National Cybersecurity
Association of Ukraine

What decides citation in generative search.

Analytical note ·

An October study of 468,000 citations finds that a page is cited when the engine's own search retrieved it, and a brand is named additionally when the model remembers it. For duplication monitoring that changes the question being asked.

In brief

  • The study covered two 90-day collections of answers to commercial questions: 158,099 answers and 468,037 citations.
  • Within a single website, page properties explain little. Which pages the engine's own search retrieved explains far more: 32 per cent against 2.8.
  • Every page the model cites is one its own search retrieved, for sub-queries the model rewrites on every answer and which rarely coincide with ordinary search results.
  • The model also names brands it knows even with no page supporting the statement, and prefers websites familiar to it.
  • Matching the meaning of the question works; FAQ blocks, structured data and page speed appeared effective in this study only across websites, not between pages of one site.

Two separate mechanisms

The study supports a distinction worth stating plainly: a page is cited when it is found, and a brand is named when it is remembered. These are two different routes into an answer, and they call for different work.

The first route is retrieval. It is driven by content, though not the content usually optimised for classical results: what matters is the match with a sub-query the model composed itself, in its own words, for one user's question.

The second route is what the model already holds. Controlled experiments in the study showed that a brand name reorders a list of products with identical facts. The model's memory shapes the answer even where the facts are the same.

Why this bears on duplication monitoring

The association tracks large-scale automated republication and assesses how it affects the answers generative systems give. The study sharpens what that effect consists of.

If a model leans towards what is familiar to it, then a large volume of near-identical material is not working on search results alone. It is working on the familiarity of a claim — on the chance that the model reproduces it even when no specific page was retrieved.

An awkward practical conclusion follows for anyone publishing a correction. A correction page that the engine's own search does not retrieve, for the sub-queries it actually composes, does not enter the answer at all, however accurate it is.

What follows

  • Write the title and the paragraphs the way people ask the question, not the way an internal document is named.
  • Test retrieval rather than ranking: whether the engine's own search finds the page for the wording questions actually take.
  • Do not rely on structured data and FAQ blocks as the main lever: in this study they did not explain the difference between pages of one website.
  • For corrections, keep one strong page per subject instead of scattered mentions: familiarity rewards repetition of a claim, not a count of domains.
  • Measure the result by how often the page is cited in answers, not by traffic alone.

The limits of this note

This is an account of one external study, published as a preprint in October 2026. Its samples concern commercial questions in two markets and its findings concern engines of one generation; the association has not independently verified the study's data and does not extend its conclusions to all generative systems. The note is not a recommendation regarding any product or service provider.

The content duplication study