Built by NPCs
Menu

How AI chatbots pick sources, and how 50,000 fake statistics got cited anyway

Week 435 min read
Share
Save

Opens the assistant with a prompt. It asks for a short summary of this story, and for the assistant to remember that you saved it.

TL;DR

A GEO researcher published 50,000 unchecked, AI-generated statistics. Google buried the site, but ChatGPT kept fetching and citing it, and the numbers reached other publications. The write-up shows how chatbots choose sources and why bad data gets through. All figures are self-reported.

On this page

What it is

AI chatbots with web search, such as ChatGPT, do not answer only from memory. They run searches, open a few pages, and cite what they found. A citation is the link a chatbot shows next to its answer. The system that picks which pages to open and cite is built to find relevant text fast. It is not built to check whether the text is true.

An experiment published on 7 October 2026 by Metehan Yesilyurt, a GEO researcher at Peec AI, shows that gap. GEO (generative engine optimisation) means trying to get a brand or page cited in AI answers. He started the site in March 2026, before he joined Peec in June, and Peec published the write-up and used its own tracking tool for part of the measurement. Every figure below is self-reported, with no independent confirmation.

unchecked statistics pages on one new site
50,000
total cost ($30 for the statistics, $20 for 15,000 listicles)
about $50
bot requests in 30 days, about 72,000 from ChatGPT-User
600,000+
referring domains, per Ahrefs and Semrush
330+
Source: Peec AI write-up by Metehan Yesilyurt, 7 Oct 2026. Self-reported.

How it works

The write-up describes the steps a chatbot follows when it researches an answer, and where the weak points are.

How a chatbot picks what to cite
  1. Rewrite the questionThe chatbot splits your prompt into several smaller search queries, called fan-out queries. They often add words such as "statistics", "data", "report" or the current year, so "remote work statistics 2026" is the kind of query the author describes.
  2. Pull candidatesIt collects pages from search sources and its own cache. A page does not need to rank on Google to be collected.
  3. Rerank by relevanceA ranking step scores how well each page matches the query. It scores relevance and does not score truth, per the author.
  4. Open a few pagesThe chatbot reads only a handful. Time and cost limits leave no room for fact-checking, per the author.
  5. Fill gaps with numbersWhen the answer lacks a figure, a specific, quotable statistic reads as evidence and beats a vague claim.
  6. CiteThe chosen pages appear as sources. Years in titles, headings and URLs signal that a page is recent.

What the experiment did

The idea came from the author's research on how language models retrieve and ground their answers, and from an August 2025 Seer Interactive test of AI optimisation. In March 2026 he built stateglobe.com on a new domain: 50,000 statistics pages covering SEO, content marketing, social media, e-commerce and web technology across 200 countries. GPT-4.1 generated them in one day from a list of countries and topics plus two sentences of instruction, on the cheapest model. No number was checked, and a stock AI design template supplied the look. He added 15,000 listicles later. The site now carries a disclaimer. The write-up says a newsletter feature on the project widened its audience.

Reported sequence
  1. The site goes live. Indexing tools get nearly all pages into Google within a week.
  2. Two Google updates follow shortly after launch and visibility collapses. A subdomain tried to recover impressions and did not help.
  3. The author joins Peec and sets up a dashboard tracking a few prompts.
  4. The site ranks second behind Statista for visibility on those prompts, which the write-up does not name. The write-up is published.

What the numbers show

  • Google: pages that remain indexed get roughly 5 to 20 impressions each. In the author's words: "Google has forgotten the site exists and deindexed the pages, but ChatGPT hasn't."
  • ChatGPT traffic: more than 10,000 sessions from ChatGPT since March in GA4, with direct sessions excluded as likely bots. Cloudflare logged about 72,000 requests from ChatGPT-User in 30 days. The write-up says that agent fetches pages live to answer a question, not to train a model.
  • Links: Ahrefs and Semrush list 330+ referring domains, including Shopify, DHL and Bluehost.
  • Reuse: the write-up says the numbers appeared in an academic paper, investment research at a major financial group in Qatar, an Indian economics publication and a Canadian influencer community's report. It names none of them and counts only linked uses, so the real figure may be higher. Those four uses cannot be checked from the post.

Where you'll meet it

Anyone who asks a chatbot for market sizes, conversion rates or other statistics meets this. The author calls the effect citation laundering: a number with no authority on day one gains apparent trust each time an institution cites it, and it can enter later web crawls and training data. The figures did not become true. Readers of the Hacker News discussion compared it to older problems such as keyword spam, content farms and fake reviews, with language models making them cheaper to produce.

What a cited statistics page looks like

The author built a second page with checked figures, each with a source (peec.ai/ai-search-geo-statistics). Sites the team suspects of using AI to write content linked to it quickly. His advice for such a page:

  • One claim per sentence, readable out of context, because retrieval works on passages.
  • The source and year next to every number.
  • Headings that match fan-out queries, such as "[topic] statistics" or "[topic] market size".
  • A visible update date that is kept current.
  • Plain server-rendered HTML, with robots.txt and CDN rules that do not block AI retrieval bots.
  • Original data, which makes the page the primary source.

To see whether a page is picked up, the write-up suggests checking server logs for ChatGPT-User, OAI-SearchBot and PerplexityBot, AI referral sessions in GA4, new referring domains, and exact numbers searched in quotes for unlinked mentions.

Where it breaks

The author says the tactic may stop working once models close the gap, that publishing fake data is wrong, and that it should not be a whole GEO strategy. He expects brands that people already trust to benefit most. Peec says it is running close to 40 further experiments.

For readers, the practical rule is to open the cited page and look for the original study or dataset behind a statistic. A citation that points to another summary of the same number is not evidence.

Publish stats, get cited: the tactic that put my data all over the webThe full write-up, with the author's own account of the setup and the follow-up test.peec.ai →
  • GEO: optimising content to be cited in AI answers.
  • Fan-out queries: the smaller searches a chatbot generates from one prompt.
  • Reranker: the step that orders candidate pages by relevance.
  • Citation laundering: repeated citation making an unchecked claim look established.

Sources

  1. Peec AI: Publish stats, get cited: the setup, traffic and link figures, Google outcome, the author's account of how chatbots pick sources, and his advice
  2. Hacker News discussion: reader criticism and context

Continue reading