In 2023, I started an experiment: I wanted to see whether a fully AI-generated website could pass as human-made, and interact on social media as a “real website.” I didn’t finish it, but still wish to document it here. The data is now almost whole two years old, but I think it is relevant to save and share – as I write more up-to-date articles on the topic.
1. Introduction
I’ve been working with computers (and amateur programming) since before the Internet. When the Internet came, I was curious to make websites and figure that out. By the time Dreamweaver was all the rage, I had already moved on to networks.
Many years later, in 2015, I decided to build a website for my notes. I had to catch up and see which tools are good for the job (mobile-friendly was already very important aspect, to name one change from my early days). I opted for WordPress as the “least-bad” option – and I still think it’s the best, after having tried and compared countless alternatives, both dynamic (database-driven) and static (“pure HTML”) ones. And – BikeGremlin was born. 🙂
A year or two after that, I built a few websites for some friends who were just starting their businesses. Unlike BikeGremlin, where my goal was to share knowledge and information (and get corrected when wrong), with these sites one of my main goals was to rank for given keywords – and help my friends make some sales (they had some very good articles, but still my main goal was to rank, not to share information).
I did all that pretty well when it comes to ranking and “Google.” BikeGremlin websites were gaining thousands of visitors each day from Google search, and my friends were getting calls and making sales when people googled for services they were offering.
Then, in 2023, “AI” boomed. Improved, perfected, it allowed people to make thousands of articles in days, without any hard work or expertise. Sure, “SEO content writers” had been at it for years already (see the writing on “The SEO scourge” – ironic!), but this was a whole new level (of stupid).
I saw the huge power (and I must say “threat”) of LLMs and AI in general, and wanted to see how powerful it has become at the time.
Google reacted to that by practically killing website ranking, and “moving” the entire Internet towards a handful of large corporate websites (like Forbes to name one) and forums (like Reddit). To be fair, Google didn’t go full-retard: the sites I built had their Google traffic cut by 30 to 50%, not totally annihilated (like some “shill “shady” sites did). So Google didn’t do a horrible job… just pretty bad. 🙂
But it keeps getting worse – more on that in the article: What killed the open Internet?
I found this pretty ironic. BikeGremlin sites are what Google has been telling us they want to see rank highly: informative and helpful articles, well-written by an expert in the field (for details, see: “Is BikeGremlin a reliable source of information?“).
However, that also tickled my curiosity. Could I build a site with minimum effort, using just the “AI tools?” Will people believe it and interact with it on social media? Could I get it to rank? Could I make its articles be at least OK (if without any soul)? Is my “SEO knowledge” still relevant for this decade? There was a way to find out…
2. The skeleton
I registered a new domain, unrelated to BikeGremlin, set up a WordPress website, and an anonymous “writer” (to simulate “SEO content writers” who aren’t real recognized experts on the topics they write about).
3. Visual design, UI/UX
For this experiment, I didn’t want to bother Maša for the visuals and the UI & UX, nor do it myself. I wanted to rely on cheap online tools.
First things first – I needed a logo. So, after asking “perplexity.ai:” “What is a good AI logo designer?” I went to logo.com.
I picked one of the template WordPress themes and did slight alterations with my AI-generated logo colours.
4. ChatGPT 4.0 AI text generator
In January 2024, I subscribed to the paid ChatGPT to make articles for the AI experiment website be as good as AI gets at the time.
The good
It was surprisingly good (got a bit more stupid at the time of writing this, but that’s a topic for a separate article). I could give it links to some of my articles, and ask it to copy my writing style. The “output” is of a decent quality. Sure, it needed a bit of touch-up, but nothing major. And some paragraphs were better than I would have come up with. Scary!
The so-so
Having said that, after about a dozen articles, I could see a pattern. The style felt repetitive, so to speak, even if the topics differed. I am still not sure if that is good or bad. Some uniformity and standardisation can be good, but over a certain degree… McDonalds built a junk-food empire on that – food for thought (pun not intended).
