Google deindexed the fake stats. AI kept citing them

Someone published a website of 50,000 fabricated statistics and watched AI assistants cite it for seven months. The site is still online. By its own banner, every page was generated with gpt-4.1-nano in under thirty minutes for less than ten dollars.

The lesson most people took from this is "AI hallucinates, check your sources." That is not what happened, and the real mechanism is worse. The model did not hallucinate. It fetched a real page and quoted it accurately. What failed is the layer that used to decide which pages deserve quoting in the first place.

What is established, and what is not

This distinction matters more than usual here, because the story is about unverified numbers.

Verified, because anyone can check it: the site exists, it serves hundreds of thousands of invented statistics, and it openly describes itself as an experiment. We checked it directly on 20 October 2026.

Self-reported, and explicitly not confirmed: the traffic and citation figures — seven months of ChatGPT citations, 600,000 bot requests in thirty days, 72,000 of them from ChatGPT-User, downstream appearances in an academic paper and in investment research at a large financial group. Those come from the author's own write-up. The one outlet that covered it noted plainly that his measurements "have not been independently verified."

So treat the stunt as an illustration, not as evidence. The evidence comes later, and it is peer-reviewed.

Ranking and retrieval are two different filters

Google did its job. Within months of launch the site collapsed in search — by the author's account, effectively invisible for most of its life.

It kept getting cited anyway.

That is the part worth understanding. For twenty years, search ranking absorbed enormous adversarial pressure. Every spam farm, every link scheme, every doorway page was an attack, and the defenses that grew up around them turned ranking into something it was never designed to be: a trust filter the whole web quietly depended on.

An AI assistant answering a question does not consult that filter. It dispatches a retrieval agent to fetch pages at answer time. Ranking is one input among several, and a buried domain is still a fetchable URL. We moved the consumption of web content off the one system hardened against exactly this attack, onto a layer with no equivalent hardening — and we did it in about two years.

The warning is on the front door. Nobody enters through the front door

Here is what the coverage missed, and it is the detail that turns a stunt into a structural problem.

That candid "this is an experiment" banner sits on the homepage.

The pages that actually get retrieved — the individual statistic pages — carry no warning at all. We fetched two of them and searched the markup: zero occurrences of the disclosure.

There is, with some irony, a page reporting Brazil's "AI Hallucination Rate" at 8.5%. It is invented. So is the 72% adoption rate beside it, and the USD 1.3 billion in R&D.

Now look at what else those pages carry.

A methodology statement. Named institutional sources — the ITU, the World Bank, Statista. A review date. A research team with a bio. A structured FAQ answering the exact question a person would type.

Every one of those is a trust signal. Every one is fabricated. The page is wearing the full costume of rigor, and the costume costs nothing to sew.

The uncomfortable part

Read that list again, then read any guide to optimizing content for AI answers. They are the same list.

Lead with a direct quotable answer. Add an FAQ. Cite named sources. Publish specific numbers rather than vague claims. State a methodology. Show a byline and a review date.

That advice is not wrong — it is how you make genuine work legible to a machine. But it describes a format, and a format is cheap to counterfeit. The signals we taught retrieval systems to reward are now the signals an attacker produces for ten dollars. Specific numbers are the worst offender: a model grounding an answer prefers "8.5%" over "a small share," because precision reads as sourcing. Precision is not sourcing.

The scale is measurable, and it has been measured

This is where the evidence stops being anecdotal.

A May 2026 study audited 111 million references across 2.5 million papers on arXiv, bioRxiv, SSRN and PubMed Central. Its conservative estimate: 146,932 hallucinated citations in 2025 alone — references to work that does not exist. The fabrications clustered in fields adopting AI fastest and among early-career authors, and they disproportionately credited already-established scholars (Zhao et al.).

That is academic publishing, which has peer review, reputational stakes and a culture of checking references. Your company has none of that around the deck someone is assembling this afternoon.

