AI When Perplexity or ChatGPT answers with citations, it’s tempting to read the source list as a quality ranking. It isn’t. Citation is a byproduct of a pipeline, and each stage of that pipeline has its own biases. If you want to be cited, it helps to know what each stage rewards.
The pipeline in four stages
1. Query rewriting. The assistant rarely searches your user’s literal question. It generates one or more search queries from it — usually shorter, more generic, keyword-ish phrasings. Practical consequence: covering the canonical phrasing of a topic matters more than covering every long-tail variant, because the long tail gets rewritten into the canonical form anyway.
2. Search. Those queries go to a conventional search index (Bing powers several major assistants; Google powers Gemini). This is where classic SEO still earns its keep: if you don’t appear in the top handful of results for the rewritten query, the model never sees your page. AEO doesn’t replace SEO — it sits on top of it.
3. Reading. The assistant fetches a few results and extracts passages. It works with a budget: partial page loads, token limits, short timeouts. Pages that front-load the answer survive this stage; pages that spend eight hundred words warming up don’t. This is also where rendering matters — content that only exists after client-side JavaScript runs is invisible to most of these fetchers.
4. Synthesis. The model writes its answer and attributes claims to the passages that supplied them. The passages that get cited are the ones that made a claim cleanly — a specific, self-contained statement the model could paraphrase without stitching together three paragraphs.
What this means for how you write
The pattern that falls out of stages 3 and 4 is simple enough to be a checklist:
- Put a direct answer in the first or second paragraph under a question-shaped heading.
- Make one claim per passage, with the supporting number or fact in the same sentence.
- Prefer specifics (“reduced crawl errors 40% in six weeks”) over abstractions (“significantly improved performance”) — specifics are quotable, and quotable is citable.
- Keep critical content server-rendered and fetchable.
The part you can’t see from inside
Every stage of this pipeline is invisible from your analytics. You can’t see the rewritten queries, which results were fetched, or which passages survived extraction — you only see the occasional referral click. The practical workaround is to test from the answer side: ask the assistants your buyers’ questions and record who gets named and cited. Tools like ChatFeatured automate that across assistants and over time; even a manual weekly spreadsheet beats guessing. Either way, treat the answer — not the ranking — as the thing you’re optimizing.