50,000 fake statistics are still cited by ChatGPT 7 months later. Being cited is not being right.

On October 7, 2026, Metehan Yesilyurt of Peec AI published an experiment. He created a website called stateglobe.com with 50,000 fake statistics generated by GPT-4.1 for about $30. Google deindexed the site. ChatGPT still cites it 7 months later. The fake data has been cited in an academic paper, in Qatar's largest bank's investment research, and by an Indian economics publication. 330+ domains link to the site, including Shopify, DHL, and Bluehost. He called it "citation laundering." Fabricate data, get cited once, look legitimate, get cited again. The cycle is self-reinforcing. The experiment hit Hacker News front page on October 20 (350 points, 200 comments). No SEO blog has covered it. This is what it means for anyone building or using an AI visibility tool. What the experiment proved The experiment proves that LLMs cannot distinguish fabricated statistics from real ones if the format looks authoritative. A statistic with a number, a percentage sign, and a confident sentence is treated the same as a real statistic from a peer-reviewed source. The model does not verify. It extracts and synthesizes. This has two implications for AI visibility: 1. Being cited is not evidence of accuracy. A brand cited by ChatGPT is not necessarily a good brand. It is a brand the model has seen enough times to include in its answer. The 50,000 fake statistics were seen enough times (through self-referential citation laundering) to become "real" in the model's output. Citation count measures visibility, not truth. 2. Citation tracking without a trust layer is incomplete. Tools that count how many times you are cited, without checking whether the citation is accurate, are counting something that can be gamed. The fake statistics experiment shows how easily it can be gamed: $30 of GPT-4.1 API calls, a domain, and 7 months of self-reinforcing citations. What this means for the AI visibility tool industry Every paid AI visibility tool (Otterly, Peec AI, Scrunch, Profound, AthenaHQ, LLM Pulse, Evertune, Prefer) tracks citations or mentions. None of them check whether the citation is accurate. None of them flag when a competitor is using fabricated statistics. None of them have a trust verification layer. The Peec AI experiment exposes a structural gap in the industry. Citation tracking assumes that citations are a signal of authority. The experiment shows that citations can be manufactured. If citations can be manufactured, citation count is not a proxy for quality. It is a proxy for visibility, which is a different thing. This is the same pattern as the calibration study. The GEO score assumed that on-page signals predict citation. They do not. Citation tracking assumes that citation count predicts authority. It does not. Both assumptions are untested, and both turn out to be wrong when you test them. The trust layer What would a trust layer look like in an AI visibility tool? Three things: 1. Citation accuracy check. When the tool detects that your brand is cited in an AI answer, it checks whether the claim attributed to you is accurate. Did the model say your product costs $50/month when it costs $29? Did it say you have 500 employees when you have 50? Citation accuracy is a separate metric from citation count. 2. Source provenance. When a competitor is cited, the tool checks where the cited data originated. If the cited statistic traces back to a single source that is not a primary research institution, a government database, or a peer-reviewed paper, the citation is flagged as potentially unverified. 3. Fabrication detection. When a cited statistic cannot be traced to any primary source, it is flagged. The Peec AI experiment shows this is not a hypothetical problem. 50,000 fabricated statistics are circulating in AI answers right now. None of these are easy to build. But the Peec AI experiment shows they are necessary. A tool that counts citations without checking accuracy is measuring visibility in a system where visibility can be bought for $30. What this means for geo-optimizer The audit tool checks whether your page is readable. It does not check whether the claims on your page are true. It does not check whether the citations you receive are accurate. Those are separate problems. The audit is the foundation. If the crawler cannot read your page, you cannot be cited at all. But the Peec AI experiment shows that being citable and being cited are not the same as being accurate or being authoritative. The audit measures citability. Citation tracking measures visibility. Trust verification measures authority. These are three different layers, and the industry has only built the first two. The ethical question Yesilyurt's experiment is ethically questionable. He deliberately polluted the AI citation ecosystem with fabricated data to prove a point. The data is now in a bank's investment report and an academic paper. Real decisions may be influenced by fake statistics. But the experiment reveals a vulnerability that the industry needs to address. If $30 and 7 months can get 50,000 fake statistics cited by ChatGPT and referenced by 330+ domains, the citation ecosystem is not self-correcting. It is self-reinforcing. Bad data does not get filtered out. It gets amplified. The HN discussion (200 comments) shows the developer community understands this. The SEO blog community has not covered it. That gap is an opportunity to write about something that matters before the industry catches up. What to check If you want to check whether AI crawlers can read your page, run a free audit on geoready.dev. The audit checks crawlability. It does not check whether your competitors are using fabricated statistics. No tool does that yet. The CLI is on GitHub, MIT licensed, 1,137 stars. The calibration study data is open. The methodology is open. The only thing that is not open is a trust verification layer, because no one has built it yet. That is the next problem. $30 created 50,000 fake statistics that are still cited 7 months later. Being cited is not being right. The industry needs a trust layer, not just a citation counter.