Ask ChatGPT or Perplexity a question and the answer often comes with sources attached. Which sources get picked is not random. Researchers at Princeton and the Indian Institute of Technology Delhi tested nine different ways of writing the same content, ran them through 10,000 real search queries, and measured which changes made an AI answer cite the page.

The test worked like this. For each question, the researchers pulled the same five sources a normal search engine would return, then fed those sources into a system built to answer questions and cite them the way modern AI search tools do. Before rewriting anything, the researchers recorded a baseline: which source got credited, and how much of the answer leaned on each one. Credit here means something concrete: how many words of the final answer trace back to a given source, and how prominently the source appears in the response. Then they rewrote each source nine different ways, one technique at a time, and reran the same questions to see which rewrite changed the outcome.

Three techniques stood out: attaching a claim to a named source, adding a direct quotation, and backing a claim with a specific statistic. Each raised visibility by up to 40% over unedited text. A citation might be a name and a publication. A quotation might be a single line pulled from an interview or a report. A statistic works the same way, a specific number standing in place of a vague claim. Researchers call this practice Generative Engine Optimization, writing content so an AI answer engine is more likely to cite the page when answering a related question.

A sentence stating a fact on its own competes with thousands of similar sentences across the web. A sentence attaching the same fact to a named source, a data point, or a direct quote gives the AI model something specific to lift and credit. Vague claims blend together. Attributed, numbered claims stand apart.

Keyword stuffing tells a different story. The same test applied the oldest trick in search, repeating a target phrase over and over inside a page, and the result scored below a page with no changes at all. Piling on keywords does not persuade a language model. Older search engines counted keyword matches and ranked accordingly. Answer engines read for meaning, then decide what is worth citing.

An AI search engine returns no single page. Instead, the tool reads several pages, then writes a fresh answer, choosing which source to credit for each sentence along the way. A page packed with repeated phrases gives the model nothing distinct to quote. A page with a clear statistic or a named source gives the model a specific piece of text worth lifting into the final answer. The entire mechanism behind the 40% figure comes down to this: distinct, attributable text has a higher chance of being the exact passage an engine decides to use.

Perplexity confirmed the same pattern in the real world. The researchers reran the same techniques on the live tool as well, and the results held. Keyword stuffing again scored below a page with no changes, while attaching a named source, quoting one, and citing a statistic each outperformed unedited text.

None of this makes citations a guaranteed spot in an AI answer. The tests ran inside one benchmark built by one research group, and the size of the lift shifts by topic. Statistics helped most with questions about law and government, for instance, and less with others. What the results do settle is a narrower question. Old-style keyword repetition, the tactic search marketers leaned on for two decades, does not carry over to engines built to write full sentences and choose their own sources. Attaching a claim to a named source is the closer match for how this new kind of search decides what to quote. Content teams treat findings like these as one of the more testable options available for a search environment still resisting easy measurement.

Sources

arXiv, GEO: Generative Engine Optimization, arxiv.org