§ I The retrieval problem
Most people assume that if they rank on Google, they'll appear in AI answers. They're wrong. LLMs don't work like search engines. They don't look up your website when someone asks a question: they draw on what they already know, or what a retrieval system has fetched and handed them.
For ChatGPT and Claude in their base modes, "what they already know" is their training data: a snapshot of the web from some point in the past, filtered and weighted in ways no one fully understands publicly. For Perplexity, there's a live web search layer on top of that. For Claude with tool use, there may be retrieval. The path to citation is different across every model, and most brands haven't been built for any of them.
The practical consequence: a brand can rank in the top three on Google for its core keyword and still be completely invisible in AI answers. The two surfaces use different signals, reward different structures, and pull from different source pools. Ranking on one does not carry over to the other.
§ II What models actually look for
The model wants a sentence it can lift verbatim and attribute clearly. If it can't find one, it skips your page.
LLMs are pattern-matching machines. When asked a question, they look for content that matches the shape of an answer. A page that hedges everything, buries the answer in the fifth paragraph, or uses passive voice throughout is much harder to cite than a page that opens with a direct, definitive claim and supports it with evidence.
This is the core insight behind AEO: you're not writing for a human reader who will scroll and evaluate. You're writing for a model that wants one thing: a clean, attributable answer to the specific question it was asked. If your page gives it that, it gets cited. If it doesn't, the next page does.
The structural corollaries are real: short introductions, clear topic sentences, Q&A formatting where relevant, comparison tables where appropriate, and a definitive thesis in the first paragraph. None of this is complicated. Most brand content is structured the opposite way, because it was written for a different era.
§ III The prompt map
Before writing a single word of AEO-optimised content, you need to know which prompts your buyers are actually asking. Not keywords: full questions, the way a person would type them into ChatGPT at 11pm when they're trying to solve a problem.
"What's the best ATS for a 50-person startup?" is a different page from "how do I choose an applicant tracking system." Both deserve coverage. They have different answer shapes, different comparison triggers, different citation opportunities. If you write one page meant to serve both, you serve neither particularly well.
A prompt map is a structured inventory of these questions, organised by buyer stage (awareness, consideration, decision), by model behaviour (which models answer this type of query, which ones cite sources), and by competitive landscape (who is currently getting cited for this prompt). It's the brief that comes before any content. We don't write an AEO page without one.
§ IV Schema and structure
Schema markup doesn't guarantee citation, but the absence of it removes you from contention in certain retrieval paths. Perplexity, in particular, pays attention to structured data when deciding what sources to surface. At minimum, you want: Organization markup with accurate name, URL, and description. Article or WebPage markup on any content you want retrieved. FAQPage where you have genuine Q&A content.
The FAQ schema is consistently underrated. It puts question-answer pairs in a format models are trained to recognise: literally the shape of conversational exchange. A well-structured FAQ on a product page can become the cited source for a dozen different prompts, none of which contain your product name. That's the compounding effect of structural citation work.
llms.txt is also worth implementing. It's a simple text file in your root directory that tells models what your site is and which pages matter. Perplexity has confirmed they crawl it. Claude has documented support for it. It takes one developer twenty minutes to ship. There's no reason not to have it.
§ V The honesty constraint
Models tend to avoid citing sources that make claims without evidence. A page that says "we're the industry leader" or "our solution is the best" gets filtered out. The claim is unverifiable and the model doesn't want to be cited as the source of an assertion it can't support.
A page that says "we reduced churn by 18% for mid-market SaaS companies by implementing X, Y, and Z" is citable. The specificity is the citation bait. The mechanism is the trust signal. The data is the proof the model needs to feel comfortable lifting the sentence and attributing it to you.
This is a harder problem for most marketing teams than it sounds. The instinct is to claim, not to prove. AEO-optimised content runs the opposite direction: prove first, claim second, or don't claim at all. Specificity over superlatives, every time.