AI reliability has a soft spot, and it isn't inside the model — it's in the sources the model reads on the way to an answer. In March 2026, a researcher named Metehan Yesilyurt launched a website packed with 50,000 statistics on SEO, content marketing, social media, e-commerce and web technology from 200 countries. He made every number up with GPT-4.1. Google quietly buried the site within weeks. ChatGPT, seven months later, still hasn't let go: that dead website has sent more than 10,000 AI sessions its way.
What happened
50,000 invented numbers for about $50
Yesilyurt bought a brand-new domain and generated 50,000 statistics pages in a single day using GPT-4.1. He never checked a single figure. He pushed the pages live with a basic AI web design template, which cost him around $30. Later he added 15,000 listicles for roughly $20 more, bringing the whole experiment in at about $50. That's it. No research team, no data pipeline, no fact-checkers — one person, one day, fifty bucks.
The trick worked because of how chatbots research. When an AI assistant needs to fill a gap in its knowledge, it goes looking for something that looks like data. Statistics are exactly that shape: a number, a year, a country, a clean sentence someone can quote. Think of it like a student writing an essay at 2 a.m. — they'll grab the first source that has the right kind of number in it, and they won't dig much deeper than the first page.
Google buried it. ChatGPT kept quoting it
At first the launch went beautifully. Using indexer tools, Yesilyurt got almost the entire site indexed within a week, and it started showing up in Google results. Then two Google updates landed back-to-back and visibility collapsed. Indexed pages now bounce between 5 and 20 impressions. A test subdomain didn't help either. Google has essentially forgotten the site exists.
ChatGPT didn't. Since launch, the site has pulled in more than 10,000 AI sessions, and more than 330 referring domains now link to it — including Shopify, Bluehost and DHL. Its invented numbers have turned up in an academic paper and in a bank's investment research. The experiment builds on an earlier Seer Interactive study from August 2025, which looked at how AI systems pick and cite sources.
What it means for you
Here's the uncomfortable part: this isn't really a story about one prankster. It's a story about where the numbers in your life come from now.
At home
You ask a chatbot how much protein is in a portion of lentils, or what a decent savings rate looks like. It answers with a confident figure and no link. That figure may have come from a page built in a day by someone who never checked it — the same way you'd trust a handwritten price tag in a shop window without asking who wrote it.
At work
Someone on your team needs a stat for a slide deck before the 10 a.m. meeting. An AI assistant hands over a confident percentage — 'most marketers say…' — and it goes straight into the deck. Nobody clicks. Nobody asks where that number came from. Six months later that slide is in a pitch you can't take back.
In business
Buyers now ask AI assistants for recommendations before they ever reach your website. If a competitor's numbers get quoted and yours don't, the chatbot has effectively made a shortlist for you — and you weren't on it. Being cited has started to matter as much as being ranked.
When you study
An academic paper already picked up these invented figures. If you're writing a dissertation, a chatbot answer is a lead, not a source. Treat a stat with no origin the way you'd treat a quote with no name attached: interesting, unusable.
For creators
Writers and video makers live on numbers — hooks, thumbnails, intros. Borrowed fake stats travel further than they should, because they're punchy and easy to repeat. The creators who'll keep their reputations are the ones who drop a link next to the number.
For your income
If you publish anything — a blog, a newsletter, a small business site — this is the flip side of the story. A plain page of well-sourced numbers can quietly become an AI citation machine for years. The barrier to entry is roughly the price of two pizzas.
How to try it right now
You don't need a budget or a developer. Start with the free version, then decide if you want to go further.
1. Test your own assistant (free). Open ChatGPT, or any free chatbot, and ask a numbers question in your niche: 'What percentage of small businesses use invoicing software?' Note the confidence level.
2. Interrogate the answer (free). Reply with: 'Where did that figure come from? Give me the exact source link.' Watch what happens. Some assistants give a real URL; some go vague; some quietly change the number.
3. Click the link. Check three things: does the page name a person, does it show a date, does it explain how the number was collected? If all three are missing, you're looking at a stat-shaped guess.
4. Reverse-search the number. Put the exact figure in quotes in Google along with a keyword. If the only pages carrying it are thin stat farms with no author, that's your warning sign.
5. Cross-check with free tools. Run the claim through MyKreaTool and its free AI tools to rewrite, summarise and stress-test the wording, then compare what you get against a second assistant. Two models disagreeing is a useful alarm bell.
6. Build the honest version. Buy a cheap domain, pick ten questions your customers actually ask, and publish ten numbers with a named source, a date and one line on method. One afternoon is enough. That's the shape of the original experiment — and it's worth noting how quickly a site like that gets ingested by indexer tools, roughly a week.
7. Track it. Once a month, ask ChatGPT and Google's AI answers the questions you published for, and see whether you get quoted.
Upsides and what changes
The upside is real and it's cheap. Fifty dollars and one day bought this experiment more AI visibility than most content teams get in a year. For a solo founder or a small marketing team, that's an enormously encouraging signal: the door is open, and it isn't guarded.
What changes is the definition of authority. In classic search engine optimisation, you fought for a ranking position. In generative engine optimisation (GEO) — the practice of getting quoted by a chatbot rather than ranked by a search engine — you fight to be the source the model reaches for. Structure beats length. A clear number with a clear date and a clear owner travels further than pages of opinion.
It also changes what 'research' means day to day. Checking a number now takes thirty seconds and one click. That's a habit worth building, because the cost of skipping it is no longer a bad grade — it's a bad decision, a bad deck, or a bad investment memo.
Limitations
Be honest about what this is: one person's experiment, published as a provocation, not a peer-reviewed study. The site was deindexed by Google, so the results are hard to reproduce cleanly, and AI answers shift month to month, so the 10,000 sessions and 330 referring domains are a snapshot rather than a law of nature. It also caused genuine collateral damage — invented figures reached an academic paper and a bank's investment research — which is exactly why running the honest version of this tactic is the only version worth running. And none of it proves that AI systems are broken; it proves they're fast, literal and trusting, which is a very human combination of traits.
Conclusion
AI reliability isn't something you can outsource to a chatbot, a search engine, or a nice-looking domain. The pipeline that fed 50,000 fake statistics into ChatGPT, into 330 linking websites, into an academic paper and into a bank's research was built in a day for about $50 — which means the pipeline that verifies what you read can be built in about thirty seconds, for free.
One action for today: the next time an AI assistant gives you a number, reply with 'show me the source' and click it. If there's nothing to click, don't put that number in your deck, your essay, or your business plan.
Comments 0