Publish stats, get cited: The tactic that put my data all over the web
- The Experiment: To test how Generative Engine Optimization (GEO) differs from traditional SEO, the author created a new website (stateglobe.com) populated with 50,000 AI-generated, completely fabricated statistics across 200 countries for around $50.
- Divergence Between Google and AI Search:
- Google deindexed the site shortly after launch following core updates, dropping search impressions to near zero.
- In contrast, ChatGPT never stopped querying the site, logging over 10,000 referral sessions and roughly 72,000 real-time retrieval requests (
ChatGPT-User) in a single month.
- The Phenomenon of "Citation Laundering":
- The fake numbers were subsequently picked up, quoted, and cited by over 330 domains, including reputable brands (Shopify, DHL), academic papers, and bank investment research.
- Fabricated figures gained perceived credibility as serious institutions cited them, creating a risk that synthetic guesses eventually pollute web crawls and future model training sets.
- Why LLMs Rely on Statistics:
- Fanout Queries: When researching a topic, LLMs break prompts into sub-queries looking for hard figures (e.g., "[topic] statistics 2026").
- Relevance over Truth: Search-grounded models operate under tight latency constraints and use rerankers that prioritize passage relevance and recency over factual verification.
- Specific Grounding: Numbers provide concrete, quotable anchors to fill knowledge gaps, which models naturally prefer over vague statements.
- Best Practices for Legitimate Data Publishing:
- Structure data with one self-contained claim per sentence to facilitate passage retrieval.
- Explicitly state the source and year alongside every statistic, match page headers to common fanout queries, and use server-rendered HTML rather than JavaScript-rendered elements.
Hacker News Discussion
- The Reward for Cheating: Commenters highlight Cory Doctorow's argument that systems designed for automated agents fail because there is an inherent economic incentive to game, spam, and falsify data.
- Context Collapse in LLMs:
- Humans naturally evaluate claims based on the source's reputation, domain authority, and commercial motivation.
- LLMs strip away this critical contextual lens, retrieving text primarily by semantic proximity and surface-level relevance rather than authentic credibility.
- Challenges in Automating Discernment:
- While some suggest teaching models to evaluate provenance like humans do, others note that achieving the requisite reliability is exceptionally difficult and expensive.
- Even a 90% accuracy rate in automated truth-filtering would lead to systemic misinformation if users blindly trust model outputs.
- "Poisoning the Well" in Agent-to-Agent Systems:
- Serious concerns were raised regarding recursive contamination, where unverified AI-generated outputs get ingested into secondary content, subsequent web crawls, and future training datasets.
- Skepticism vs. Systemic Safeguards:
- A recurring debate centered on whether source evaluation must remain the end user's personal responsibility ("caveat emptor") or whether societal and technical guardrails are needed to prevent widespread data pollution.
