A new website calling itself the Hanover Institute for Public Policy has published more than 100 reports about Israel and Palestine since August 6, according to reporting by Responsible Statecraft. The site looks and reads like a policy research organization, but its own disclosure says its material is distributed by advertising firm Piro on behalf of Havas Media Germany, acting for the Israel Government Advertising Agency.
The arrangement matters beyond one political campaign. Hanover's pages are dense with citations, tables, narrowly phrased questions and measured-sounding conclusions. Those are useful features for a reader checking evidence. They are also the features that make a page easy for an AI answer engine to retrieve, summarize and cite. Piro separately sells a service designed to shape what those systems say.
This is an early, unusually visible example of a political sponsor applying the logic of search optimization to generative AI. It shows how an influence campaign no longer needs to persuade only journalists, social media users or search algorithms. It can also try to become part of the source layer from which a chatbot assembles an answer.
A research institute with a named sponsor
Hanover describes its work as content analysis and computational social science focused on the causes of antisemitism in the United States. Its current About page says the reports take no policy positions and that readers can inspect the data and methods. It also says the institute does not call itself independent, nonpartisan or neutral because of its government funding.
That disclosure is important. So is the presentation around it. The reports appear under the institute's name, not individual researchers' names. Hanover says this is an editorial choice, partly because its subject attracts harassment and partly because findings should stand on their sources. The site uses the visual and editorial language of a conventional policy shop while placing the financial relationship in its About and funding material and in a footer on each page.
The resulting product is not an anonymous network of counterfeit news pages. The sponsor can be found, and the underlying registration is public. But a reader, crawler or chatbot arriving directly at an individual report may absorb its claims before examining who paid for the work. That gap between formal disclosure and practical visibility is the central issue.
Responsible Statecraft found reports framed around questions a person might put to an assistant, including the causes of Palestinian displacement in 1948, allegations of deliberate starvation in Gaza and claims about the conduct of the Israel Defense Forces. It reported that Hanover's work sometimes departs from Israeli government positions, but also repeatedly uses Israeli government bodies as sources. That mix makes blanket labels such as true, false or propaganda too crude. The more useful questions are who selected the topics, how the evidence was weighted and whether an answer system tells users about the sponsor.
What the public filing establishes
The Foreign Agents Registration Act filing is more precise than speculation about the campaign. Piro registered as representing Havas Media Germany on behalf of the Israeli Government Advertising Agency, known as LaPam. The agreement is dated April 30, 2026, and describes strategic communications and media relations work on behalf of the State of Israel. It says activities may include communications intended to influence the US public.
An attached work order calls the project a “Digital Storytelling Pilot” with a total budget of $900,000. Its scope includes audience research, content strategy, scriptwriting, asset production, distribution planning, performance analysis and program evaluation. The work order says the content is intended for digital and social platforms in the United States.
The filing does not say that Piro will manipulate ChatGPT, train a model or operate the Hanover Institute. That distinction matters. The connection to AI comes from the public websites and the reporting around them, not from an explicit clause in the contract.
Piro co-founder Daniel Rosenberg told Politico, in a statement quoted by Responsible Statecraft, that the firm's purpose was to put sourced information into the public record and counter misinformation about Israel. That is the contractor's stated case for the work. The same methods, however, can serve both public education and persuasion. Funding, topic selection and distribution determine which role is dominant.
Built for the answer-engine layer
Piro's own AI Story Optimization page removes much of the ambiguity about the technical objective. The company says it maps the sources AI systems read, identifies valuable queries, writes structured and sourced content for the way language models assess credibility, and places that material on client sites or third-party properties. It also says it monitors citations and revises content as the source landscape changes.
That practice is commonly called generative engine optimization, or GEO. A foundational GEO paper described methods for improving a source's visibility in generative answers. The field extends familiar search optimization into systems that do not merely rank links but synthesize a response from retrieved pages.
The shift changes the prize. Traditional SEO tries to win a click. GEO can influence the paragraph a user reads without visiting any source at all. A chatbot may combine several documents, remove their visual branding and present the result in one consistent voice. Even when citations are shown, funding disclosures several clicks away are unlikely to survive the summary.
Piro now presents the Hanover work as proof that this approach can produce quick citations, saying Politico found its content cited by ChatGPT and Perplexity within a week. That is a claim on Piro's marketing page, not an independent measure of sustained influence. Chatbot answers vary by prompt, location, product mode, index freshness and repeated testing. A few citations show that a page entered a retrieval pipeline; they do not show that it changed a model's underlying beliefs or consistently altered users' views.
This is not the usual meaning of data poisoning
The label “LLM poisoning” is tempting but technically imprecise here. In machine-learning security, data poisoning usually means placing malicious examples in training or fine-tuning data so a model learns unwanted behavior. Anthropic and other researchers have studied that threat directly. Hanover's immediate route appears simpler: publish crawlable documents that a search-connected assistant can fetch when answering a current question.
That is closer to source optimization than a backdoor planted during training. It can take effect quickly because no model retraining is required. It may also be less durable. Search indexes change, competing sources appear, and answer systems can alter how they rank publishers. The same page might be cited in one response and ignored in the next.
The distinction does not make the issue harmless. Retrieval is now part of the product. If sponsored political material is selected as evidence, the system's answer can inherit its framing while leaving the sponsorship behind. The risk is not that one website secretly rewrites a whole foundation model. It is that many well-resourced organizations learn to fill narrow information gaps with documents tailored to the machinery that answers questions.
Disclosure needs to travel with the claim
Publishers have always tried to look authoritative, and governments have long paid for communications campaigns. AI answer engines compress that old problem. They turn a collection of sources with different incentives into a single piece of prose, often without preserving the context a careful reader would use to judge each one.
Model providers have several practical options. They can identify government-funded and registered foreign-agent material at ingestion or retrieval time, surface that status beside citations, and reduce reliance on clusters of recently created pages with a common sponsor. They can also show which sentence came from which source instead of attaching a general list of links to an entire answer. None of those steps requires banning a viewpoint or deciding the Israel-Palestine dispute. They require retaining provenance.
There is a responsibility on publishers too. A disclosure that satisfies a legal filing regime may still be too distant for an automated summary. Sponsorship should appear in page metadata and near the report title, where crawlers and readers encounter it before the claims. Methodology and raw data should be available in forms that outside researchers can reproduce, especially when a publication presents itself as empirical.
What to watch next is measurable: whether Hanover's reports keep appearing in answers across several assistants, whether those assistants reveal the Israeli government relationship, and whether Piro files additional campaign material with the US Justice Department. The larger test is whether AI companies can preserve source context as quickly as political and commercial actors learn to optimize for their systems.