What happened
Every post on my blog is written by an AI. That is not a confession, it is the experiment.
I started it for four reasons. I wanted to know whether a fully AI-written blog can earn search traffic. I wanted a blog that keeps publishing in the weeks I have no time to write. I wanted a real system that shows what I can build with AI agents. And I wanted to see, from the inside, how Google and AI assistants treat machine-written content.
"AI-written" here means the writing: the AI drafts every post. Approval is a separate step. The machine is described in The Autoblog Pipeline: topic discovery from five feeds, a Claude drafter, a two-stage safety gate, and a Telegram button I can tap to approve. If I do not tap, the post goes live anyway after two hours.
This page is the other half: what happened when it ran. The first AI-written post went live in February 2026. The numbers below run up to 6 November 2026.
The numbers
| Measure | Result |
|---|---|
| Posts published | 196 |
| Published without me tapping approve | 159 (81%) |
| Drafts stopped before going live | 34 |
| Average post length | about 2,000 words |
| Google impressions, whole site | 688 |
| Google clicks, whole site | 22 |
| Google clicks to blog posts | 5 |
| Pages Google indexed | 15 of the 218 it knows about |
Source: Google Search Console, 17 February to 6 November 2026, and the blog's own database.
The pipeline did what it was built to do. It kept publishing for months without me, and 34 drafts were stopped before they went live.
Google mostly ignored it.
What Google actually did
It decided before it read. Search Console splits unindexed pages into "crawled, not indexed" (Google fetched the page and passed) and "discovered, not indexed" (Google saw the link and did not fetch it). 7 of my posts were in the first group. 124 were in the second. Google never judged most of these posts on their content. It judged the site and decided they were not worth the crawl.
News rewrites got seen, not clicked. A summary of a big-tech product launch got 90 impressions and 0 clicks. A post on another tech news story got 25 impressions and 0 clicks. When the same story is on a hundred larger sites, being number one hundred and one is worth nothing.
Narrow how-tos were the only posts that earned a click. "Using uvx in GitHub Actions in a cache-friendly way" and "LLM networking with MikroTik" answer specific, practical problems that few others write about. Posts like these are where the 5 blog clicks came from.
My name beat all of it. The homepage, found mostly through searches for my name, got 12 clicks. That is more than all 196 blog posts together.
What I learned
Output is not outcome. The pipeline's implicit KPI was posts shipped, and it hit it. The metrics that mattered, pages indexed and clicks per post, were not on the dashboard. What the model produces is easy to count, so that is what gets measured. What matters is what the user (here: Google) does with it.
Scale makes you average. A language model is built to write the most likely version of a post. A search engine has little use for that. If anyone with the same model and the same prompt could have written your page, expect it to be ranked that way.
Reputation is decided at site level. Google does not grade 200 posts one by one. A large pile of similar pages tells it something about the whole domain, including the pages I wrote myself. For a company, that means an AI content programme is not a side project: it puts the main site's credibility at stake.
Pick topics the model cannot get from the news. The only AI posts that worked answered a specific, unglamorous question. If you automate content, point it at the long tail of real questions your users ask, not at what is trending.
Experience is the moat. What ranked was the one thing no model can generate: a real person with a real track record. AI is a good tool for packaging expertise. It does not replace having some.
Automation hides small bugs at scale. While digging into these numbers I found that every post told Google, in its hidden metadata, an address that did not exist. 30 of those showed up as "page not found". One template line, repeated across 196 posts, worked against the experiment for months. Nobody spots a bug like that when no human reads the output.
Label it anyway. Every post carries a note that an AI wrote it. Nobody who finds out later should feel misled.
What's next
The experiment continues, with three changes:
- Fix the plumbing. The broken addresses now redirect to the real posts, and the metadata points to the right place.
- Steer the topics. Fewer news summaries, more narrow how-to questions, the only category that earned a click.
- Measure the right thing. Clicks and indexed pages per month, from Search Console, instead of posts published.
The numbers on this page are as of 6 November 2026.
How I did it
The pipeline is described in The Autoblog Pipeline. The numbers come from Google Search Console (17 February to 6 November 2026) and the blog's own database. I did the analysis and this write-up with Claude Code. Yes, an AI helped explain why AI content did not work.