I published 95 AI-written blog posts. They got 3 clicks.
Eight months of automated blogging for OppFinder, measured honestly: 95 posts, 3 clicks from Google, 0 sales. What went wrong, what I deleted, and what I'm doing instead.
In February I set up an n8n workflow to write blog posts for OppFinder. Three times a week it picked a topic, asked Gemini for a 1,000-word article, and published it to the site. Nothing in the workflow required anyone to read a post before it went live.
I have a day job and my evenings were already full, so I didn't have time to write articles (it's also not my strength). I automated it instead, thinking the volume would compound. It didn't.
By October there were 95 posts. This week I finally sat down with Search Console and the database.
The numbers
| Feb to Oct 2026 | |
|---|---|
| Blog posts published | 95 |
| Clicks from Google to any blog post | 3 |
| Best-performing post | 77 impressions, 0 clicks |
| Visitors to the whole site (last 90 days) | 153 |
| Of those, from search | 23 |
| Accounts created | 3 |
| Sales | 0 |
Google indexed 94 of the pages, so it did see them. It just ranked someone else's page above mine almost every time. Having now read the posts properly, I understand why.
What was actually in those posts
I went through all 95 with a few simple checks.
Fifteen were "case studies" about founders who don't exist. At least two were about the same imaginary founder, Sarah. Some admitted a few paragraphs in that she was hypothetical, and some never did. One was titled "Case Study: The Solo Dev Making $14k/mo in Commercial Refrigeration." Neither the solo dev nor the $14k a month is real.
The most recent post was called "I Analyzed 100 AI Startups." Nobody analysed 100 AI startups. The model wrote that title and the workflow published it.
Posts from this month still recommend GPT-4o and Claude 3.5. The writing model only knew what was in its training data, and nothing in the workflow told it the date or checked its facts.
Many of the posts were the same post. There were three versions of "Gemini vs Claude vs GPT", three of "how to compete with funded startups as a bootstrapper" and two of "finding your first paying customers". The topic list feeding the generator kept returning the same eight niches, so the posts circled the same ideas and competed with each other in search.
And every post was signed "AI Opportunity Finder", with no person's name on any of them.
There was a technical problem as well. A trailing slash in an environment variable meant every page told Google its address was a slightly different URL, which redirected back to the page itself. That's fixed now. It wasn't what held the posts back, though. They had nothing in them worth reading.
Why this matters more for a product like mine
OppFinder exists to stop founders building on vibes. It runs grounded web searches, keeps the sources, and won't invent a market size it can't find. I spent most of September making it stricter about this: if a number can't be traced to the research, it says "Unknown".
Meanwhile the blog was publishing an imaginary founder's revenue three times a week. Anyone who found OppFinder through one of those posts and then read how the product works would have had good reason not to trust it.
What I did about it
I unpublished 61 posts: every invented case study, every post with outdated model advice, every unsourced revenue claim, and the fake study. They're archived rather than deleted, but they won't come back in that form.
I kept the other 34 for now. They had no obvious red flags but they're still generic. Most of them will go too, and the few that survive will be rewritten with my name on them.
The generator is switched off. From now on it's two posts a month instead of twelve, and each one starts from something OppFinder actually produced: a real scan with real sources, plus what I think of the result, including where the tool got it wrong.
The next two posts are written that way. One runs OppFinder across six South African industries and reports what came back. The other follows one legal tech idea through the full analysis.
If you're about to automate your blog
Automating the publishing is fine. The problem was letting a model decide what was worth saying. It can write a fluent 1,000 words on almost anything, and so can everyone else's model. The posts people click on have something the reader can't get elsewhere, like your own data, a mistake you made, or an opinion you'll stand behind. If I can't add something like that to a post, I won't publish it.
I don't regret trying it. I regret not making proper time to check the numbers regularly.
OppFinder runs grounded market research on a niche and region you choose, and shows the sources behind every claim. If you want to see what it says about your idea, try a scan.
Related Resources: