The contest to influence what people see online is moving beyond search rankings and social feeds. A reported Israel-backed campaign built around a purported American think tank offers a striking example of a newer strategy: publishing large volumes of apparently authoritative material designed not only for human readers, but also to be discovered, cited and potentially absorbed by artificial intelligence systems.
According to Ynet, citing an investigation by The Guardian, the Hanover Institute for Public Policy published 124 reports between August 6 and August 14, producing more than 560,000 words in nine days. The reports addressed politically charged questions involving Israel, Gaza, antisemitism, genocide allegations, war crimes and Palestinian displacement. Their significance, however, lies as much in how the material was packaged and distributed as in the arguments it contained.
Content engineered for the chatbot era
Many Hanover report titles were framed as direct questions resembling the prompts users enter into conversational AI systems. The site also reportedly used technical infrastructure associated with making content easier for large language models to interpret. Ynet reported that an early version of the site contained an llms.txt file associated with Res, a company described as an AI-native content platform focused on helping material gain visibility in services such as ChatGPT, Perplexity, Claude and Gemini.
This approach belongs to an emerging discipline commonly called generative engine optimization, or GEO. Traditional search engine optimization aims to improve a page's visibility in search results; GEO attempts to make information more likely to appear in AI-generated answers. In ordinary commercial use, that can mean structuring useful, well-sourced material so an AI system can understand and cite it. The Hanover case illustrates the more controversial side of the same mechanism: interested actors can try to manufacture an information environment that looks authoritative enough to influence automated answers.
Why volume and apparent authority matter
The reported publication pace is central to the story. Hanover released dozens of lengthy reports within days, presenting them with the visual and rhetorical conventions of policy research. Ynet said the organization listed no physical address, named no researchers or employees and published reports without bylines. The site was nevertheless presented as a research institute, while its articles cited established organizations and statistical sources.
That combination creates a difficult problem for AI retrieval systems. A chatbot searching the live web may encounter a professionally presented report alongside journalism, academic research, government records and advocacy material. Unless the system evaluates provenance, ownership, conflicts of interest and the quality of the underlying reasoning, superficial signals of authority can carry disproportionate weight. The challenge becomes even greater when large volumes of similarly framed content reinforce one another across the web.
The campaign's sponsorship is therefore important context. Ynet reported that the site was registered with the U.S. Justice Department under the Foreign Agents Registration Act as material distributed on behalf of the Israeli government by New York media production company Piro Inc. The same reporting says Piro markets an "AI Story Optimization" service focused on producing content for the way large language models assess credibility. Israel's Foreign Ministry, meanwhile, has said that the state does not conduct influence operations in the United States and that it carefully follows U.S. law. Piro co-founder Daniel Rosenberg has characterized the company's work as an effort to place accurate, sourced facts into the public record and counter misinformation about Israel.
Retrieval today, training data tomorrow
There are at least two distinct ways campaigns of this kind could affect AI output. The immediate route is retrieval: an AI assistant with web access finds a newly published page and cites or summarizes it in response to a question. The longer-term concern is that strategically produced material could enter large web archives or datasets that later contribute to model training, potentially making its framing harder for users to trace back to a particular source.
The distinction matters because retrieval is comparatively visible. A user can inspect a citation, research the publisher and compare competing sources. Information incorporated during model training is much less transparent at the point of use. A model may reproduce a framing or association without providing a direct citation explaining where it originated. This is why provenance and dataset quality are becoming information-security questions rather than merely publishing concerns.
There are also signs that current AI systems do not simply accept such material at face value. The Guardian reported that when it asked ChatGPT for recent Hanover Institute research, the chatbot returned several Hanover links but also warned about controversy surrounding the institute's funding and origins and surfaced critical reporting. That example suggests that attempts to influence AI can collide with the broader information ecosystem: investigative reporting about an influence campaign can itself become information that AI systems retrieve.
GEO creates a new trust problem for the open web
The larger lesson extends far beyond any one government or geopolitical conflict. Companies, political organizations, lobbying groups and individual operators all have incentives to become part of the corpus from which AI systems construct answers. As conversational interfaces increasingly mediate how people research products, politics, health, history and current events, visibility inside an AI response becomes economically and politically valuable.
That does not make GEO inherently deceptive. Publishers have always adapted content to new discovery systems, from newspaper indexes to Google search and social platforms. The crucial distinction is between improving the accessibility of genuinely attributable information and manufacturing the appearance of independent authority. For AI companies, the response will require stronger source evaluation, better disclosure of provenance and more robust mechanisms for detecting coordinated content networks rather than simply judging individual pages in isolation.
For readers, the emerging rule is equally important: an AI citation is not the same thing as independent verification. A source can be real, accessible and extensively referenced while still representing the interests of a sponsor. The next phase of online influence may therefore be less about persuading a person to click a particular result and more about becoming one of the sources from which machines assemble reality. The Hanover episode shows why the provenance of AI-visible information is rapidly becoming one of the defining trust questions of the generative web.