Where AI answers get their sources

The most common mistake in AI search strategy is assuming the model knows things. People picture a giant memory, ask how to get their brand into it, and conclude the answer involves training data, which they cannot influence, so they give up or buy something.

That picture is wrong in a way that matters. When someone asks ChatGPT or Perplexity or Google's AI mode a question about your market, the interesting part usually is not recall. It is retrieval. The system runs a search, fetches a handful of pages, reads them, and writes an answer from what it just read. The citations you see under the answer are the pages it opened, seconds ago.

That changes the problem. You are not trying to get into a model. You are trying to be one of the pages it opens, and to be quotable once it is open.

The three pools an answer gets pulled from

A search index the vendor rents or runs. Most answer engines sit on top of a conventional web index. If a page cannot be found by a keyword query, it will not be found by the retrieval step either. Classic SEO has not stopped mattering here; it has become the qualifying round.

The vendor's own crawler. These are separate bots with separate appetites, and they are less patient than Googlebot. They tend to want HTML they can read on the first pass. A page that assembles its content in the browser can be indexed perfectly well by Google and still be a blank document to the thing writing the answer.

Whatever the user handed over. A pasted URL, an uploaded PDF, a connected tool. This pool is invisible to you and enormous in aggregate, and it is the reason two people asking the same question get differently-sourced answers.

You have real leverage over the first two. None over the third.

Citations attach to passages, not pages

Watch an answer engine work and you will notice it rarely cites a page for the page's sake. It cites a specific claim, and the citation points at the passage that supports it.

This is why a comprehensive guide that covers eleven subtopics often loses to a thin page that answers one question in its first paragraph. The retrieval step is looking for a span of text that resolves the question on its own. If your answer only makes sense after four paragraphs of context, it is not a span. It is a journey, and nothing is going to quote it.

The practical version of this:

• Put the answer in the first two sentences under the heading, then explain.
• Make headings the questions people actually ask, not the topics they fall under. "How long does a site migration take?" is retrievable. "Migration timelines" is a filing label.
• Keep the claim and its evidence in the same paragraph. Split across two, and a quoted span arrives without its support.
• Write numbers and dates inline rather than in an image or a chart alone. A chart is not text.

What you can control, honestly

You control whether your page is fetchable, whether it is legible without JavaScript, whether it states its claims in extractable form, and whether those claims are specific enough to be worth quoting. That is a short list, and it is most of the job.

You do not control which engine crawled you last, what the user pasted, how a model weighs two conflicting sources, or whether you get named in the sentence or buried in a footnote pill. Anyone selling certainty about those is selling you a story.

The gap between those two lists is why AI visibility is measured rather than predicted. You ask the engines the questions your buyers ask, you record who gets cited, and you watch that set change. It is closer to polling than to ranking.

How to check this yourself, today

Pick five questions a real prospect would type. Not your head keyword — the question underneath it.

Ask each one in two or three different engines. Record which sources get cited, in what order, and which sentence of each source the answer leaned on. That last part is the one people skip, and it is where the lesson is.

Then open the cited pages and read them for shape rather than authority. You will keep finding the same thing: the cited paragraph is self-contained, states something specific, and sits directly under a heading that matches the question. Often the site is smaller than yours.

Do that five-question sweep once a month and you have a real visibility signal, built from the same evidence the engines used. If you want it running on a schedule instead of by hand, that is what aeotrace's MCP server is for — it puts keyword, SERP, ranking and Search Console data where an agent can reach it, so the sweep becomes a prompt rather than an afternoon.

The part worth remembering

Being cited is not a reward for being important. It is a reward for being easy to quote correctly at the moment someone asks. Those are different games, and the second one is much more winnable.