What happened when OtterlyAI created two websites, published approximately 1,000 AI-generated articles on each and waited to see which search and AI platforms would pay attention.
schedule
Read time:
short_text
Word count:
AI has made it possible to produce content at a volume that would once have required a newsroom, a generous budget and industrial quantities of coffee. Faced with that capability, it is understandable that some marketing teams have begun to wonder whether the quickest route to visibility is simply to publish as much as possible.
OtterlyAI recently tested that theory in unusually ambitious fashion. It created two completely new websites about capybaras and filled each one with roughly 1,000 fully AI-generated articles. The entire publishing exercise took less than a day. The sites were then monitored across Google and seven AI search platforms for almost four months.
At first, parts of the experiment looked remarkably promising. One site generated almost 150,000 Google impressions and achieved an average position of 9.7. Around 960 of its pages appeared in search. For anyone checking the numbers during those early weeks, mass production might have looked like a triumph.
Then the visibility collapsed.
One of the sites lasted around eight days before Google effectively removed it. The other continued to grow for approximately 12 weeks and then suffered the same fate. Neither received a manual penalty or a useful warning. They simply disappeared from Google’s organic results and from its AI Overviews, AI Mode and Gemini experiences.
The platforms did not agree
The most revealing part of the experiment was the difference between platforms. Google filtered the sites out, while Bing continued to index and reward the same material. Copilot cited it heavily. Perplexity initially used it and then began reducing its citations. ChatGPT crawled the websites 5,332 times and cited them precisely zero times.
The same content therefore achieved an 8% citation share on one platform and no measurable citation visibility on four others. It is difficult to imagine a clearer demonstration that ‘AI search visibility’ is not a single result. Each system applies its own retrieval methods, source preferences and quality filters.
Early traction was equally deceptive. One site enjoyed ten weeks of growing impressions before its decline. A short reporting window could easily have presented the tactic as an outstanding success, leaving the rather awkward collapse for a later slide deck.
AI content was not the only problem
The experiment does not prove that search engines automatically reject anything written with AI. Many strong articles are now produced with some form of AI assistance, including research, transcription, analysis, planning and editorial development.
The danger came from combining automation with enormous scale and very little original value. The articles reworked information the models already knew. They did not contain first-hand experience, proprietary evidence, named expertise or a distinctive perspective. Publishing 2,000 versions of familiar information created volume without creating much additional knowledge.
This distinction matters for GEO because generative search systems already have access to endless summaries of common subjects. Another general article has little reason to become a cited source. Original research, credible expert commentary, detailed case studies and useful evidence have a far stronger claim to inclusion because they contribute something the answer cannot easily reconstruct from elsewhere.
What it means for brands
Volume still matters when it is used to build genuine topical depth. A brand may need a connected body of content covering the different questions customers ask as they move from discovery to comparison and purchase. That does not require publishing on every possible variation of a phrase or filling the website with articles created solely because the technology makes it cheap.
A safer and more valuable approach begins with the knowledge the organisation genuinely owns. Customer insight, specialist experience, original data, documented processes and a clear point of view all provide material that AI can help develop and distribute. Human experts supply the substance, professional writers turn it into communication people want to read, and AI helps the process move at a useful pace.
OtterlyAI conducted its test on disposable capybara domains, which made the eventual collapse interesting instead of commercially disastrous. Most brands do not have that luxury. A business website carries years of investment, customer trust and commercial dependence. Using it to test whether 2,000 unedited AI blogs can outrun a platform’s quality filters would be a brave decision in all the wrong ways.
The experiment’s most useful lesson is wonderfully old-fashioned. Producing more work is not the same as doing better work, and copying out the textbook 1,000 times still does not prove that you understand the subject.
Source: OtterlyAI’s 2,000 AI Blogs Experiment
About the author
What we do
Frontier15 helps brands achieve visibility in the era of AI search, creating content strategies that make your expertise discoverable and trusted by both people and machines. Get in touch and ask us how we can map a GEO-informed content strategy that helps future-proof your brand.