A synthetic think tank, an openly one-sided source library and contracts worth tens of millions of dollars reveal how public diplomacy is being rebuilt for the age of chatbots.
A website presented as an American public-policy institute published more than half a million words about Israel and Gaza in nine days. It named no researchers, listed no address and appeared to have no legal existence, but it offered the outward signs of an academic institution: a sober logo, long reports, dense citations and carefully qualified conclusions.
The Hanover Institute for Public Policy was not simply built for human readers. Its reports were framed around questions people ask chatbots, its pages included a machine-readable guide for artificial-intelligence systems, and the company behind its publishing platform sells a service designed to make content more likely to appear in answers from ChatGPT, Perplexity, Claude and Gemini.
Now the operation can be seen more clearly. An analysis of website records, source code, US foreign-agent filings and the newest Common Crawl index shows that Hanover’s material entered a widely used web dataset while the site was still publishing. It also reveals a second Piro-linked website, IsraelFact.com, which uses the same technical infrastructure but makes a strikingly different disclosure: Israel pays for this site.
IsraelFact says it presents only Israel’s official record and does not place opposing accounts beside it. Its launch marks a shift in presentation, from an ostensibly independent institute to an explicitly state-funded source library. The underlying objective appears consistent across both sites: to create large quantities of structured, citation-rich material that search engines and AI assistants can retrieve when users ask contested questions about Israel, Palestinians and the war in Gaza.
The evidence does not establish that any particular AI company has trained a model on the material or that the sites have altered a model’s underlying parameters. It does show an organized effort to influence the information environment around those systems, first by making favorable material easy to retrieve and, potentially, by placing it in datasets that can later be used to build or refine models.
A think tank built at machine speed
The Guardian reported on 26 August 2026 that Hanover had published 124 reports between 6 and 14 August, totaling more than 560,000 words. Seventy-three reports, nearly 354,000 words, appeared on 12 and 13 August alone. The volume, equivalent to several substantial books in two days, would be extraordinarily difficult for a conventional research institute to produce through normal commissioning, research, editing and review.
Hanover’s current sitemap contains 125 reports. The additional page, dated 19 August, is titled “What Would a Palestinian State Require to Function?” The sitemap’s modification dates are highly concentrated: 18 entries on 6 August, 10 on 7 August, six on 10 August, seven on 11 August, 39 on 12 August, 31 on 13 August, 10 on 14 August, three on 17 August and one on 19 August.
Almost every headline is phrased like a natural-language query. The subjects include whether Israel is committing genocide in Gaza, whether anti-Zionism is antisemitic, whether Palestinians were expelled in 1948, whether Israel deliberately starved Gaza and how allegations of torture or war crimes should be evaluated.
The format gives the pages an appearance of dispassionate inquiry. Reports cite international institutions, human-rights organizations, demographic data and legal instruments. Their recurring method, however, is to narrow or postpone conclusions unfavorable to Israel. Findings by rights groups are frequently answered with the observation that a court has not issued a final verdict. Casualty and aid figures are surrounded by caveats that leave the reader with the impression that the basic record is too uncertain to support firm judgment.
The Guardian found that a 2022 paper about antisemitic Facebook comments was cited 454 times across 88 Hanover reports, including reports on subjects only loosely connected to the paper. That repetition matters in machine-oriented publishing. A citation does not merely support a sentence for a human reader; repeated associations can also make a page look topically dense and authoritative to automated systems evaluating relevance.
Hanover’s institutional identity remains opaque. The website names no employees or authors, provides no physical location and does not identify a jurisdiction of incorporation. Its terms refer only to the unnamed state in which the institute is established. Every page does, however, carry a footer saying that the material was distributed by Piro Inc on behalf of Havas Media Germany and the Israeli Government Advertising Agency, known as LaPam, with further information available through the US Department of Justice.
That disclosure is important, but it sits at the bottom of pages otherwise designed to resemble independent policy research. A disclosure can tell a diligent reader who paid for distribution without neutralizing the authority created by the site’s name, design and research style.
