Publish Fake Statistics Experiment Shows LLMs Cite Non‑Google Indexed Sites

TL;DR

A site with 50,000 invented statistics was quickly de‑indexed by Google, yet ChatGPT has fetched it over 600 000 times in seven months, earning 330+ backlinks and citations in academic and financial reports. The experiment proves that LLMs retrieve content independently of Google rankings, and that statistics are a high‑value signal for AI‑driven search.

Experiment Overview – Building a Fake‑Stats Site in One Day

Result: A $50 website (stateglobe.com) generated 50 000 statistic pages with GPT‑4.1, then added 15 000 listicles. Within a week the whole domain was indexed by Google, but two subsequent Google updates caused an immediate traffic collapse.

Key metrics after 7 months:

10 000 AI sessions originating from ChatGPT‑User agents.
600 000 total bot requests (Cloudflare logs).
330+ referring domains, including Shopify, DHL, and Bluehost.
10 000 GA4 sessions attributed to ChatGPT traffic.

"In the last 30 days alone, the site got more than 600,000 requests from bots. Around 72,000 of them came from ChatGPT‑User."

Google vs. ChatGPT – Divergent Retrieval Paths

Conclusion: Ranking on Google and retrieval by an LLM are independent processes.

Google removed the site from SERPs after algorithm updates; impressions dropped to single‑digit levels.
ChatGPT continues to open the pages in real time to ground answers, regardless of Google visibility.

"When ChatGPT searches, it doesn't just copy Google's top 10. It rewrites your prompt into sub‑queries, pulls candidates from its sources and cache, reranks them, and then opens a handful of pages to ground the answer."

Real‑World Citations of Fabricated Data

Conclusion: Even wholly invented numbers can propagate into scholarly and commercial outputs.

Cited in an academic paper (see screenshot).
Used in investment research by Qatar’s largest financial group.
Referenced by an economics publication in India and a Canadian influencer research report.

These citations illustrate a citation laundering effect: a statistic starts with zero authority, gains a citation from a reputable source, and then appears more trustworthy to subsequent users and models.

Why LLMs Chase Statistics

Conclusion: Statistics satisfy a specific retrieval need for LLMs when drafting content.

Fan‑out queries – LLMs automatically append modifiers like "statistics" or "2026" to sub‑queries.
Numbers act as evidence – A concrete figure is more useful for grounding than a vague statement.
No built‑in fact‑checking – Latency and cost constraints prevent models from verifying each number; relevance rerankers prioritize match over truth.
Freshness signals – Year‑rich URLs and headings signal recency, which LLMs weight heavily.

Building Statistics Pages the Right Way

Conclusion: Properly sourced, well‑structured stats pages can earn LLM citations without resorting to fabrication.

One claim per sentence – Ensures each statistic can be retrieved in isolation.
Place source and year next to the number – Improves model relevance and human credibility.
Match headings to likely fan‑out queries – Use patterns like "[Topic] statistics 2026".
Show and maintain an update date – Freshness is a strong ranking signal for LLMs.
Serve data as server‑rendered HTML – Avoid JavaScript hiding that blocks AI bots.
Keep the design minimal – Simple lists reduce crawl overhead and focus on content.
Add original data when possible – Unique numbers can become primary sources.

Monitoring LLM Retrieval

Conclusion: Traditional SEO tools miss AI‑bot traffic; dedicated signals are needed.

Server logs / Cloudflare – Look for agents like ChatGPT-User, OAI-SearchBot, PerplexityBot.
GA4 referral sources – Track sessions from chatgpt.com, perplexity.ai.
Citation dashboards – Monitor visibility across a set of prompts (Peec’s internal tool).
Ahrefs / Semrush – Track new referring domains.
Unlinked mentions – Search exact quoted numbers to find usage without backlinks.

Ethical Takeaways

Conclusion: Publishing false data is dangerous; the experiment serves as a warning, not a playbook.

The site now carries a disclaimer acknowledging the fabricated nature of its content.
Verification of sources must become a standard step for writers and researchers.
The tactic’s efficacy may diminish as LLMs improve detection of low‑quality or self‑generated sources.

Community Reactions (Selected HN Comments)

Citation laundering analogy: "On Wikipedia it's called 'citogenesis'" – @acadapter.
Historical precedent: "Trap streets were used to catch plagiarists on maps" – @jfrbfbreudh.
Broader risk: "Fake stats can snowball into real‑world decisions" – @KronisLV.
Model behavior: "AI doesn't know or care if the data is fake" – @jqpabc123.
Future mitigation: "We need a Google for LLMs" – @tkdb.

Final Thought

The experiment demonstrates that LLMs treat the web as a retrieval pool distinct from Google’s ranking algorithm. Statistics, even when invented, are a high‑value retrieval cue, leading to real citations and potential misinformation cascades. Brands should consider publishing accurate, well‑structured stats to capture this channel, while the industry must develop robust verification and detection mechanisms to prevent citation laundering.