What an AI-Run Content Network Actually Produces: 130 Days of Data

Most numbers you read about AI content production are projections. These are not. This is the actual output of a network of six sites where the writing, the cover images, the build and the deploy are done by autonomous agents on a schedule,

Most numbers you read about AI content production are projections. These are not. This is the actual output of a network of six sites where the writing, the cover images, the build and the deploy are done by autonomous agents on a schedule, with a human in four specific places and nowhere else.

Everything below is counted from the repositories themselves. Where a publication date could have been backdated for SEO, we ignored it and used the date the file first entered version control instead, which is the date the work was actually done.

The network

Property	Live posts	Produced on	Median words
A (daily cadence, EN)	105	45 days	1,662
B (bilingual, EN + DE)	80	16 days	892
C (long-form, EN)	40	16 days	2,387
D (long-form, EN)	38	14 days	2,441
E (regional, DE)	15	7 days	1,530
F (regional, DE)	5	2 days	1,380
Total	283	47 days	1,609

The properties are deliberately unnamed. What is being measured is the production pattern, and naming the sites would add nothing to it.

283 live posts. 507,909 words. Produced between 18 April and 25 August 2026, on 47 distinct days of actual production inside a 130 day window.

In the last 30 days alone the network produced 126 posts.

The number nobody publishes: the abstention rate

Property A has the longest daily history. Its agent is scheduled every single day. Over the last 30 days it published on 26 of them.

That gap is the most useful number here. The agent is instructed that if no topic clears both of its gates, it writes nothing at all and reports two lines saying so. Roughly one day in seven, that is what happens.

An agent that publishes every single day it is asked to is not a well-configured agent. It is an agent with no gate, and thin daily content at volume is what gets a site deindexed. The abstention rate is the quality control, and it should be visible in your logs.

Cadence is bursty, not smooth

The naive mental model is one post per site per day, forever. The real production histogram is lumpy: days with a single post, days with eight or ten across the network, and dead days.

Two things cause the lumps. Multi-language properties produce in pairs, so property B emits an English and a German version of the same topic in one run, which reads as "two posts" in any count. And a new property starts with a seeded launch batch written in one sitting, then drops to routine cadence. Of the 283 posts here, 71 came from routine days and 212 from seeding batches. Anyone quoting a per-day average without separating those two is quoting a number that means nothing.

What the human still does

Four jobs, and they have not gone away:

1. Setting the gates. What counts as a publishable topic, what counts as thin. This is a judgement call and the agent inherits it from you.
2. Verifying claims. Every named statistic or company event has to link to a primary source. This is enforced as a hard gate before save, because it failed once: a post shipped citing three companies and one report with a single external link in the whole piece.
3. Approving anything that carries someone else's name. Non-negotiable.
4. Auditing. Even agents auditing themselves miss glitches, in AI-generated images and in logical breaks in a draft. Someone with domain knowledge has to read the output and ask confirmation questions.

Method, so you can check the work

Post counts come from the content directories of each repository. A post counts as live if it has a publication date in frontmatter and is not marked as a draft.

Two dates exist for every post and they are not the same thing. The frontmatter date is what the site displays, and on seeded launch content it is deliberately backdated. The production date is the first commit that introduced the file, or the filesystem timestamp where a property is not under version control. Every claim on this page uses production dates. A day is counted as a routine day for a property when that property produced fewer than three posts on it, which separates scheduled agent runs from human-triggered seeding batches.

We publish the counting method because a content statistic without a method is a press release. If you run something similar, you can produce the same table from your own repositories in an afternoon, and we would rather compare notes than compare claims.