How answer engines choose what to say

A search engine ranks pages and lets the user choose. An answer engine reads pages and writes one answer, naming a few products and citing a few sources. The unit of competition has changed from a position on a page to inclusion in a paragraph.

This is why classic rank reports go quiet exactly where buyers now start: "what is the best X for Y", "X vs Y", "is X worth it". Those questions are increasingly answered inside ChatGPT, Perplexity, Gemini and Google's AI Overviews, and nothing in a ranking report shows whether you were named.

For most buying questions the engine does not answer from memory. It runs one or more web searches, fetches the top results, extracts passages that bear on the question, and writes the answer from those passages, attaching the pages it used as citations. This is grounding, or retrieval-augmented generation.

Two consequences follow. First, what you can influence is which pages the retrieval finds and whether they are quotable. Second, the loop between publishing and being cited closes in days, not in model training cycles.

Engines often split one prompt into several searches: a list of tools, their pricing, reviews, comparisons. Pages that answer a sub-question well get cited even if they never targeted the original prompt. Covering a topic's sub-questions explicitly, under headings a retriever can match, is therefore worth more than one broad page.

Generation is probabilistic and retrieval results shift daily, so the same prompt yields different answers on different days. A single screenshot proves nothing. Visibility has to be measured as a rate over repeated runs, which is the subject of the next chapter.