What building an autonomous website has taught me so far

I began this project with a fairly ambitious idea: a website that could collect information, turn it into useful articles and look after itself. Nearly a year later, I am pleased with how far it has come. It can monitor sources, filter large amounts of material, develop ideas and produce detailed articles. But the part I find most impressive is not the writing itself. It is what the system can do with the content that already exists.

Cheap enough to filter properly

One of the practical changes that made the project more viable was having a cheap model workhorse. GPT Luna has been useful for working through large amounts of data, discarding junk and applying filtering logic that would be awkward to express as a rigid list of rules. In my experience, the cost can sometimes be around fifty cents per website per day. That is an observation from my own use, not a benchmark or a promise about what another site will cost.

I have been very happy with the journalism workflow: the system can take ideas gathered through monitoring and develop them into detailed articles. That is a powerful capability, but the steps between a source and a published article still matter. Filtering and drafting do not, by themselves, establish that the source is accurate, that an interpretation is fair or that attribution and privacy have been handled well.

Why internal linking stands out

Internal linking has always been a laborious job. It requires remembering what exists across a site, finding relevant places to connect it and keeping those links useful as the catalogue grows. Older automated approaches often depend on exact keyword matches and rules that somebody has to prepare in advance.

Combining a vector database with an LLM changes the shape of the problem. The index can help find related content by meaning, and the model can make a small change to a paragraph so an appropriate anchor fits naturally. The result is not limited to waiting for the perfect prewritten phrase to appear. Indexing the site and running the tool can then update posts at a scale that would take substantial manual work.

I find that combination remarkably capable. The cleverness I am reacting to is in what the model and vector search can do together, not a claim that every link is flawless. I deliberately made the system edit the stored post directly and track the changes for reversal. I was worried that a bug in a content filter could affect the page at display time; having the changes recorded in the post gives me a more concrete way to inspect and undo them.

The new linking work has not been running long enough for me to establish its effect on search performance. I expect useful internal links to help readers navigate and hope they will improve the site, but that is an expectation to test, especially as search changes in the age of AI.

Convenience has a resource cost

The video builder was a useful counterexample to the idea that a working feature is automatically a viable one. I liked the prospect of starting a render from my phone and letting a server do the work. The videos did render, at roughly a minute of processing for a minute of video, but on a server without a GPU they consumed the CPU and slowed every other site on the machine.

I would have accepted slower renders if I could have limited their resource use. The point was a convenient central workspace, not the fastest possible renderer. In practice I could not make that approach viable on the shared server, so I removed the tools and started extracting the work into a separate project. I may return to it, although full automation is unlikely and that makes it less compelling for this particular website.

What still needs an answer

The project has shifted from generating content towards managing it as well. That is a broader and, for me, more interesting problem. A system that can collect sources, decide what to develop and publish at low cost also raises questions about what survives each transformation: context, attribution, uncertainty and the interests of the people represented in the source material.

I do not have a simple conclusion about what more autonomous publishing means for the web. The next useful step is to keep examining the output and the links, measure their effects where I can, and make it practical to inspect and correct the system when it gets something wrong.