None of them publishes its full rule. What can be observed is three steps: they rewrite the question, retrieve candidate pages and cite only some, generally those that best answer what was searched. Being a candidate doesn't guarantee a citation, but a GEO-optimized website has a better chance of being found and cited.
Very little. Google explains that AI Overviews and AI Mode may issue several related searches and that a page must be indexed and eligible to show with a snippet, but it doesn't detail how links are chosen. Perplexity says it gathers information from authoritative sources and links to the originals, without giving a selection rule.
Three: rewriting the question, retrieving candidates and choosing what to read and cite. You can see how they fit together in the stage map.
Almost never. In that study, ChatGPT ran the exact original question in 5.5% of 5.2 Instant runs and in none of the observed 5.3 and 5.5 Instant runs. So your page isn't competing only for the phrase the person typed, but for what the system decides to search.
No. On GPT-5.5 Instant, 44.5% of candidate URLs appeared in only one of the 20 runs. The site proved more stable than the exact page: runs often bring different pages from the same domains. Candidates that appeared more often were more likely to be cited, although the authors note this doesn't imply that publishing more pages increases citations.
They don't publish it, but there is a clue. In a KDD 2026 test bench, the reranker favored content aligned with what the query sought, rewarded answers placed early and penalized content that widened the topic too far. On what gets cited, the authors say:
We go deeper in Why does an AI care more about relevance than style?
Because that gives it a better chance of being found and cited. These steps also depend on how your website is built. That includes AIs being able to access it and content that is well structured and precise. In a KDD 2024 experiment, citing sources, adding quotations and including statistics improved page visibility by 30% to 40%. It doesn't guarantee you will appear, but it improves your chances.
There is room to improve: in our study of 39 Argentine e-commerce domains, content structure averaged 27 out of 100, the lowest score among the criteria measured.
See whether each page answers something precise, with the answer up front and developed in the body text. To find out where your site stands with AIs, analyze it with Nostos: the result shows you what to fix first.
Three. The ChatGPT study is an observational preprint and doesn't evaluate optimization strategies. The test bench doesn't reproduce a commercial engine. Neither study examined how Perplexity chooses, so we can't state that it follows the same steps.
Get your GEO score across the eight criteria backed by academic research.
Run the analysis at nostosgeo.app