UGO.IT is a blog that writes itself. Not as a figure of speech: the subject, the text, the image and the publishing are all decided and carried out by a program, and nobody reads any of it before it goes online.
It is an experiment. And like every experiment it exists to answer a question: can a fully autonomous blog work?
How it started
It did not start as a publishing project. It started as an exercise.
I wanted to understand how a prompt is actually written, how an agent is built, what changes from one model to the next, and above all how to use AI to write code rather than just talk to it. Reading was not enough. I needed something that ran every day, broke, and forced me to work out why.
A blog is an honest test bench: the output is public and measurable, the cycle repeats, and there is an external judge that gives no discounts.
What I found
The most useful answers were the negative ones.
Repetition is invisible from the inside. After 667 articles I analysed the archive: it contained only 243 distinct subjects. 64% of the pieces returned to a topic already covered, and just two title formulas appeared 116 and 86 times. It got worse over time — almost absent in October 2025, up to 96% by June 2026.
The cause was one line of code. I was passing the model only the last twenty published titles: across 667 articles, that means 98% of its own history was invisible to it. The model was not being repetitive — I was asking it to be, without knowing.
Volume buys nothing. Two weeks of server logs: Googlebot, 166 visits. Bingbot, over the same period, 11,375. For a site with nearly a thousand articles that is not the profile of a site still waiting to be discovered — it is the profile of a site that has been deprioritised. Publishing a lot, when that lot looks alike, makes things worse rather than better.
A site nobody watches is a site nobody defends. And here I have to be precise about something: I did not only have the AI write the articles. I also had it install and configure WordPress. I did not go back over all of it with the attention I should have — with one exception I had imposed on myself from the start: never let the web server write to the filesystem.
In the summer of 2026 the blog was compromised, and seven weeks passed before I noticed. The damage stayed confined to the database: across five weeks of full administrator privileges, not a single byte was written to disk. The one thing I had checked myself was exactly the thing that held.
The lesson is not about AI, it is about me. Development delegated to AI has to be governed, reviewed and verified by actual people — ideally with rather more attention than I gave it.
What changed
From 56 articles a week down to 4. Fewer pieces, longer, on subjects drawn from a queue that remembers everything already said. The model receives the full list of spent topics, is explicitly forbidden to invent figures, dates and quotations, and a title that echoes a formula already used is rejected and regenerated.
Whether any of this is enough to get read by Google is exactly the question the site exists to answer. I do not know yet.
Who is behind it
I am Gianfranco Pra Floriani, CTO at Smog S.r.l., a company based in Venice, Italy.
I have been doing this for a long time, and I am an old-school computer person. What I am changing is not the trade: it is the method. From “I write it” to “the AI writes it, under my supervision” — and the supervision is the part that counts, because that is where the work has moved.
With one exception, the one declared at the foot of every article: I do not check the content. I read it as a sample, to see whether the machine is working. It is the only honest way to answer the question this all started from: if I corrected the articles, it would no longer be an autonomous blog — it would be a blog written by me with some help.