How do you get cited by LLMs? The evidence, lever by lever

What is the single strongest signal for getting cited by LLMs?

This page is the evidence, not a to-do list. This page walks the real research on AI citations lever by lever, with the sample sizes, the effect sizes, the sources, and the caveats each study carries. If you just want the plain steps, read the practical version, our action guide on how to get cited by AI. If you are the skeptical or technical reader who wants to see the studies before you believe the advice, you are in the right place.

One honest warning up front. The research here is thin and uneven. The strongest on-page study was run on a simulated engine, not live ChatGPT. The strongest off-page study is from a tool vendor and shows correlation only. So we lead with what is independent, flag every vendor, give ranges instead of single numbers, and name what has been debunked. That honesty is the whole point of this page.

How do you get cited by LLMs, according to the research?

Two levers have the best evidence behind them. On the page, content packed with specific stats, citations, and quotations. Off the page, earned brand mentions on trusted third-party sites. Everything else is either weaker, unproven, or debunked, and the table below lays out each one with its source and its caveat.

Start with the one finding that reframes the whole question. Ahrefs looked at 15,000 queries and found only about 12 percent of the URLs cited by AI also rank in Google's top 10. Ahrefs sells SEO tools, so read it with that in mind, but the result cuts against their own product, which raises its credibility. The takeaway is that most AI citations go to pages that do not rank at all. Getting cited is not the same game as ranking, so the levers are not the same either.

A quick note on the r values, because they carry a lot of this page. Spearman's r is a correlation score from 0 to 1, where 1 means two things move together perfectly and 0 means no link at all. It measures whether two things rise together, not whether one causes the other. Ahrefs, the source of these scores, says so plainly. So treat r=0.664 for brand mentions as the best hint we have, not final proof.

Do backlinks still matter for AI citations?

Yes, but far less than earned brand mentions. In the Ahrefs 75,000-brand study, branded web mentions correlated with AI-Overview visibility at r=0.664, while backlinks came in at r=0.218. That is about 3 times weaker. The single most useful thing you can read into that gap is this. A plain mention of your brand name, even with no link attached, tracked AI visibility more closely than a link did.

Why this matters. For twenty years the link was the currency of search. This is the first solid dataset suggesting that for AI answers, the link is not the main event. The mention is. That changes where you spend. A press quote, a directory listing, a mention in an industry roundup, all of these can carry weight even when they do not link back to you.

The honest caveat. This is one vendor study, and it measures correlation, not cause. It is possible that brands with lots of mentions also do many other things well, and the mentions are a symptom, not the driver. Ahrefs says as much in the study itself. So do not read it as "links are dead." Read it as "mentions look more important than links, on the best public evidence we have, and that is worth acting on."

Does domain authority matter?

The evidence says domain authority is a weak signal, and possibly a confounded one. In the Ahrefs study, domain rating correlated at only r=0.326, less than half the strength of brand mentions. Seer Interactive found lower-authority pages out-cited higher-authority ones in their sample. And the 12 percent overlap number tells the same story from another angle. Most cited pages are not the ones ranking at the top of Google.

Why we say weak, not zero. High-authority brands often have more mentions, more content, and more history, so some of their citation edge may come from those things, not from the authority score itself. That is what "confounded" means here. The authority may be riding along with the real drivers rather than causing anything. The Ahrefs overlap data points the same way. AI often cites pages sitting well below Google's top 10, not just the winners at the top.

What to do with that. Do not pour your budget into chasing a big domain rating for its own sake. It is a slow, expensive number that the evidence links only weakly to getting cited. The mentions and the quotable content are where the stronger signals sit.

What on-page changes have evidence?

One real study anchors this, and it is the Princeton GEO paper (KDD 2024). It found that adding statistics, citations, and quotations to a page lifted its generative visibility by roughly 30 to 40 percent, with quotations and stats scoring highest and fluency adding another 15 to 30 percent when combined.

The methodology, so you can weigh it. The researchers built a benchmark of about 10,000 queries across 9 domains, then rewrote source content with different tactics and measured how visible each version became in the generated answer. The results are consistent and the mechanism is believable. Quotable, fact-dense text is easier for a model to lift and attribute.

The caveat that governs the whole finding. This was measured on a simulated GPT-3.5 style engine the team built, not on live ChatGPT, Perplexity, or Google AI Overviews. So it is directional mechanism evidence, not a live effect size. Treat the 30 to 40 percent as "this kind of change helps in a controlled setting," not "you will get 30 to 40 percent more citations tomorrow." The same paper found keyword stuffing scored below the plain baseline, and an authoritative tone on its own showed no significant improvement. So the win is real content quality, not tricks.

What has been debunked?

Three things get sold as citation levers that the evidence does not support. Each one has a specific study or statement behind the debunking, so here they are.

Schema markup on its own. Ahrefs ran a controlled test on 1,885 pages using a matched difference-in-differences design, and adding schema moved AI citations within noise, roughly minus 5 to plus 2 percent across AI Overviews, AI Mode, and ChatGPT. Google says no special schema is needed for AI Overviews or AI Mode. Schema is still worth having for rich results and entity clarity, so it is not useless, it just is not a citation lever by itself. Our full take is at schema markup for AI search.

The llms.txt file. Google's John Mueller called it "purely speculative," and Gary Illyes said in mid-2025 that Google does not support it. No AI system is on record reading it. Adding one does no harm, but the evidence gives you no reason to expect a citation from it. Our full take is at does llms.txt work.

Keyword stuffing. In the Princeton study, stuffing keywords scored below the plain baseline. It reads badly to humans and it does not help with AI either. The old SEO habit is a net negative here.

What is the honest bottom line?

Put your effort where the evidence is strongest, and range everything else. The two best-supported levers are quotable content full of stats, citations, and quotes on the page, and earned brand mentions off the page. Backlinks matter less than mentions. Domain authority is weak and maybe confounded. Schema alone and llms.txt do nothing on their own.

What makes this trustworthy is that the independent and vendor sources mostly agree in direction, even when they disagree on the exact number. That is why we give you the ranges and the flags instead of a clean headline figure. If you want the plain, do-this-first version of all this, read our action guide on how to get cited by AI. And to track whether any of it is working on your own site, see our guide to how to measure AI visibility.

How to apply the evidence

Here is the research turned into four moves, in order of how strong the evidence is. This numbered list matches the plan, so you can act on it without re-reading the studies.

1. Make your content quotable with real stats and citations. Fill your pages with specific numbers, named sources, and short quotes from credible references. This is the on-page move with the best evidence, from the Princeton GEO paper, even though that study used a simulated engine.
2. Earn brand mentions on trusted third-party sites. Get your business named on the sites AI reads, like directories, review sites, and industry pages. Ahrefs found brand mentions correlated about 3 times stronger with AI visibility than backlinks did. This is slow work, and it is the highest-value work.
3. Keep your site crawlable by AI bots. Access is the floor under everything above. If AI crawlers cannot fetch your pages, no lever helps. Confirm your robots.txt and CDN allow them, since a silent block is the most common cause of zero visibility.
4. Ignore schema-only and llms.txt as citation levers. Do not spend citation budget on schema markup alone or an llms.txt file. Controlled tests and Google's own statements show neither moves AI citations by itself. Keep schema for rich results, not for citations.

Common questions about how you get cited by LLMs