QWE AI Academy The #1 mistake after “publish fake stats, get cited in LLMs” blew up
The story ripped through Hacker News and timelines: about $50, tens of thousands of invented statistics, and ChatGPT still fetching them months after Google mostly dropped the site. The #1 mistake is treating that as a green light for more fiction – or the opposite shrug, “AI is broken, nothing I publish matters.” Neither moves the lever.
The lever is retrieval shape. Models hit a gap, rewrite the ask into fanout queries packed with “statistics,” “data,” and a year, then snag a few passages that look specific. Under normal latency and cost budgets they do not run a second truth pass. Package real, citable numbers the way those queries already look. Inventing more pages is the wrong half of the lesson.
Why numbers win fanout (the experiment is proof, not a recipe)
A vague claim is hard to paste into an answer. “X% of Y did Z in 2026” is already a passage. Year in the title, URL, or H2 reads as freshness. Rerankers score relevance, not audit trails.
That is what the March 2026 stateglobe.com run accidentally stress-tested. Metehan Yesilyurt (Peec AI) spun roughly 50,000 fabricated stats pages in a day with GPT-4.1 plus a template, then about 15,000 listicles – generation around $30, listicles around $20, total ~$50. Google went hard, then visibility collapsed to something like 5-20 impressions. Live answer fetching did not.
Turns out the boring logs were the story: Cloudflare later showed 600,000+ bot hits in 30 days, ~72,000 from ChatGPT-User; GA4 passed 10,000 ChatGPT sessions; 330+ referring domains showed up in Ahrefs/Semrush (Shopify, DHL, Bluehost among them). On Peec’s own tracked-prompt dashboard the domain sat #2 behind Statista on October 7, 2026. Linked downstream hits included an academic paper, Qatar-linked investment research, an Indian economics outlet, and a Canadian creators report. The site now carries a disclaimer; the author frames it as research, not a content farm manual. Full write-up is on the Peec AI blog.
None of this is brand-new GEO theory. The GEO paper (Aggarwal et al., arXiv:2311.09735, KDD 2024) already put Statistics Addition, Quotation Addition, and Cite Sources among the strongest lifts – about 30-40% on their generative-engine metrics. Fake mass pages only proved how hungry the pipe is when every sentence is a snappy number.
Watching a zero-authority domain outlive classic deindexing while serious outlets repeat the figures still feels unsettled. Citation laundering is an old citogenesis pattern; the speed is what changed.
Step-by-step: one real stats page retrieval can use
Build one cluster you can defend. Not a 50k fiction grid.
- Pick one tight cluster. “Remote work adoption statistics 2026 by country” or “[niche] market size and conversion benchmarks.” Same shape as fanout, not a lifestyle essay title.
- Collect only verifiable numbers. Primary reports, government releases, your own product instrumentation. Lock source + year per figure before you write prose.
- One claim per sentence. A retriever may snip a single line. “According to [Source] (2026), 34% of mid-market teams run weekly async standups.” Nested setup paragraphs die in the cut.
- Mirror fanout in headings. H2/H3 like “[Topic] statistics,” “market size,” “adoption by country,” “year-over-year change.” Densest answers high on the page.
- Show a real last-updated date and refresh it when numbers move. Stale chrome with fresh claims is a smell humans notice even when models do not.
- Server-rendered plain HTML. No only-percentage locked in tabs, images, or client widgets. Latency budgets favor boring markup.
- Watch retrieval separate from Google rank. Log live answer agents, re-run a fixed prompt set, fix passage clarity before you chase more referring domains.
Pro tip: Reuse a number in a deck or blog? Keep the original source URL beside it. If a model cites you later, an editor can check the trail in under a minute.
Peec later shipped a clean real-data page on their own domain and again saw AI-written sites link over. Format transfers; fiction was never the requirement.
Edges that bite (read these before you ship)
ChatGPT-User ≠ training crawlers. It fires on demand for live answers. OpenAI’s own bot notes say user-initiated fetches may not honor robots.txt the same way as GPTBot or OAI-SearchBot – so blocking the training/search bots does not stop answer-time pulls. Details live in the OpenAI bots documentation. Check CDN/WAF allowlists too; a “silent” 403 looks identical to “we were never relevant.”
JS burial. Known URL, invisible stat: if the only useful % lives in a tab, canvas, or client-rendered chart, passage fetch often returns empty chrome. Put the sentence in HTML.
Google-quiet, AI-loud. A page can sit at 5-20 Search impressions and still rank near the top of tracked AI visibility while collecting tens of thousands of ChatGPT-User hits. That is not a bug in your GSC mental model – it is a different system.
Laundering after the first cite. Once a number lands in a paper or bank note, later humans and models see institutional backing. Those pages can re-enter crawls and training. Hygiene up front beats arguing with a citation graph later.
Here is the awkward part nobody prices in: independent verification across everyday prompts is expensive. Even if flagship modes grow stricter, most answer paths still grab a few candidates and write. So what is the ground truth when earlier laundered figures already litter the open web? Your source-beside-number habit is the boring answer that still works.
How this stacks against other GEO moves
| Approach | Effort | Risk | What it buys |
|---|---|---|---|
| Fake mass stats site | Low $ / high moral + legal | Very high (disclaimer territory, trust collapse) | Short-term retrieval proof only |
| Real stats page (passage-ready) | Medium research | Low if sourced | Repeatable citations + links from AI writers |
| Quotes + cite-sources only | Low-medium | Low | Strong GEO lift in academic tests; less raw number-hunger |
| Generic long blog, no numbers | Medium | Low | Weak match for research fanouts |
Want brand mentions inside answers? Pair one tight stats page with clear entity naming and comparison copy. Want to be the URL others paste? Extractability beats word count. Start with a cluster you already measure internally – cheaper to defend, faster to update when the quarter closes.
Common pitfalls
- Useful percentage three scrolls down inside a narrative.
- No year and no source, so the snipped line looks orphaned.
- Rate-limiting or blocking user-triggered answer agents while allowing everything else.
- Scaling fiction because one HN thread went nuclear – the public write-up and disclaimer say the opposite.
- Shipping charts without a plain-language sentence the fetcher can quote.
FAQ
Is publishing fake stats a viable GEO tactic?
No. Research with a public disclaimer is not a growth playbook. Downstream laundering is the reason to refuse the fake half.
Do I still need Google rankings if ChatGPT keeps fetching quiet pages?
Different jobs. Classic search still feeds discovery and some of the corpora models lean on; live retrieval can keep using a URL Search buried. Practical case: almost no Search Console impressions, yet the URL shows up in answer citations – improve passage clarity and fetcher access first, do not delete the page.
Will better models just fact-check every number?
Maybe in high-stakes modes. Everyday paths still budget for a handful of fetches and a relevance score. Full independent audit at internet scale is the expensive option, which is why teams skip it. HN threads under titles like “Publish fake stats, get cited in LLMs” kept looping the same gap: fact-check against what, once laundered figures already look institutional? Source, year, and primary links next to the claim beat hoping the model becomes your auditor.
Pick one topic you already measure. Ten self-contained stat sentences, source and year on each, fanout-style headings, clean HTML, allowlist the live fetchers you care about. Same five prompts in ChatGPT a week later. That is the loop.