"Publish fake stats, get cited in LLMs": how AI answers get gamed, and what AI search companies say

Update, October 23, 2026: Google says AI Overviews can't be turned off; our guide covers the Web filter and the settings you can change.

"Publish fake stats, get cited in LLMs" is how a Hacker News submission summed up a blog post by Metehan Yesilyurt, a researcher at Peec AI, posted on October 20, 2026, that had 366 points and 206 comments by October 23. The post itself, published October 7, describes building a website of 50,000 AI-generated statistics that its author never checked, then watching ChatGPT send it traffic and other websites cite it. The experiment is real and the author is candid about it, but its evidence is narrower than the headline. It fits a pattern in AI search that published research and earlier incidents have documented.

This analysis is based on the Peec AI post, the GEO paper (KDD 2024), Google's Search Central documentation and May 2024 AI Overviews post, OpenAI's and Perplexity's help centers, and studies by NewsGuard, Columbia's Tow Center and the European Broadcasting Union, checked October 23, 2026. It builds on our analysis of websites blocking AI agents, which covers the user-triggered fetchers, such as ChatGPT-User, that this experiment relied on.

The short version

• What was done: one person generated 50,000 statistics pages with GPT-4.1 in a day for about $30, on a new domain, and didn't verify any of them. He calls the data "fake."
• What's reported: 10,000+ ChatGPT sessions since March 2026, about 72,000 ChatGPT-User requests in 30 days, links from 330+ sites, and citations in an academic paper and investment research.
• What isn't shown: citation counts for Google AI Overviews, Gemini, Perplexity or Claude, or a measured rate of how often ChatGPT repeated the numbers.
• Why it isn't a one-off: a 2023 paper found that adding statistics to a page lifts its visibility in AI answers, and propaganda networks have fed chatbots false claims at scale.
• What companies say: Google bans mass AI-generated pages under its spam policy; OpenAI tells users to check the cited sources; Perplexity says its source labels don't certify accuracy.

What did the Peec AI experiment actually do?

Yesilyurt says he bought a new domain, stateglobe.com, and "generated 50,000 statistics pages in a single day with GPT-4.1," at a cost of "around $30." He later added 15,000 listicles, for a total of "about $50." The pages claimed to cover SEO, marketing, social media, e-commerce and web technology across 200 countries.

The numbers weren't researched. "Every one of them was fake," he writes, and "I never checked a single one." Strictly, they were generated by a model and never verified, not deliberately falsified; the author calls the result fake data, and nothing in his account suggests any figure came from a real measurement.

He says he started building the site in March 2026. Peec's post says the site "now carries a disclaimer." The site's homepage, checked October 23, 2026, describes its numbers as "2026 estimates and projections based on industry trend analysis and historical data." It doesn't say they were generated by AI and never checked.

Which AI assistants cited the fake stats, and how was it measured?

Only ChatGPT figures are reported. Yesilyurt's evidence comes from his own analytics, not from a controlled study:

• Google Analytics: "Since March, GA4 shows 10,000+ sessions from ChatGPT alone." A session is a visit referred from ChatGPT, not a count of answers that cited the site.
• Cloudflare logs: more than 600,000 bot requests in 30 days, about 72,000 from ChatGPT-User, OpenAI's agent for fetching pages during a conversation. A fetch shows ChatGPT retrieved a page, not that it quoted it.
• Peec's visibility dashboard: on October 7, the site ranked "second for visibility" behind Statista on "a few prompts." The post doesn't say how many prompts or which ones.
• Google: after "two Google updates came out right after launch," the site's Google visibility "collapsed," and indexed pages drew about 5 to 20 impressions.

The post reports no citation data for Perplexity, Google AI Overviews or AI Mode, Gemini, Claude or Copilot. Its central claim is that ChatGPT's retrieval kept using pages that Google had largely stopped showing: "Google may have forgotten about the site, but ChatGPT clearly hasn't."

The spread beyond AI answers is the more striking part. Ahrefs and Semrush show "links from 330+ other sites," Yesilyurt writes, including Shopify, DHL and Bluehost. He names four uses: "an academic paper," "investment research by Qatar's largest financial group," "an economics publication in India," and "a research report from a Canadian influencer community." He doesn't link to them, so they can't be checked from the post. He calls the effect "citation laundering": "The fake data certainly didn't become true, but every citation made it look more trustworthy."

Is gaming AI answers new?

No. In a paper first posted in November 2023 and presented at KDD 2024, researchers from IIT Delhi and Princeton named the practice "generative engine optimization" (GEO). They had GPT-3.5 Turbo rewrite web pages nine ways, then measured how much of a generated answer drew on each page, on a benchmark of 10,000 queries.

