There are two distinct questions hiding inside “why doesn’t AI mention us?”, and confusing them is the reason most AEO work stalls. The first is whether an engine knows your brand at all. The second is whether, on a specific question, it chooses to cite your page rather than someone else’s.
The two have different causes and different fixes. This post is about the second.
Citation is a retrieval problem, not a memory problem
When an assistant answers from what it absorbed in training, there are no citations to win: it produces a claim with no source attached. Citations appear when the engine goes and fetches something. Perplexity does this on nearly every query; ChatGPT and Gemini do it when the question looks current, specific, or commercial; Claude does it when browsing is available to it.
That retrieval step is the whole game. It runs roughly like this:
- The question is rewritten into one or more search queries.
- A conventional search returns a candidate set, typically small.
- Each candidate is fetched and reduced to the passages that look relevant.
- The answer is composed from those passages, and the pages whose text actually contributed get cited.
Step two is where classical SEO still matters: you cannot be cited if you were never retrieved. But steps three and four are where most brands lose, and no amount of ranking fixes them. Your page was fetched. Nothing in it was usable.
What makes a passage usable
The model is looking for text that answers the question on its own, without the surrounding page. That single constraint explains most of what follows.
The answer appears before the argument
Journalism calls this the inverted pyramid, and it is the single highest-leverage change most pages can make. A section that opens with “Answer Engine Optimization is the practice of making a brand appear inside AI-generated answers” has given the engine something to lift in one sentence. A section that spends three paragraphs establishing why the topic matters before defining it has not.
This runs against how a lot of marketing content is written, where the payoff is held back to keep people reading. Engines do not read to the end.
The claim is self-contained
Pronouns and back-references break extraction. “This makes it considerably more effective” is unusable in isolation, because this and it live in a previous paragraph the engine may not carry along. Name the subject again, even when it feels repetitive to a human reader. Repetition is a style cost; ambiguity is a citation failure.
Headings read as questions
Retrieval matches the user’s question against your page’s structure. A heading that reads How much does an AEO audit cost? matches a question of the same shape far more strongly than one reading Pricing. This is also why FAQ schema works: it hands the engine an explicit list of question-and-answer pairs, pre-extracted.
Figures carry their own context
“Visibility rose 12 points” is not quotable. “Across 40 tracked prompts on five engines, visibility rose from 6% to 18% over sixty days” is, because every qualifier a careful engine needs is inside the sentence. Numbers without their denominator get dropped, and rightly.
The part nobody controls directly
Retrieval leans on sources that other sources already trust. In practice this means a small number of publishers dominate the citation pool for any given category: review platforms, established trade publications, documentation, and forums where the question gets discussed in the open.
This is uncomfortable, because it means the fastest route to being cited in your category is often not to publish more on your own domain. It is to be accurately described on the domains the engines already read. A single corrected listing on a review site that engines consult can move more answers than a quarter of blog posts.
The practical step is to find out which sources those are for your category rather than assuming. They are rarely the ones you would guess, and they differ by engine.
Why engines cite different sources for the same question
Each engine rewrites the question differently, searches a different index, and applies different rules about how many sources to consult. So the citation pool for “best CRM for small teams” on Perplexity may overlap only partly with the pool on Gemini.
The practical consequence: a single blended “AI visibility” score averages away the thing you need to act on. Being invisible on one engine and strong on four is a completely different problem from being weak on all five, and the average looks the same.
How to tell which problem you have
Pick one question a real buyer would ask. Ask it on each engine from a fresh session, with no personalisation. Then read the answer twice, for two different things.
- Are you named in the text? If not, this is a presence problem. The engine does not associate you with the category.
- Are you in the citation list? If you are named but not cited, something else is the source of what the engine knows about you. If you are cited but not named, your page was read and contributed nothing memorable.
Do that across a few dozen questions and the pattern stops being anecdotal. That is the point at which it becomes a work list rather than a worry.
What to change first
In rough order of return on effort:
- Rewrite the opening of each key page so the answer precedes the argument. Cheapest change, largest effect.
- Turn section headings into the questions they answer, and add FAQ schema that mirrors them.
- Make every claim self-contained. No orphan pronouns, no figures without their denominator.
- Find the sources the engines actually cite for your category, and get accurate about your brand on them.
None of this is exotic, and that is rather the point. Citation rewards clear writing that states its conclusions. The novelty is only that the reader is a machine that stops reading early and never scrolls back.