Hanover has now entered Common Crawl
When Hanover was first investigated, its reports had appeared too late for Common Crawl’s July index. That is no longer the case.
A search of Common Crawl’s August 2026 index returns 123 records for the domain. Of those, 122 have a successful HTTP status, and the dataset includes successful captures of 30 Hanover research-page URLs. The captures were made on 13 August between approximately 02:12 and 03:49 UTC, while Hanover was in the middle of its most intensive publication burst and one day before Politico first reported on the campaign.
Common Crawl maintains a massive open repository of web data. Search systems, researchers and AI developers use filtered portions of it for many purposes, including the construction of datasets for large language models. Inclusion therefore moves Hanover beyond a live website that a chatbot may find during a search. It makes at least some of the material available to downstream data pipelines.
That distinction should not be overstated. A Common Crawl record is evidence of availability, not evidence of use. Model developers select, filter, deduplicate and weight data differently. Public records do not show that a named developer downloaded these Hanover pages, retained them after filtering or used them to train or fine-tune a model. Claims that the site has already “poisoned” model weights go further than the available evidence.
The immediate pathway is easier to demonstrate. AI assistants with live web search can find, summarize and cite the pages. Piro’s own commercial page now promotes the campaign as a case study, saying its material was cited by ChatGPT and Perplexity in less than a week. The company advertises content engineered for how LLMs evaluate credibility
and distribution through trusted third-party properties we build
.
Those phrases describe a new form of search optimization. Traditional search-engine optimization tries to move a page higher in a list of links. Generative-engine optimization tries to place a source inside the answer itself, where its framing may be presented in a synthesized voice and where many users will never open the original page.
IsraelFact makes the sponsorship explicit
A second website connected to the same campaign is more candid about what it is.
Verisign’s domain record shows that IsraelFact.com was registered on 20 August 2026 at 20:44 UTC. Its sitemap listed 30 reports as of 27 August: 10 dated 20 August, six dated 21 August, one dated 25 August and 13 dated 26 August.
On its About page, the site states that Israel pays for it and says, We give you one side’s record and we say that is what it is.
It explains that it relies on official Israeli sources, does not place the other side’s figures alongside them and does not independently verify Israeli bodies’ own counts.
That admission makes IsraelFact more transparent than Hanover, but not neutral. It describes a deliberately selective evidence policy. The site can use real documents and accurately reproduce official claims while still shaping conclusions through source selection, omission and the absence of adversarial context.
IsraelFact also has an llms.txt file listing its reports with compact, machine-readable descriptions. Its robots file permits automated crawlers. These choices do not force an AI company to use the material, and llms.txt remains a proposed convention rather than a universal standard. Together, however, they reduce friction for any automated system seeking to map and retrieve the site.
The timing does not prove that IsraelFact was created in response to the exposure of Hanover. Its domain was registered after Politico’s 14 August report but before the Guardian’s 26 August investigation. The available evidence supports describing it as a parallel or successor-style property within the same campaign, not as a confirmed emergency replacement.
Two identities, one publishing system
The connection between the sites is not based on subject matter alone.
When their source code was reviewed on 27 August, Hanover and IsraelFact were being served with the same Vercel deployment identifier and the same res-light theme marker. Their page structures, machine-readable files and publishing patterns closely matched. Both sites also carried foreign-agent notices naming Piro, Havas Media Germany and LaPam.
An archived 13 August capture of Hanover contained application metadata naming Res, the AI-focused content platform founded by Hai Tran. The visible reference was later removed, although the theme marker remained. Res markets itself as a platform that helps business teams earn citations from leading AI assistants.
Technical markers can change, and shared code by itself does not identify every person who commissioned or edited a page. In this case, the matching deployment evidence is reinforced by the common statutory disclosure and the public filings connecting Piro and Tran. It is a strong indication that the two sites were produced through the same publishing operation.
The contracts behind the content
Piro is a New York production company founded by Daniel Rosenberg and Tim Piper. Its company biography emphasizes experience in film, advertising and branded entertainment. The firm registered under the Foreign Agents Registration Act, or FARA, on 2 June 2026 for work ultimately benefiting Israel’s Ministry of Foreign Affairs.
