How Generative Engines Choose Citations? James Dooley Interviews Sergey Lucktinov

James Dooley Hi, today I’m joined with Sergey Lucktinov and today’s question is how generative engines choose citations. Sergey Lucktinov The way it works is a multi stage process. It starts with fan out queries. When someone asks an LLM a complex question, that question is normalised and expanded into related or implied queries. Sometimes there are only a few fan out queries and sometimes there are many, depending on how complex the original question is. Those fan out queries are then sent to a search engine. In the case of Gemini, that is Google. In the case of ChatGPT, that is Bing. In the first stage, the system analyses metadata from the search results. This includes the website name, the page URL, the meta title, the meta description, and internal signals such as spam scores and other trust data assigned by Google or Bing. The goal at this stage is to remove irrelevant sites, so typically between forty and eighty percent of results are filtered out. In the second stage, the remaining websites are lightly checked. The system looks at domain structure, content type, relevance, and content stability, including whether the layout shifts. If a site passes this stage, it reaches the final stage. In the final stage, usually between three and ten websites remain. These are fully parsed and evaluated by an LLM. The model checks meaning, stability, and trust, then selects roughly five to eight passages that are used in the final LLM response. James Dooley If someone wants to be cited by generative engines, is there anything they can do to optimise for that, or things like prompt injection to get into AI answers? Sergey Lucktinov You can use prompt injection, but I would not recommend it. It can work in the short term, but it is likely to be penalised later because it is not user friendly from an LLM perspective. If you want long term results, you need to follow a semantic SEO approach. That means building a high quality website from a macro semantic perspective. You need proper macro pages, seed pages, and node pages that reflect fan out queries. You also need strong micro semantics, which means clarity, concise writing, and focused explanations of the entities you are describing. Cover only the topic you intend to cover and define it clearly at the beginning of the article so it is cheap and easy for an LLM to retrieve. James Dooley Some websites now have buttons for ChatGPT, Perplexity, Claude, or Gemini that say summarise this page and ask the model to cite it. Does that help future citations, or is that a form of prompt injection? Sergey Lucktinov It is closer to light prompt engineering, but it does not really help. LLMs do not have memory in that way. You cannot tell them to remember something for future use. That process simply fetches the page through an API and summarises it for a single user. It has nothing to do with the fan out process when someone else asks a different question later. James Dooley So it is more about engagement. Is there anything else you would recommend to increase the chance of being cited by generative engines? Sergey Lucktinov It depends on the goal. If the goal is a product, listicles are often effective because you do not need to rank first and you can appear across multiple sites. In some cases you do not even need your own website. If the goal is a proprietary strategy or service, then optimising your own website is critical. The most important factor is macro semantic structure. These structures work well because they mirror how LLMs organise information internally. When your content matches that structure, it becomes easier and cheaper for the model to understand. LLMs want the most stable and satisfying answer at the lowest possible cost. If your content is good but difficult to extract, it will not be cited. Speed also matters. If time to first byte is above roughly three hundred milliseconds, the page may be removed early in the process. James Dooley It always seems to come back to the cost of information retrieval, whether for search engines or LLMs. Thanks a lot, Sergey Lucktinov.