The Answer-First Rule for AI Citations Is Advice, Not Research

Fifty-three percent of pages cited by Google AI Overviews are under 1,000 words.

That number stopped me. It comes from an analysis of 560,000 AI Overviews and 1.6 million cited URLs. The average cited page is 1,282 words. The median is below 1,000.

I should say up front: I'm not a GEO expert, and I'd argue nobody really is yet. The field is barely a year old and everyone, me included, is running small experiments and reading each other's. What follows is a compilation of what I've tested on my own writing and what I've read from people running actual analyses. Take it as a starting point for your own checking, not a verdict.

I'd been collecting GEO advice for my own writing and because I'm building tools for teams that publish content daily. The consistent instruction across every guide: answer the question in the first two paragraphs. Put the conclusion first. AI engines extract answers from the top and cite the pages that make extraction easy.

So I went looking for the study behind that claim. There isn't one.

The inverted pyramid rule has not been tested

I couldn't find any primary research that isolated answer placement as an independent variable and measured its effect on AI citation rates. If it exists, I'd genuinely like to see it.

The closest structural research — published in September 2025, specifically on what drives AI citations for product content — doesn't mention inverted pyramid, answer-first, or first paragraph at all. Its structural guidance is about machine-readability: schema markup, comparison tables, scannable specifications. Not where on the page the answer sits.

The advice isn't wrong. Answering early is good writing. It respects the reader. It probably gives AI engines a cleaner passage to extract from than burying the punchline in paragraph eight. But "probably better" and "confirmed to increase citation rates" are different claims. The second one has no study behind it.

The position that matters is inside the AI's output

The foundational GEO research, from 2023, introduced a citation metric using exponential decay. Sources cited earlier in an AI response receive more visibility weight than sources cited later. The paper states it directly:

Sentences that appear first in the response are more likely to be read.

Almost every practitioner summary applies this to article structure. Put your answer early because early position gets more weight.

That's a misread. The decay applies to the AI's output ordering — where within the model's own response your source ends up. Whether your article gets cited first or fourth in a Perplexity answer is a property of the engine's ranking, not of where within your article you placed the answer. The ordering is happening inside the model, not inside your document.

You can't reorder yourself into first position by restructuring your page.

The one tactic confirmed to hurt

The same research that introduced the citation metric also tested what happens when you apply classic SEO tactics in a generative engine context. Keyword stuffing — adding more instances of the target term — scored measurably below no optimization at all. About a 9% decline on their citation metric.

This matters because a lot of GEO advice is recycled from SEO playbooks with different framing. Some of it actively reduces citation chances. The two disciplines aren't equivalent, and the most expensive mistake in the space right now is treating them as if they are.

Generic optimization can hurt specific articles

Research published in March 2026 looked at why standard GEO methods underperform. The finding:

Existing approaches apply generic rewriting rules uniformly, failing to diagnose why individual documents are not cited.

For long-tail content, generic optimization can actively harm citation rates.

The alternative — diagnosing why each specific article isn't being cited and making targeted changes to about 5% of the content — outperformed blanket rewrites significantly. The same structural change that helps a broad-topic page can hurt a page on a specific, low-volume question.

The tactic needs to match the document. There is no universal formula.

What the measurement gap means for you

There is a distinction almost nobody in the practitioner space acknowledges: contribution versus citation.

Most GEO research measures contribution — how much of your article's language appears in the AI's output. That is not the same as citation — whether the AI attributes the source, links to your page, and sends readers your way.

You can be heavily contributed (your phrasing paraphrased without credit throughout a ChatGPT answer) and receive zero traffic from it. Research published in 2026 identifies this as a systematic flaw in how GEO performance has been measured. The tactics that maximize content presence in AI output don't necessarily produce the attributed citations that drive visits.

If you're writing for traffic, not just presence, that gap matters more than almost anything else in the GEO conversation.

The short version, in one table

If you skim nothing else, here is where each common tactic stands against the research I could find:

What I'm doing with this

Answer-first structure is still the right default. Nothing here disproves it — the research just hasn't confirmed it as a citation mechanism. Meanwhile, shorter and more focused is confirmed. An 800-word page that directly answers one specific question outperforms a 3,500-word overview that covers the topic and everything adjacent.

I'm writing shorter articles because of this. Not shorter in the sense of stopping before the idea is complete — shorter in the sense of not adding extra sections because a longer piece feels more authoritative. The data suggests that instinct is wrong. The median AI-cited page would not be described as "comprehensive" by any SEO playbook.

The articles that get cited by AI engines look less like SEO content and more like precise technical documentation: one specific question, a direct answer, no padding, machine-readable structure. That's not a radical approach. It mostly means dropping the habits that made articles longer without making them better.

One thing I'm watching closely: whether "answer first" eventually gets tested properly. Right now it's a reasonable prior with no primary evidence. If a controlled study isolates it as a variable and shows an effect, the guidance changes. Until then, the lever with the most data behind it is scope: a focused page on a narrow question, kept short, with clean markup.

And treat everything here the way I'm asking you to treat the answer-first rule.

This is a compilation of my own tests and what I've read from people running real analyses, not a ruling from someone who has this figured out. GEO is moving fast and a lot of it is still guesswork dressed up as method. Go read the studies I'm pointing at and run the experiments on your own pages. Believe what you can measure over what reads confidently online, including this article.