The contract exhibits filed with the Justice Department contain two work orders totaling $1 million.
- A $900,000 “Digital Storytelling Pilot,” dated 30 April 2026, allocates $190,000 to strategy and content development, $80,000 to production oversight, $400,000 to production costs, $95,000 to distribution advice and evaluation, $35,000 to media consultation, and $100,000 to measurement and analytics.
- A $100,000 “Fact-Based Content Initiative,” dated 20 July 2026, provides two monthly installments of $50,000 for strategy, research, editorial development, creative work, production and distribution planning for the Israeli foreign ministry.
The filings do not assign every Hanover or IsraelFact page to a particular budget line. They do establish that Piro was retained for research, editorial production and distribution during the period in which the sites were built.
Additional FARA forms identify specialists around the project. Hai Tran registered on 22 July as a Piro subcontractor and digital strategist, reporting a flat fee of $50,000. The form states that his activities were intended to influence the American public. Ashira Solomon registered on 19 August as an on-camera host and interviewer for a flat fee of $5,000, suggesting that the campaign extends beyond text into video or interview formats.
On 23 August, Piro filed 16 items of informational material with the Justice Department. They represented 13 unique IsraelFact articles, with three duplicate filings. This was the first clear FARA paper trail tying the new brand to the campaign.
Hanover’s own funding statement raises a further question. The page says the institute was established and publishing before the relevant external funding existed. The public chronology is difficult to reconcile with that assertion. The first Piro work order is dated 30 April, Piro registered on 2 June, the Hanover domain was registered on 26 June, the second work order is dated 20 July, and the site’s earliest reports are dated 6 August.
That sequence does not prove which pot of money paid for each element, and it cannot rule out an earlier offline project. It does mean that Hanover’s public claim requires a more specific explanation than the website provides.
A much larger influence apparatus
Piro’s $1 million in work orders is only one part of a wider campaign routed through Havas Media Germany, which appears in US filings as an intermediary for at least five American firms working for Israeli interests: Piro, Clock Tower X, Targeted Communications Global, Bridges Partners and Davis Media NY.
Justice Department records show that Havas paid Clock Tower X, a company led by former Trump campaign manager Brad Parscale, $15 million in six installments between 24 October 2025 and 3 March 2026. It paid Targeted Communications Global $9,892,335 between October and December 2025. Those two payment streams alone total nearly $24.9 million.
Clock Tower X’s contract is unusually explicit about scale and purpose. It calls for at least 100 original creative assets and 5,000 variants each month, with 80 percent of the content tailored to Generation Z and a target of 50 million monthly impressions. It also requires monthly search optimization, at least 20 pages focused on priority search terms, and the deployment of websites and content to deliver GPT framing results
.
The wording moves the activity beyond a conventional advertising buy. It describes an attempt to influence both the web pages people encounter and the answers generated by AI systems that ingest, retrieve or summarize those pages.
The Guardian found that seven sites associated with the older Clock Tower X network appeared 294 times in Common Crawl’s July index. Hanover’s arrival in the August index shows another component entering the same public-data ecosystem.
The ultimate flow of public money is less transparent. The Guardian reported that Israeli procurement records showed large earlier, noncompetitive awards to Havas entities for overseas media planning and placement, including a €50 million contract awarded in March 2023 that expired in March 2024. It found no subsequently published Havas award covering the period of the recent US payments. That absence is not proof that no authorization exists, but it leaves a significant gap in the publicly visible chain between Israeli appropriations, Havas and the American contractors.
How AI influence differs from ordinary propaganda
State-funded messaging is not new, nor is the use of front groups or neutral-sounding institutions. What changes in the generative-AI era is the target.
A traditional campaign tries to persuade a person who can see the advertisement, identify its sponsor and compare it with other accounts. A campaign aimed at AI retrieval tries to become source material for an intermediary that speaks in a single, fluent voice. The user may receive a synthesis without noticing which framing came from which page.
Research into “generative engine optimization” supports the basic commercial premise, although not every claim made by vendors. A study first posted in November 2023 found that techniques such as adding citations, statistics and attributed quotations could increase a source’s visibility in generated answers, in some conditions by roughly 40 percent. The result does not prove that Hanover achieved that effect, but it helps explain the site’s dense citations and report-like format.
AI systems also use different kinds of crawlers for different purposes. OpenAI’s published crawler documentation, for example, distinguishes its search crawler from GPTBot, which may be used to collect material for model development. A page appearing in a search answer today and a page being incorporated into future training data are separate events with separate technical pathways.
This is why the most defensible description of the campaign is not that it has secretly reprogrammed chatbots. It is that contractors are manufacturing an AI-friendly evidence layer around one government’s preferred account, then distributing that layer across sites and formats designed for discovery, citation and reuse.
The tactic is selective truth, not necessarily invented facts
The most effective influence material does not need to fabricate evidence. It can shape understanding by deciding which facts are foregrounded, which legal threshold is treated as decisive, which uncertainty is amplified and which source is excluded.
IsraelFact makes that editorial logic unusually plain. It tells readers that it presents one side’s official record and does not independently check some of the state’s own counts. Hanover uses a different presentation, wrapping related arguments in the conventions of independent policy analysis.
That difference matters. IsraelFact’s sponsorship is visible before a reader evaluates its claims. Hanover’s name and structure create an additional layer of borrowed institutional credibility, even though its foreign-agent footer eventually identifies the state-backed distribution chain.
Automated writing may also explain the volume. 404 Media reported on 14 August that an AI-text detector flagged 11 of 12 sampled Hanover articles. Such detectors are not reliable enough to prove authorship, and their outputs should not be treated as forensic evidence. The publication pace is independently consistent with heavy automation, but the precise mix of generated text, templates, human editing and quality control remains unknown.
Disclosure is not the same as accountability
FARA is a disclosure law. It requires agents acting in the United States for foreign principals to register and to label and file covered informational materials. Justice Department guidance says such materials generally must be filed within 48 hours of transmission.
Registration does not mean the US government has verified the material, endorsed it or determined that every relevant legal obligation has been satisfied. Nor does foreign-government sponsorship by itself establish that a claim is false. The filings are valuable because they expose relationships that the design and branding of a website may otherwise obscure.
As of 27 August, Piro’s FARA record contained 64 documents, with the most recent batch filed on 23 August. Publication dates on the live sites are not precise enough to draw conclusions about later pages that may still fall within a filing window, may have been revised or may have been assigned dates retrospectively.
The harder accountability question concerns AI platforms. A chatbot can cite Hanover while also warning that the institute is controversial, as the Guardian found in a test of ChatGPT. That is preferable to presenting the source without context. Yet users do not consistently inspect citations, and systems can carry a page’s framing into a summary even when they decline to repeat its conclusion verbatim.
Platforms could respond by treating sponsorship, FARA notices, absent authorship and implausible publication velocity as source-quality signals. They could also show clearer provenance at the sentence level, distinguish state advocacy from independent research and retain warnings when material is summarized rather than directly cited. None of those measures requires suppressing a government’s argument; they require labeling it accurately.
The Guardian said it sought comment from Piro, Havas Media, LaPam, Res and Hanover’s published contact address but received no response before its article appeared. Piro co-founder Daniel Rosenberg had earlier told Politico on 14 August that the company had been hired to place accurate, sourced information in the public record and counter misinformation about Israel. Israel’s foreign ministry has told Ynet that the state does not conduct influence operations in the United States and complies carefully with American law.
A marketing response to a policy crisis
The campaign is expanding as Israel’s position in American public opinion deteriorates, particularly among younger voters.
Gallup reported in February 2026 that Americans sympathized more with Palestinians than Israelis for the first time in its trend, with 36 percent siding with Israel. Pew Research Center reported on 7 April that 60 percent of Americans viewed Israel unfavorably, an increase of almost 20 percentage points since 2022, with the most negative movement among younger adults.
An AP-NORC survey published on 7 July found that 31 percent of Americans described Israel’s military actions in Gaza as genocide. The figure rose to 52 percent among Democrats, while 58 percent of Democrats said the United States was too supportive of Israel.
Israeli officials have themselves questioned whether large public-diplomacy budgets are producing political results. The Guardian, citing reports by Ynet and Kan, described officials complaining that the government had spent heavily while public sentiment continued to worsen. Foreign Minister Gideon Sa’ar allocated hundreds of millions of dollars to overseas public diplomacy for 2026 and 2027.
This points to the campaign’s central limitation. Optimization can change which pages a machine finds, which source it quotes and how an answer is framed. It cannot, by itself, resolve the underlying events that are driving public judgment: civilian deaths, hunger, displacement, attacks on aid and medical infrastructure, hostage policy, settlement expansion and legal allegations arising from the war.
That is the tension running through the Hanover and IsraelFact project. Its operators are treating a political and moral crisis partly as an information-distribution problem. On its own terms, the tactic may secure citations and temporary visibility. Whether it can restore trust is a different question.
What the new evidence establishes
Three conclusions can now be stated with confidence.
- The Hanover operation did not end when it was exposed. Its site remains online, its sitemap now contains 125 reports and its material has entered Common Crawl’s August index.
- The campaign has added a second, more openly sponsored publication. IsraelFact launched in August, states that Israel pays for it, publishes only one side’s official record and is technically tied to the same Piro-linked publishing system.
- The objective is documented, not speculative. Corporate marketing and FARA-filed contracts describe content designed to earn AI citations and websites intended to affect the framing of chatbot conversations.
What remains unknown is equally important: who wrote or approved each report, how much of the prose was automated, which work order paid for each property, how AI companies are weighting the sites, and whether any model developer has included the pages in training data.
Those unanswered questions should not obscure the broader development. Foreign influence campaigns no longer need only advertisements, social accounts or sympathetic commentators. They can build entire libraries of source-shaped material for machines, give those libraries the visual grammar of expertise and wait for AI systems to carry the framing to users.
Hanover tried to look like an institution. IsraelFact is more direct about being an instrument. Together, they show a campaign adapting in public and a new contest over not only what people read, but what the machines between readers and the web are taught to retrieve, trust and repeat.
Selected sources and primary documents
- The Guardian investigation into Hanover and the wider campaign, 26 August 2026.
- Common Crawl CC-MAIN-2026-34 index results for HanoverInstitute.com, captures dated 13 August 2026; checked 27 August 2026.
- IsraelFact About page and funding disclosure, checked 27 August 2026.
- IsraelFact machine-readable content index, checked 27 August 2026.
- Piro Inc FARA registration and document index, registration 7732, checked 27 August 2026.
- Piro contract exhibits and work orders, filed 29 July 2026.
- Hai Tran FARA short-form registration, 22 July 2026.
- Ashira Solomon FARA short-form registration, 19 August 2026.
- Clock Tower X contract exhibits, filed 18 September 2025.
- Clock Tower X supplemental statement and payment records, filed 18 May 2026.
- Targeted Communications Global supplemental statement and payment records, filed 30 June 2026.
- Piro AI Story Optimization service page, checked 27 August 2026.
- Politico Influence report on Piro, Hanover and AI citations, 14 August 2026.
- Responsible Statecraft reporting on Israel’s effort to influence AI answers, 17 August 2026.
- 404 Media investigation of Hanover’s synthetic-content indicators, 14 August 2026.
- US Department of Justice FARA guidance, checked 27 August 2026.
- OpenAI crawler documentation, checked 27 August 2026.
- “GEO: Generative Engine Optimization” research paper, first submitted 16 November 2023.
- Gallup polling on American sympathies in the Middle East, February 2026.
- Pew Research Center polling on views of Israel, 7 April 2026.
- AP-NORC polling on Israel, Palestinians and Gaza, 7 July 2026.
Methodology: This report compared live sitemaps, page source and machine-readable files for HanoverInstitute.com and IsraelFact.com; queried the August 2026 Common Crawl index; reviewed Verisign domain-registration records; and examined US Justice Department FARA filings available through 27 August 2026. Publication dates shown by the sites were treated as publisher-supplied dates unless independently corroborated.
