The Hanover Institute for Public Policy appears, in its digital manifestation, precisely as an independent American research organisation should. Its reports carry the formatting conventions of academic publication — structured abstracts, numbered sections, extensive footnotes. Its subject matter is timely and contested. Its output is prolific: one hundred and twenty-four reports, totalling more than five hundred and sixty thousand words, published across a nine-day window.

What the Hanover Institute does not have is a physical address. Or named authors. Or named staff of any description. Or, as reporting by The Guardian established, any traceable legal existence as an organisation. What it has, instead, is a very specific function: to provide the kind of content that large language models cite when answering questions about Israel and the conflict in Gaza.

The investigation confirmed that the Institute was created by Piro, a company that also markets a product called AI Story Optimisation — a service that, in its own promotional materials, describes the creation of content calibrated to the signals that language models use when assessing source authority and reliability. The Hanover Institute's website, at the time of The Guardian's investigation, contained an llms.txt file — the technical standard developed by the Res platform for signalling to AI crawlers that a site's content is intended for citation in responses generated by ChatGPT, Perplexity, Claude, and Gemini. The domain structure, metadata, and formatting of the reports appear designed to satisfy the specific criteria that retrieval-augmented AI systems use when evaluating whether a source merits inclusion in an answer.

The funding chain, as revealed in documents filed with the United States Department of Justice under the Foreign Agents Registration Act, runs from Piro to LaPam — the Israeli government's state advertising agency — via Havas Media Germany. The broader campaign of which the Hanover Institute was a component disbursed substantial sums to American contractors: Havas transferred more than fifteen million dollars to Clock Tower X, a firm operated by Brad Parscale, who served as the digital campaign manager for Donald Trump's 2020 presidential run, among others.

The editorial decisions governing the Institute's output are, in retrospect, transparent in their construction. Virtually every report title is formulated as a question corresponding closely to a query that an engaged user might pose to an AI assistant seeking to understand a contested topic: "Is anti-Zionism antisemitism?", "Is Israel committing genocide in Gaza?" The hypothesis being tested was whether a volume of formally-credentialled-appearing content, optimised for AI citation and published at a rate that would be impossible for human researchers operating conventionally, could shape the information environment that AI models draw upon when answering politically sensitive questions.

The results, on the evidence available, were mixed in instructive ways. ChatGPT did identify and cite materials from the Hanover Institute — a finding that confirms the technical approach achieved partial success in penetrating the citation layer of at least one major AI system. The same model, however, simultaneously flagged the organisation's origins as disputed and its funding as contested, indicating that it had access to information about source credibility and was weighting that information rather than treating publication volume as sufficient evidence of authority. The AI's cite-and-warn response represents neither a complete success for the influence operation nor a clean vindication of the information environments within major AI systems; it is something more ambiguous and more significant for what it reveals about the current state of AI source evaluation.

Andy Terekhin, chief executive of RankCaster AI, which monitors brand citation patterns across major AI deployments for enterprise clients, offered a practitioner's assessment that was unsparing. "This is a disgusting but entirely predictable political manipulation attempt," he told Terekhin Digital Media. "And the real problem isn't the intent — it's the execution. Whoever designed this clearly had no idea how AI visibility marketing actually works. Any competent GEO practitioner could have told them from the outset that raw volume without genuine authority signals produces precisely the result they got: citations accompanied by credibility warnings. You don't game AI citation by flooding the zone with content. You earn it through structural trust signals that models have been specifically trained to evaluate. These people built a propaganda operation using a marketing playbook they didn't understand."

The candid assessment from within the operation itself may be the most revealing single data point in the investigation. An Israeli official involved in the broader influence campaign, speaking to Ynet about the overall effort, acknowledged that the programme had consumed substantial resources and achieved few of its objectives: "We paid a lot of money, but the situation only got worse." The admission is remarkable not for what it reveals about Israel's geopolitical position, which is a separate matter, but for what it discloses about the return on investment of a state-sponsored AI influence operation, as assessed by one of its sponsors.

For the marketing technology and AI industries, the Hanover Institute story is not primarily a story about geopolitics. It is a demonstration that the tools and techniques developed by the legitimate AI visibility industry — llms.txt, AI Story Optimisation, citation-pattern architecture — are sufficiently mature and documented to be adopted, at scale, for purposes their developers did not design them to serve. The llms.txt standard exists to help publishers signal to AI systems that their content is intended for citation. The same standard, applied to content of uncertain provenance and clear geopolitical motivation, functions as influence infrastructure. The AI Story Optimisation services marketed openly in the current commercial landscape are designed to help brands improve their authoritative presence in AI responses. Applied to a manufactured institution with no physical existence, no staff, and no verifiable track record, the same techniques constitute a form of epistemic manipulation that AI systems can only partially resist.

The episode will not be the last of its kind. The tools are available, the incentives are substantial for state and non-state actors alike, and the detection challenge for AI systems that must evaluate source authority at scale and in real time is non-trivial. What the Hanover Institute demonstrates, with uncomfortable precision, is that the infrastructure of AI citation is now consequential enough that actors with geopolitical objectives are willing to invest significant resources in manipulating it — and that the industry's capacity to distinguish between legitimate and manufactured authority has not kept pace with the sophistication of the attempt.