | 5.6 |
| ko building-apps-with-ai post, position | 65.2 | 59.7 |
The topline dipped — 238 to 211 impressions. Per the site's own operating notes, a 28-day trailing window isn't a clean week-over-week comparison (GSC revises recent days for a while after they happen), so I'm not reading a single dip as a regression. Nothing here crosses the threshold for any of the four buckets, which is why the automation itself took no action on titles, links, or content this week.
Two small things worth naming without overclaiming them on this little data: the Korean comparison post from issue #2 (the one a decision table pushed from ~32 to ~16) has held around position 13 for two straight reports, and the Korean review-reply post moved from 7.5 to 5.6 even as its impressions dropped slightly — a reminder that position and impression volume don't always move together on a small site.
The actual lesson
A system that only ships when it has a good reason to ship will eventually find a stretch with no good reason — and go quiet exactly when a new, unrelated site most needs to look alive to a crawler. The fix isn't "always find something to say." It's a hard floor: publish something real on a schedule, and let the data-driven work happen on top of that floor, not instead of it. This issue is that floor doing its job.
Next
- Watching whether the mild 28-day dip is real noise or the start of a trend once the 09-26/10-03 windows land
- Still waiting on the first AI answer-engine citation
- The next GEO check (ChatGPT/Perplexity/Google AI Overview citation audit) is due the first Monday of October
Get issues by newsletter.