What this costs you, in two directions

Inbound — the numbers your team brings back. If anyone researches with an AI assistant, some share of what they return is sourced from pages nobody evaluated. The failure is quiet: a plausible figure with a real link behind it. Nothing looks wrong. The number goes in the board deck, the pricing model, the market-size slide.

Outbound — your content strategy. If "publish specific numbers, get cited" works as cheaply as this suggests, it works for everyone, including people with reason to misrepresent your market. Being cited is no longer a proxy for being credible, for you or your competitors.

What to do about it

In the order we would add them:

1. Treat a citation as a lead, not a verdict. This is the whole fix in one line. A link is an invitation to check, and it always was — the retrieval layer kept the link and removed the person who follows it. Put that person back for anything load-bearing.
2. Require the primary source for any number that enters a decision. Not the page that states the figure: the organization that produced it. If the trail ends at an aggregator, the number is not sourced.
3. Know which of your numbers are load-bearing. You cannot verify everything. You can list the handful of figures your pricing, roadmap or fundraising actually rests on, and hold those to a different standard.
4. In your own products, separate retrieval from authority. If you ship anything that browses, cites or does RAG, you are now responsible for the trust layer Google used to supply for free. Keep a source allowlist for claims that matter, store provenance with every retrieved fact, and show the user where a number came from instead of laundering it into clean prose. The model supplies the text. It does not get to confer authority on it — the same boundary we argued for with AI-written code.
5. Say what you could not verify. Internally and publicly. A report that marks its unverified claims is more useful than one that reads uniformly confident, and it is the only honest way to write about something like this.

None of that is novel. It is ordinary source discipline, which went unpracticed for a decade because a ranking algorithm was absorbing the cost. That subsidy has ended, and the bill now lands inside your company — in the research your team trusts and in the systems you ship. It is the kind of unglamorous boundary work we end up doing whenever AI-built software has to survive production.

A citation was never proof. It was a lead to verify.

Key takeaways

• A website of 50,000 fabricated statistics, generated for under ten dollars, kept being cited by AI assistants long after Google buried it in search.
• The model did not hallucinate — it retrieved a real page and quoted it correctly. Search ranking and AI retrieval are different filters, and only one of them was hardened by twenty years of adversarial pressure.
• The site's honest "this is an experiment" disclosure appears only on its homepage. The statistic pages that actually get retrieved carry a fabricated methodology citing the ITU, World Bank and Statista, a fake research byline and a review date.
• The trust signals AI-optimization guides recommend — FAQs, named sources, specific numbers, stated methodology — describe a format, and formats are cheap to counterfeit.
• Measured scale: 146,932 hallucinated citations in academic papers in 2025, from an audit of 111 million references.
• The defensible rule is old: a citation is a lead to verify, not proof. Retrieval kept the link and deleted the person who follows it.

Frequently asked questions

Did the AI invent these statistics?

No, and that is the point. The figures were published on a real, fetchable webpage. The assistant retrieved it and reported it accurately. The failure was in source selection, not generation — which means better factuality training does not fix it.

Why did removal from Google search not stop the citations?

Because ranking and retrieval are separate systems. An AI assistant fetches pages at answer time rather than deferring to search position, so a domain that ranks nowhere is still reachable. Cached crawls and copies on other sites extend the reach further.

Are the traffic and citation numbers from this experiment reliable?

They are self-reported by the person who ran it and have not been independently verified — the outlet that covered the story said so explicitly. The verifiable parts are that the site exists, serves fabricated statistics, and omits its disclosure from its inner pages. The independently measured evidence of the broader problem is the 2026 audit of 111 million academic references.

What is the single highest-value habit to adopt?

Requiring a primary source for any number that enters a decision — the organization that produced the figure, not the page repeating it. Most fabricated statistics fail that test immediately, because the trail ends at an aggregator with no underlying study.

Does this mean we should stop optimizing content for AI answers?

No. It means being cited is not evidence of being credible, for anyone. Keep making genuine work legible to machines, and stop reading citation counts as a quality signal — yours or a competitor's.

Sources: Peec AI — the author's write-up · Runtime Wire · Zhao et al., arXiv (2026) · Hacker News discussion