Browsing-mode citations are the realistic target
ChatGPT (with browsing), Perplexity, Claude (with web search), and Google's AI Mode all use live retrieval: the assistant runs a search query against a SERP backend, fetches the top results, and asks the model to summarise and cite. Perplexity uses its own index; ChatGPT primarily uses Bing; Claude uses Brave; Google AI Mode uses Google. Each retrieval backend has different ranking biases, but the high-level pattern holds across all of them.
Practically: if your page is in the top 5–10 organic results for the query the assistant fires, it has a real chance of being cited. Below that, citation probability collapses. So traditional SEO ranking is the prerequisite — there is no separate AEO ranking system that bypasses it.
What makes a page extractable
Assistants prefer passages that answer a question concisely without requiring much context. Pages that open with a clear definitional or directly-answering paragraph (40–80 words) are cited disproportionately versus pages that bury the answer or front-load brand storytelling.
FAQPage, HowTo, and Article schema all help citation probability in our engagements, both because they signal structure to retrievers and because answer-passage extractors use them to choose snippet boundaries. Use schema honestly — fabricated FAQ entries are penalised by both Google and several LLM retrievers.
Third-party signals matter more than people think
LLMs lean heavily on entity-recognition: if the assistant has seen your brand mentioned consistently across reputable third-party sites, it's more likely to surface and trust your domain. Wikipedia, established industry publications, well-known directories, Reddit threads with substantive discussion, and GitHub/Stack Overflow presence all contribute.
This is why pure on-site optimisation hits a ceiling for AI citations. Ongoing PR, podcast appearances, original research that gets quoted, and presence in the communities your buyers read are leading indicators of citation share for queries in your category.
Training-data citations are slow and hard to influence
When ChatGPT cites you 'from memory' (no browsing), it's pulling from training data that's months to years old. The realistic way to influence this surface is to be mentioned authoritatively across the web before the next training cycle — which is essentially the same playbook as browsing-mode citations, just on a longer feedback loop.
We don't recommend optimising specifically for training-data inclusion. The browsing-mode loop is faster, more measurable, and overlaps the training-data loop almost entirely.