Should you be original or say what everyone else says to be cited by an AI?
Published on
Say what reputable sources say first, then bring what you alone know. With Claude, a page whose content overlaps with that of other pages found is cited up to 1.5 times more often, and a page found through several different searches for the same answer is cited in 88 to 100% of cases depending on the corpus, compared to 50 to 60% for the others. Originality alone does not get a page chosen. But once the page is selected, proprietary data, figures, and cited sources are what set it apart.
For what type of question? It all depends on the question. On an informational or market question, the AI looks for the common ground among sources. On a question about you, your product, or your data, it is your source it will look for, and originality becomes the answer.
What we measured: consensus wins with Claude
For each page found by Claude, we measured how much its vocabulary overlaps with that of other pages retrieved for the same answer. The pages most redundant with the rest are cited 1.5 times more often in HVAC and artificial intelligence, 1.4 times across a dozen sectors in French; the effect is still present in English, though weaker. When removing pages from the same site, the effect still holds, from 1.3 to 1.5 times. And among the cited pages, the share of vocabulary unique to the page is lower than among the discarded pages.
A second signal points in the same direction. When the same page is retrieved by two different searches for the same answer, Claude almost always cites it: 100% of cases in HVAC, 93% in artificial intelligence, 93% across a dozen sectors in French, and 88% in English, compared to 50 to 60% for a page found only once. A page that answers multiple facets of the question, using the expected words, is found more often and selected more often.
Domain vocabulary in the paragraph
Consensus is also visible at the level of the cited passage. With Claude, a paragraph dense in vocabulary specific to the domain being covered—the terms that sources on the subject share—is cited 3 to 13 times more often than a paragraph containing none, depending on the corpus. The AI selects the passage that speaks the language of the subject.
This does not mean copying others. A paragraph can use the common vocabulary of the field and say something precise, dated, and supported by figures. This is what the most cited pages do: they align with what the domain says, then they build upon it.
What research says about originality recipes
The paper that coined the term GEO, presented at the KDD conference in 2024, tested content modifications in a simulated generative engine based on GPT-3.5 across a benchmark of 10,000 queries. Adding citations, statistics, and sources increased visibility by up to 40%, whereas keyword stuffing brought almost nothing. On Perplexity, under real-world conditions, the best measured gain was 22%. In this context, the page was already provided to the engine: the study measures the page's share in the response, not its chances of being found.
A benchmark published at NeurIPS in 2025, C-SEO Bench, conversely concludes that most conversational optimization methods are largely ineffective and often have a negative effect, that gains decrease when multiple actors adopt them, and that traditional SEO strategies are significantly more effective. The two findings do not necessarily contradict each other: enriching a page helps once it is read, but does not get it found.
Why AI prefers what is shared
An AI that provides answers puts its credibility on the line. Retaining the common ground of several sources reduces its risk of being wrong, and an isolated page asserting something different from all the others is a risk. With ChatGPT, this behavior takes another form: the model names the sources it deems established in its search and cites those first.
Researchers from the University of Toronto measured in September 2025 a systematic bias in AI engines toward independent and recognized media, to the detriment of brand websites and social networks. For a company, being original on its own website therefore matters less than being consistent with what third-party sources say, and being mentioned by them.
What Google calls "banal" content
Google insists on the opposite approach. Its guide on optimizing for generative AI, updated on July 10, 2026, recommends "non-commonplace" content: it contrasts content based on common knowledge, which could come from anyone and adds little, with content that provides an expert's point of view or lived experience, going beyond the ordinary. According to Google, content that people find unique, compelling, and useful will likely influence a site's presence in its AI responses.
Our measurements and this recommendation do not contradict each other. With Claude, cross-referencing the domain's vocabulary helps a page to be found and retained; with Google, repeating what everyone knows adds nothing. A useful page does both: it speaks the language of the subject, and it provides what others do not say.
What this changes for a business
A page that wants to be cited begins by covering what the domain says about the topic, using its vocabulary, in developed paragraphs. It then provides what it alone can bring: a measurement, a dated figure, an experience, a method. Originality comes after the foundation, not in its place.
This section follows this structure: each page first presents what other studies publish, supported by sources, then our own measurements with their context. The consensus helps find and retain the page; our own measurement is what deserves to be cited.
What we do not know
Vocabulary cross-referencing and multiplicity are measured with Claude, which reads entire pages; with ChatGPT, we have not formalized these measurements. Our reports describe associations on the models' API, not the effect of rewriting a page.
Sources
- Aggarwal et al. — "GEO: Generative Engine Optimization", KDD 2024, .
- Puerto et al. — "C-SEO Bench: Does Conversational SEO Work?", NeurIPS 2025, .
- Chen, Wang, Chen, Koudas (University of Toronto) — "Generative Engine Optimization: How to Dominate AI Search", .
- Google Search Central — "Optimizing your website for generative AI features on Google Search", .
- FaireDuBruit measurements, from September 4 to 16, 2026 — see "How we measured".