Quotes and statistics move AI citations. Keyword density actively hurts.
Last time thread here ended on a question I didn't have a good answer for: if llms.txt does nothing, what actually changes whether an assistant cites you?
There's a controlled study on it — Aggarwal et al. tested nine content interventions and measured how much of the generated answer each one earned. The ranking isn't what most GEO advice suggests.
Adding direct quotations from credible sources: +42.6%. Replacing vague claims with specific numbers and their sources: +32.8%. Plain prose improvements — shorter paragraphs, active voice, sentences that resolve instead of hedging: +28.7%. Citing and linking out to credible sources: +27.7%. Keyword stuffing: -8.7%, meaning it actively makes things worse.
The outbound-link result is the one that flips an old instinct. Classic SEO treated links off your site as leakage. For citation the opposite holds: quoting and linking a credible source makes your own page more quotable, not less.
Two caveats, since the last thread rightly pushed me on evidence. The study measures share of the generated answer, not traffic or revenue — nobody has clean numbers connecting citation share to anything downstream. And it predates the current generation of models, so treat the ordering as directional rather than a scoreboard.
The direction is a relief though: it's mostly "write clearly and show your sources," and nothing on the list is a file you drop at your site root.
https://webpixie.io/blog/post/geo-vs-seo-content-changes
Curious whether anyone here has done the quote-and-cite thing deliberately and then actually seen themselves show up in assistant answers.
Use a fixed five prompt set across ChatGPT, Perplexity and Gemini, then log the exact answer, citation, competitor and date for each run. Repeat it weekly and keep a control page or competitor beside your own so a single lucky answer does not look like a trend. If you have never built that baseline, the free five prompt snapshot lives on the jessie_geo profile.
The outbound-link point is the most interesting bit to me. It makes sense if the model is looking for text it can safely reuse: a claim with a source attached is easier to trust than a claim floating on its own.
I reckon the hard part is measuring this without seeing patterns that are not really there. Assistant answers vary a lot by prompt wording, model, region and timing, so a before/after test probably needs a fixed prompt set and a few competing pages tracked at the same time.
The practical takeaway still feels useful though: replace vague authority with specific claims, numbers, and sources. Even if citations do not turn into traffic immediately, that kind of content is usually better for humans too.
The trust framing makes sense to me too — cite a claim, and now there's an external thing the source can be checked against, which is presumably worth something to a model optimizing for correctness. On the noise problem: that's basically why I didn't want to publish a "before/after our traffic" post here — answers shift enough per run that a single comparison could easily just be noise dressed up as a result. Agree on the last bit too — even if citation share never moves a bit, specific-claims-with-sources is the same thing that makes content better to read, so it's hard to construct a downside case.
The study gives strong citation evidence, but the business link is still missing. Have you seen higher citation frequency translate into qualified traffic or conversions yet?