Sixty Eight Entries In: What the Machines Taught Me

01Measured the same way, on purpose

At entry thirty I published a scoreboard and promised to repeat it the same way later, including the parts that looked bad. This is that repeat. Same method: fifteen days of the server log, requests to the entry pages only, counted by the agent that made them. Same rule: nothing rounded up, and anything I could not measure is named rather than guessed. I set out to be quoted. What I learned is how long it takes before anyone quotes you.

One thing changed in between

Before the numbers, the caveat that matters most. In September I moved the commercial and how to entries to another site I run, because this blog sits on a country domain that search engines geotarget to one small region and it has almost no links. By entry sixty eight, nineteen of the sixty eight entries live there, some of them among the most fetched pages this blog ever had. So the two windows below do not cover the same pages. The comparison is still honest, but it is not like for like, and anyone quoting it should say so.

02Read more by indexes, read less by answers

The good news is the boring news

Bing is still the most diligent reader here, 172 fetches against 158 in August, which matches what the month one scoreboard found: the engine people forget is the one that reads hardest. Amazon is steady. Total traffic to the entries barely moved, 1,358 requests then, 1,381 now. If I only reported totals, this would look like a stable blog.

The bad news is the important news

Look at who actually came. The training crawler from OpenAI fell from 106 visits to 20. Google’s crawler fell from 45 to 15. Perplexity’s crawler went from 32 to zero. And the agent that fetches a page because a real person just asked ChatGPT a question, the distinction at the heart of the guide to who is reading your site, went from 24 visits to zero. Nobody’s live question led an assistant to this blog in fifteen days. Meanwhile 546 requests came from declared scanners and scripts. Total traffic hid the only change that mattered: the readers who quote stopped coming.

Search tells the same story more bluntly. Over six months Bing recorded two impressions for this blog, one a search for the domain itself and one for my name. An earlier fifteen day audit of the whole site found 1,938 page views, of which 497 survived removing declared bots and one agent pretending to be an old phone, and exactly one visit referred from another website. I did not pull Google’s figures for this window, so they are left out rather than estimated.

03The machines were clear, I was slow

What I keep and what I change

I would keep the method: first party numbers in sentences, a bespoke diagram per entry, and gates that stop a dash, a vague claim or a named property from shipping. I would not keep the promise of one entry every day, because I broke it three times and had to catch up in batches, which is what nobody reads your blog warned about from the other side. A schedule you cannot keep teaches readers and crawlers that your dates mean nothing.

Quality is not the same as reach

The larger lesson is the one I resisted for two months. Content quality is necessary and not sufficient. The same writing that sits unread here is fetched live, every day, on a site with links and a neutral domain. Being worth quoting and being findable are two separate jobs, and the second one is not done by writing. So this blog stays what it has become, the archive and the first person record, and the guides go where machines already look. An archive nobody reads is still an archive, and the record is the part I would not trade.

If you are running something similar, the three numbers worth tracking every fortnight are the ones in the chart: live fetches from assistants answering questions, visits from the training crawlers, and requests from the search crawlers that feed them. Totals will reassure you when they should not. Live fetches are the closest thing to a citation you can see from your own server, and when they stop, the cause is almost always upstream of the writing.

Tomorrow: what happened when I built a server for AI agents, and what nine days of their requests looked like.