A Think Tank That Doesn't Exist Wrote 560,000 Words for ChatGPT

TL;DR

A website calling itself The Hanover Institute for Public Policy published 124 research reports totaling more than 560,000 words between August 6 and August 14, according to a Guardian investigation published August 26. The institute has no legal entity in any jurisdiction, no listed address, no staff, and no named report authors. Its own funding page states that its materials are distributed by Piro, Inc. on behalf of Havas Media Germany GmbH, acting for the Israel Government Advertising Agency, registered under the Foreign Agents Registration Act as registration 7732. The part that should interest you is not the politics. It is that the entire corpus was built as a delivery mechanism for language models, and you can read the machinery yourself in the site's llms.txt.

Nine days, 124 reports, one very busy Wednesday

The publication curve is the first thing that gives it away. I pulled the site's own sitemap.xml: 148 URLs, of which 125 sit under /research/. Grouping those by the lastmod timestamps the site publishes for each page produces a shape no research organization has ever produced.

The Guardian reports 73 of those documents landed on August 12 and 13 alone, roughly 354,000 words in two days. Responsible Statecraft ran twelve randomly sampled articles through GPTZero and got eleven flagged as high-confidence AI-written, with the twelfth at moderate confidence. Nobody is pretending a research staff did this, least of all because there is no research staff to pretend about.

The tell was in the file written for machines

llms.txt is a young convention: a Markdown file at your domain root that hands a language model a clean, curated summary of your site instead of making it guess from your navigation. Useful, well-intentioned, and now demonstrably load-tested by people you would not invite to the working group.

Hanover's is 45KB and about 5,400 words, which is roughly fifteen pages of prose written for an audience of zero humans. It indexes all 125 reports with a one-line abstract each, then adds a section explaining how to read a report, which is not a thing you write for a person who has already opened the page. The Guardian and Cybernews found that the file originally did not describe the institute at all: it carried the boilerplate of Res, an "AI-native content platform" that sells getting clients cited by ChatGPT, Perplexity, Claude, and Gemini. The one file on the site written exclusively for machines was the one that named the humans.

Fetch it today and it opens with the disclosure instead:

This material is distributed by Piro, Inc. on behalf of Havas Media Germany GmbH on behalf of the Israel Government Advertising Agency (LaPam). Additional information is available at the Department of Justice, Washington, DC.

GEO is SEO pointed at the answer box

Generative engine optimization is the working name for the practice, and Piro sells it under the label "AI Story Optimization," promising content "engineered for how LLMs evaluate credibility." Res founder Hai Tran registered with the Justice Department in July as a Piro subcontractor on the account.

Read the artifacts and the design intent is not subtle. Nearly every report title is phrased as a question, because a question is what a user types. The prose carries footnotes, tables of contents, a stated methodology, explicit source ranking, and a flat affectless register. The robots.txt is four lines long and allows everything. This is not content written to persuade a reader. It is content shaped to survive a retrieval filter.

Here is the uncomfortable mechanic. Picture a librarian who has never once left the building and decides what to trust purely by counting footnotes and checking that the tone stays even. You have not defeated that librarian's reasoning. You have flooded its intake. A model ranking sources by surface markers of credibility is doing exactly that, and surface markers are cheap now: 560,000 footnoted words cost nine days and a work order.

What a corpus looks like when it only has one thing to say

The llms.txt groups its 125 reports into five topic sections. The distribution is its own disclosure.

What the site says for itself

To its credit, the funding page does not hide behind much. It concedes that an earlier version claimed the institute operated without external funding, notes that "the earlier wording remains findable in search results and archives," and states plainly: "The registration is why this site does not call itself independent, nonpartisan or neutral." It also asserts the funder "does not select topics, does not see drafts, does not review findings, and approves nothing before publication."

Piro co-founder Daniel Rosenberg told the Guardian the company was hired to place "accurate, sourced facts" and counter misinformation. None of the outlets covering this, Gizmodo included, independently measured how often chatbots actually cite Hanover. Ynet reported that querying ChatGPT about the institute's research returned links to four Hanover reports, alongside a flag about controversy over its funding. That is one probe, not a study, and I am not going to inflate it into one.

What this means if you build on retrieval

The campaign is disclosed, filed, and now publicly reported, which makes it the least dangerous version of this technique. The dangerous version is the one nobody registers under FARA.

• Treat llms.txt as an assertion, not a fact. It is an unverified, self-authored summary that you are inviting into your context window ahead of the page it describes. That is a privileged position. Log which domains you honor it from.
• Domain age and entity existence are cheap signals you are probably not using. A site whose entire corpus postdates a single fortnight should not carry the same weight as one with a decade of archives, and "does this organization legally exist" is a resolvable question.
• Publication velocity is a usable feature. One hundred and twenty-four long reports in nine days from an entity with no staff is a signal available from a sitemap parse.
• Corpus concentration matters. A neutral research body does not put 88 percent of its output in one topic bucket.
• Stop rewarding formatting. Footnotes, methodology sections, and an even tone were once expensive proxies for rigor. They now cost a prompt.

Key Takeaways

• A site with no legal entity, address, or named authors published 124 reports and more than 560,000 words in nine days, with 70 of its 125 research pages carrying lastmod dates of August 12 and 13.
• Its own funding page discloses the chain: Piro, Inc. for Havas Media Germany GmbH for the Israel Government Advertising Agency, under FARA registration 7732.
• The site's 45KB llms.txt reportedly began life carrying the boilerplate of Res, a vendor that sells getting clients cited by ChatGPT, Perplexity, Claude, and Gemini.
• Question-shaped headlines, footnotes, stated methodology, and a permissive robots.txt are optimization choices, not editorial ones.
• GPTZero flagged eleven of twelve sampled articles as high-confidence AI-written, per Responsible Statecraft.
• No published study has measured how often chatbots actually cite this corpus. The technique is documented; the yield is not.

Sources: The Guardian, hanoverinstitute.com/llms.txt, Hanover Institute funding disclosure, hanoverinstitute.com/sitemap.xml, Ynetnews, Responsible Statecraft, Gizmodo, Arab News, DOJ FARA unit, llmstxt.org