Overview

When an AI assistant answers a question and shows links, it has made a series of decisions: what to search for, which pages to read, which to trust and which to quote. Nobody outside the companies behind AI search engines can see those decisions in full, and they change over time. Still, the broad mechanics are well understood, and they point to practical things you can do.

This guide explains the typical pipeline, the signals that tend to matter, and how to read your own citation data without over-interpreting it.

The typical pipeline

Most AI search experiences follow a similar flow, even though the details differ by product.

  1. Interpret the question. The assistant works out the intent and often rewrites it into several narrower searches. These are sometimes called query fan-outs. A question like “best CRM for a small agency” may become searches about pricing, integrations, reviews and alternatives.
  2. Retrieve candidates. The system pulls pages from a search index, a live web search, or both. If your page is not retrievable, nothing else in this list matters.
  3. Read and rank passages. Models look at passages rather than whole pages, and favor the ones that answer the sub-question directly.
  4. Compose the answer. The model writes a response, drawing on a handful of sources, and attaches citations to some of them.
  5. Decide what to show. Not every source it read is displayed, and different engines are more or less generous with links.

Two things follow. First, you can lose at any step, so a diagnosis has to start with whether you were retrievable at all. Second, fan-outs mean your page competes for many narrow questions, not just the headline one.

Signals that tend to matter

The exact weights are private, but the same themes appear again and again in how these systems behave.

  • Relevance to the sub-question: pages that answer a specific question in plain language get picked over pages that merely mention the topic.
  • Freshness where it counts: for fast-moving topics such as pricing, versions or recent events, newer pages are favored. For stable topics, age matters less.
  • Authority and corroboration: claims that are echoed by several independent sources feel safer to a model than claims that appear in one place.
  • Clear structure: headings, short paragraphs, lists and definitions make passages easy to lift without losing meaning.
  • Specific, checkable facts: numbers, names, dates and steps are quotable. Generic marketing language is not.
  • Accessibility to crawlers: pages blocked by robots.txt, hidden behind scripts or slow to respond are less likely to be used.
  • Source type fit: for some questions engines lean toward reviews or forums, for others toward documentation or news. Matching the source type the engine prefers helps.

Why different engines cite different sources

Each product has its own index, its own retrieval method and its own habits. Some lean heavily on a conventional search index. Others run their own crawling or use live browsing. Some surface many links per answer, others only a few.

The practical result is that the same question can cite quite different sets of domains in different assistants. A page that is cited in one engine may be absent from another. That is why tracking several engines separately matters, and why a single “AI ranking” number can hide the real picture.

What the citation mix tells you

If you collect citations for your tracked questions, you can spot useful patterns.

Which domains win your category

Look at the domains cited most often for your topic. They might be review sites, community forums, vendor blogs, documentation or news outlets. This tells you the kind of source engines trust for your questions, and where you may need a presence beyond your own site.

Where competitors are cited and you are not

A page cited for a competitor but not for you is a concrete gap. Read it. Does it answer the question more directly, offer a comparison table, or carry more evidence? Match or beat it on substance, not on length.

Which of your pages already get cited

Your existing winners show what works for you. Extend them: keep them fresh, add related questions, and link to them from relevant pages.

How to improve your odds

You cannot buy a citation, but you can remove reasons to be skipped.

  1. Be retrievable. Allow the AI crawlers you want, keep your sitemap current and make sure key pages return quickly with real content in the HTML.
  2. Answer first. Open each key page or section with a direct, self-contained answer, then add detail.
  3. Cover the fan-outs. For each core topic, list the narrower questions around it (pricing, alternatives, how-to, pitfalls) and make sure a page addresses each one.
  4. Add evidence. Include real numbers, named examples, dates and links to primary sources.
  5. Mark it up. Use schema such as FAQPage, Article, Organization and Product so the page type and facts are unambiguous.
  6. Earn corroboration. Get accurate descriptions of your brand into the places engines already trust for your category.
  7. Keep it current. Update pages that cover changing facts, and show the update date.

Tip: When you review a competitor page that keeps getting cited, copy its structure of answer, not its wording. The goal is to be easier to quote, not to be similar.

Be careful with certainty

Because outputs vary from run to run, one answer is an anecdote. Judge citations across many prompts and repeated runs, and be cautious about drawing causal conclusions from a single change. If your visibility moves after you update a page, check a few weeks of data and look at whether competitors moved too.

Where aeoora helps

aeoora records which domains and URLs each engine cites for your tracked prompts, shows where competitors are cited and you are not, and flags lost citations as they happen. From there, the Action Center turns the gap into prioritized tasks, and agents can draft the fixes for you to approve.

If you want to see your own citation picture, start a free trial or read how the product works in the docs.

Written by

AI search research

Share this article