Perplexity cites somewhere between three and six sources per answer, depending on the query. ChatGPT's browsing mode rarely surfaces more than five. The researchers who coined the term "Generative Engine Optimization" — a 2023 paper out of Princeton, Georgia Tech, and Allen AI — found that pages with citable statistics and direct quotations saw visibility gains of up to 40% in AI-generated answers, while stuffing in more keywords, the thing that used to move a Google ranking, did almost nothing. Those three numbers get quoted a lot in GEO writeups. The fourth one doesn't, and it's the one that should actually change how you write a page: when you trace which part of a document a model pulled its answer from, it's overwhelmingly the top third. Not the best-written paragraph. Not the section with the most detail. The top.
Why "comprehensive" stopped being the flex
For fifteen years, the SEO advice was some version of "write the definitive resource." Cover every angle, hit 2,500 words, out-depth the competition, and Google's ranking systems would eventually notice you'd built the most complete answer on the internet. That instinct doesn't transfer to how a generative engine picks sources, and understanding why requires understanding what's actually happening under the hood when someone asks ChatGPT or Perplexity a question instead of typing it into a search box.
A search engine ranks whole documents. A generative engine doesn't read whole documents at inference time — it retrieves chunks. Your page gets split into overlapping windows, typically somewhere in the 300–500 token range, each one turned into an embedding and scored against the user's query for similarity. The system pulls the top few chunks across all the candidate pages, stuffs them into the model's context along with the question, and asks it to synthesize an answer. Your 2,500-word guide isn't competing as a single unit anymore. It's competing chunk by chunk, and most retrieval pipelines only pull one or two chunks per source before moving on to the next citation.
Here's the part that surprised me when I first started paying attention to which pages actually got cited for queries I was tracking: the opening chunk wins disproportionately often, even when a later section answers the query more precisely. Some of that is genuine relevance — good writers put the answer near the top anyway. But a lot of it is structural. Titles, H1s, and meta descriptions get indexed with extra weight in a lot of retrieval systems, and that weight bleeds into whatever paragraph sits immediately below them. A page that buries its actual claim under six paragraphs of throat-clearing is handing its best material to a chunk boundary it doesn't control.
What this looks like in a real comparison
| Factor | Classic SEO weight | GEO weight |
|---|---|---|
| Total word count | Correlated with rankings for years | Roughly irrelevant past the first chunk |
| Keyword density | Moderate signal, historically | Near zero — embeddings match meaning, not string overlap |
| Position of the answer on the page | Minor UX factor | Dominant — determines which chunk gets retrieved |
| Citable stats / direct quotes | Nice-to-have for backlinks | Up to +40% visibility per the Princeton/GT study |
| Backlink count | Primary ranking signal | Indirect at best — still matters for crawl priority, not retrieval scoring |
The part people get backwards
The common mistake I see when builders try to "do GEO" is treating it like a keyword problem with a new vocabulary — they go hunting for the phrases people type into ChatGPT and try to work those into a headline. That's not nothing, but it's optimizing for the query side of a match when the leverage is almost entirely on the document side. The move that actually shows up in citation counts is answering the question in the first 150–200 words, in plain declarative sentences, before you get to the nuance, the caveats, or the origin story of why you're qualified to answer it.
I rewrote one of our own help-center pages purely to move the answer above the fold — same content, same word count, just reordered — and it started showing up in Perplexity answers for a query it had never appeared for in three months. Nothing else changed. That's a small, single data point, not a study, but it's the kind of thing that makes you stop writing intros.
The second-order effect worth knowing about: because the retrieval unit is a chunk, not a page, a long page doesn't get penalized for depth — it just doesn't get credit for it in the citation itself, even though the depth may be why a human clicks through afterward. That means the old habit of front-loading SEO fluff ("In this comprehensive guide, we'll explore...") is actively costing you the one slot that matters, while the actual substance further down is doing unpaid work for click-through and dwell time but never making it into the model's context window at all.
What to actually change
Three things, in order of how much they moved the needle for pages I've watched: put the direct answer to the implied question in the first paragraph, not the third. Use a real number or named source in that same paragraph if you have one — the 40% figure isn't a fluke, models seem to weight concrete claims over general statements when picking what to quote. And structure headers as if each one might be the only sentence retrieved, because for a meaningful share of AI-answer traffic, it will be.
None of this replaces writing something worth reading start to finish. It just means the opening 200 words are no longer throat-clearing real estate — they're the whole pitch, made to an audience of one model deciding whether you're worth citing at all.