Share of AI answers drawn from a rewritten page (GEO paper)

Position-Adjusted Word Count, overall, GEO-bench test set (higher means more of the answer cites the page)

Quotation Addition27.2

Statistics Addition25.2

Cite Sources24.6

No optimization (baseline)19.3

Keyword Stuffing17.7

Show the data

Rewrite method	Position-Adjusted Word Count
Quotation Addition	27.2
Statistics Addition	25.2
Cite Sources	24.6
No optimization (baseline)	19.3
Keyword Stuffing	17.7

The three methods that added evidence worked best. On the paper's Position-Adjusted Word Count metric, a baseline page scored 19.3; Quotation Addition scored 27.2, Statistics Addition 25.2 and Cite Sources 24.6. Keyword stuffing, the old SEO trick, scored 17.7, below doing nothing. The paper sums up the top three as a "relative improvement of 30-40%." On a 200-query test of the live Perplexity.ai, Statistics Addition improved a second metric, Subjective Impression, by 37%.

The limits matter. These results come from the researchers' own test engine built on GPT-3.5 Turbo and the top five Google results, plus a small Perplexity sample, in 2023. Statistics Addition "modifies content to include quantitative statistics instead of qualitative discussion, wherever possible," and the paper describes no step that checks whether the added numbers are true. It shows models favor pages with numbers in them; it doesn't test fabricated numbers.

Have AI answers repeated planted or bad data before?

Yes, in several documented cases, each with its own method and limits.

Google's May 2024 post is the clearest company admission. It traced the rock answer to "satirical content on this topic" republished on a geological software provider's website, and said it had "limited the inclusion of satire and humor content" and user-generated content in answers that "could offer misleading advice." NewsGuard said 10 leading chatbots repeated false narratives laundered by the Pravda network 33% of the time, and that the network, which it credits with 3,600,000 articles in 2024, was "deliberately infiltrating the retrieved data of artificial intelligence chatbots." The Tow Center's test of 8 AI search tools found wrong source details in more than 60% of 1,600 queries; tools often cited syndicated copies rather than originals, and Perplexity answered 37% incorrectly against Grok 3's 94%. In the EBU study, journalists at 22 public broadcasters rated more than 3,000 answers from four assistants and found serious sourcing problems in 31%.

What do OpenAI, Google and Perplexity say about source quality?

Google says pages appear in AI Overviews and AI Mode under the same rules as Search: a page must be "indexed and eligible to be shown in Google Search with a snippet," and there are "no additional requirements" or "special optimizations necessary." Its spam policies, updated August 28, 2026, list "scaled content abuse," including "using generative AI tools or other similar tools to generate many pages without adding value for users." By Yesilyurt's account, Google's systems did largely drop his site.

OpenAI's help article on web search says ChatGPT "ranks search results using multiple factors intended to help users find relevant, reliable information," but doesn't list those factors; a site becomes eligible by letting OpenAI's search crawler in, and "placement is not guaranteed." The same article says search results and citations "can be incomplete, outdated, or incorrect" and tells users to open a cited source to check that it supports the answer. Its "Does ChatGPT tell the truth?" article lists "fabricated quotes, studies, citations or references to non-existent sources" among known failure modes and tells users to "verify important information from reliable sources."

Perplexity says it gathers information "from authoritative sources like articles, websites, and journals." It also marks some citations with source labels (Government, Academic or Trusted) based on a review of whole websites. It says a label doesn't certify any article's accuracy, most sites have none, and "no label does not mean a site is low quality."

None of the three documents a check on whether a retrieved statistic is true. That matches the Peec post's own diagnosis: rerankers "score relevance, but not truth."

What the evidence supports, and what it doesn't

Supported: AI search tools can retrieve and send traffic to pages Google has largely dropped; pages with statistics, quotations and citations get used more in generated answers, at least in the GEO paper's setup; and once numbers are repeated by other websites, they gain the look of authority.

Not shown: how often ChatGPT repeated StateGlobe's numbers in answers, whether other assistants did, or whether the effect lasts. The Peec data is self-reported by a company that tracks AI-search visibility, and the four named citations aren't linked. The author himself says the results "were extreme" and that "LLMs might close this gap."

How can you check a statistic in an AI answer?

The same habits that work for any number apply, and AI answers make the first step easy because they cite sources:

• Open the citation. If it's a statistics page, look for who collected the number, how and when. StateGlobe's pages offered "estimates and projections."
• Find the original. A number that appears on many sites may trace back to one unchecked page; that's the "citation laundering" Yesilyurt describes.
• Be wary of sites that cover everything. A single site claiming statistics for every topic in 200 countries is a warning sign in itself.
• Compare assistants. Different tools retrieve different sources; our comparison of Claude, ChatGPT and Gemini covers which plans include search. If you build with search APIs, our comparison of AI web search APIs covers each one's citation rules.

What's confirmed, reported and uncertain

• Confirmed: the Peec post's description of the method, cost and author's own admissions; the GEO paper's results; Google's and OpenAI's published policies; the HN discussion on October 20, 2026.
• Reported (self-reported by the author): 10,000+ ChatGPT sessions, 72,000 ChatGPT-User requests, the Statista comparison, 330+ linking sites and the four named citations.
• Uncertain: whether ChatGPT still cites the site, whether other assistants ever did, and how widely the numbers spread without links.