The bad
With some paragraphs (usually at least one in every article), it was clear they weren’t written by a human expert in the field. I’ve also noticed some patterns, like the term “in conclusion” used way too often, for practically every article (this is fixed at the time of writing).
That’s why I still did some minor edits. That took a fraction of the time needed to write a whole article from scratch, “manually.” Still, instead of just copy/pasting and taking less than 5 minutes per article, I spent about 30 minutes per article to do some minor edits. That’s a huge time and effort save, compared to spending 4 to 40 hours per article (over 20 hours is my average).
I was curious to find out: will this be good enough for Google, and for visitors? It would have taken dozens of articles and at least a couple of months (if not a year or two) to see what Google “thought” of the “robot” website. What about human visitors? That leads us to the next chapter:
5. A social experiment to boot
The website I had built had no contact form or contact information whatsoever. The author is anonymous too (not using an alias, but really with no name).
Visitor comments were also disabled for the entire website.
People (or companies) wouldn’t thave known how to contact the author, ask questions, or comment on any articles. Would this have driven people to publicly comment on various social networks, forums etc. (and provide more backlinks to the website, thus improving its ranking)? I don’t know, but I was curious to find out – for a while. Answering visitor questions is helpful to the visitors, but apparently not very helpful in terms of website ranking, not at the time of my experiment start – nor at the time of writing this.
6. Google ranking algorithm discussion
I’ve created a “Google alert” to notify me whenever the new project is mentioned:
https://www.google.com/alerts
Of course, since I gave up on the experiment, this was of no use.
7. Giving up
I usually see things through – even if it takes years. However, this was a very rare exception where I simply gave up! Why?
There are several reasons:
- It felt empty.
Writing articles like that was not fulfiling, not inspiring. Felt like “getting it done” – like throwing out the thrash (or making a fast, crappy meal instead of cooking something delicious, with a bit more effort). - I saw it done on a mass scale on the Net, with success, across multiple websites and domains, so no need to add a 101st experiment to boot.
In 2025 – apart from countless “SEOs” bragging and showing graphs, a social study was performed on Reddit too (irony, since Reddit had started with bot/fake accounts until it gained popularity; and now has a million dollar deal with Google; and is allowed by Google to AI-translate all discussions to any language – without Google penalties that other websites would face!):
https://www.newscientist.com/article/2478336-reddit-users-were-subjected-to-ai-powered-experiment-without-consent/ - It felt wrong.
Can’t put my finger on all the exact reasons, beyond the rational ones I could list. It just felt wrong to commit to doing what I had planned, despite my curiosity.
8. Document it?
Even though I gave up, I saved most of the articles – they seemed worth preserving, explaining the basic technical stuff and terms in a short, clear way. So I copied them and re-published them here, on this very site. They are “authored” by the “author” “Ghost Writer AI” – so it is clear what they are, and you can see if they are any good (and if it is really clear that they are “AI-written” – which is getting more and more scary, but that’s a topic for a separate article).
The experiment has helped me learn how to use LLM/AI tools to make my work faster and easier, but without reducing the human part (spell checking, icons and formatting etc. – more on that in a separate article). It also showed me how scarily human-like AI can be. But I have decided to not build another AI website – and don’t like the idea of letting robots do my writing – for better or for worse.
For a deeper look at SEO, AI tools and what they’re doing to the Web, see: SEO, LLMs, and the Death of the Open Web?
There are exactly 27 such AI-written articles on this website. Those are the ones I saved. Mostly short, to the point, discussing computer hardware. This may sound ironic, and perhaps even hypocritical (another topic for a separate article), but while I see no use in saving the whole “fake website”, saving and publishing the articles publicly can show the quality (or “quality”) of the AI work and all the nuances (see: “Can we still recognize “AI” “content”?“). Also, I find those articles to be helpful in defining certain hardware specs and, with the sketches and photos I added in those articles, I’m even happy to link to them when explaining stuff! If that is not irony, I don’t know what is. 🙂 Go figure…
Last updated:
Originally published:
