Ask an assistant about your product category and you get back one paragraph. It sounds settled, it cites a few sources, and most brands have no idea how it was assembled.
Some people have made a business out of knowing. There is now a service category for producing content designed to be picked up and cited by language models, and at least one firm markets it under its own name.
The pitch, in the seller’s words
The clearest description of the technique comes from someone selling it rather than from any critic.
Responsible Statecraft reports a LinkedIn post by Daniel Rosenberg, co-founder of the firm Piro Inc., in which he describes asking ChatGPT, Gemini or Perplexity about a category and getting back “one confident paragraph.” Most brands, he writes in the passage quoted there, have no idea how that paragraph gets built, and his firm spent months reverse-engineering it. The same reporting quotes the firm’s website as offering to author “content engineered for how LLMs evaluate credibility,” under the label “AI Story Optimization.”
Both quotations reach us through that reporting rather than from the sources directly. The firm’s site returns no readable content when fetched, so the wording above is what Responsible Statecraft published, not something we confirmed on the page itself.
Strip away the branding and the claim is straightforward: the inputs that decide which sources an assistant reaches for can be manufactured on purpose.
What makes a source look credible to a model
The mechanics are not mysterious, which is part of the problem.
Alice Lee, an analyst at the disinformation tracking company NewsGuard, described the pattern to Responsible Statecraft: “LLMs favor concrete statistics and data, as well as strong citations and sources, which these articles all have.” She characterised the presentation as “a perfect mimicry of a typical credible American think tank, right down to the generic name, the site layout, and the red-white-blue color scheme.”
So the recipe is a neutral tone, footnotes, a table of contents, concrete figures, and a name that sounds institutional. Every one of those signals is a proxy for credibility rather than credibility itself, and every one can be produced by anyone willing to spend the effort.
What the technique looks like at scale
The reporting that brought this to light examined a site presenting itself as a research institute, publishing reports with footnotes, tables of contents and a neutral register.
The volume is the part that matters. More than 100 reports appeared in the weeks after the site began publishing on 6 August 2026, none of them carrying an author byline. Responsible Statecraft ran 12 randomly selected articles through the detection tool GPTZero: 11 came back flagged as AI-written with high confidence, one with moderate confidence.
A hundred plausible-looking documents is not an expensive undertaking any more. That is the shift. Producing a body of work large enough to look like an established institution used to take an institution.
What separates this from SEO spam
Classic search spam competed for a ranking position. A reader could still see ten results, notice that one looked thin, and click something else.
An assistant does not present ten options. It composes one answer, and the sources it drew from are compressed into a sentence or two of attribution, if they are shown at all. A source that successfully looks authoritative does more than sit near the real ones. It becomes part of the answer, and the reader has no obvious way to weigh it.
The second difference is durability. A page can be delisted from a search index. What a model has absorbed during training, or what a retrieval system keeps reaching for, does not get cleaned up on the same timescale.
What is not known
Nobody has published a number for how often this material actually surfaced in AI answers. The reporting says that assistants including ChatGPT, Gemini and Copilot have been influenced by similar sites, which is not the same as a measured result for this operation.
If someone quotes you a percentage for how effective this campaign was, ask where it came from. As far as the public reporting goes, that figure does not exist.
What this means if you are the brand
You cannot audit a competitor’s citation strategy the way you can audit their backlinks, and no classical SEO tool will show you a synthetic source that a model happens to trust.
What you can do is check what assistants currently say about your category, which sources they lean on, and whether those sources are what you would expect. If the answer includes an institution you have never heard of, that is worth ten minutes of looking into who runs it. The technique described here only works while nobody checks.
And on the temptation side, the obvious one: the same recipe is available to any brand willing to build a fake authority for itself. It is also the kind of thing that becomes a story with your name in it when a disinformation analyst opens the disclaimer at the bottom of the page.
Sources
- Responsible Statecraft (Quincy Institute), reporting by Nick Cleveland-Stout: https://responsiblestatecraft.org/israel-influence-chatgpt/ (retrieved 23 August 2026). Source of the Rosenberg LinkedIn quotation, the Piro Inc. service description, the Alice Lee quotations, the report count and start date, the funding figure, and the GPTZero results. Note: Responsible Statecraft is the magazine of the Quincy Institute, which has a stated foreign-policy position; the facts cited here are its reporting.
