How AI answers are built

An AI answer comes from two places: what the model learned during training, and what it retrieves from the web or another index at the moment you ask. Many answers mix both. Knowing which one an answer leans on tells you where your influence lies: you shape model knowledge slowly and indirectly, and you shape retrieval through pages that can be found and read today.

Model knowledge: what the assistant already believes

A language model is trained on a large body of text collected up to a cut-off date. From that text it learns associations: which brands appear with which problems, which descriptions recur, which claims are repeated by many independent sources. When an assistant answers without searching, it is drawing on those associations.

You cannot edit model knowledge directly, and changes reach it only when a model is retrained. What you can do is make sure the public record about you is consistent, specific and repeated by credible independent sources, so that the associations a future model learns are the right ones.

Live retrieval: what the assistant looks up

When an assistant searches before answering, it sends one or more queries to a search index, reads a selection of the results and writes its answer from them. This is where most short-term AEO work pays off, because the pages it retrieves are pages you can improve, and the answer often shows them as sources.

Retrieval is also where small technical problems hurt most. A page the crawler cannot render, a fact that lives only in an image, or a robots rule that blocks the bot all keep your evidence out of the pile the assistant reads from.

Fan-out: one question becomes several searches

Assistants rarely search for the exact words the user typed. A question with several parts is broken into narrower queries, each answered separately, and the results are combined. A buyer asking for the best CRM for a small sales team with a tight budget may trigger searches about small-team CRM features, pricing models, reviews and alternatives to one well-known product.

This matters because your page does not have to match the whole question. It has to be the best answer to one of the sub-questions. A clear pricing page, an honest comparison page or a specific feature page can each earn a place in the final answer by winning one branch.

Why answers change from run to run

Generation involves sampling, so the same model can phrase the same answer differently and pick different examples. Retrieval adds more variation: search results shift, the assistant may choose different sub-queries, and the sources it reads change as the web changes. Location, language, account settings and earlier messages in the conversation all play a part.

Worked example: Lumen CRM

Lumen CRM is a fictional CRM for small sales teams. Its marketer asks an assistant, “What's a good CRM for a five-person sales team that needs email tracking and costs less than our current tool?” The answer recommends three products, not Lumen, and cites a comparison article and two pricing pages.

Working backwards, the team lists the likely sub-questions: CRMs for small sales teams, CRMs with email tracking, affordable CRM pricing, and alternatives to the product they named. Lumen has a strong feature page for email tracking but no public pricing page and no honest comparison page. Two of the four branches have nothing from Lumen to retrieve. That is the gap to close first, and it is far more useful than a general wish to “appear in AI”.