What Actually Drives AI Citations, and What's Just Sold to You

In the space of five weeks, three independent studies landed on the same conclusion. The two tactics sold hardest across the AEO and GEO category, schema markup and llms.txt files, aren't the citation levers they're marketed as. One of those studies analysed 137,000 domains and found that 97% of llms.txt files are never read by anything at all. Here's what the evidence says actually drives AI citations, and what's just being sold to you.

If you have bought, or been pitched, an AI visibility product in 2026, the pitch almost certainly led with one of two things: add structured data so AI can understand your pages, or publish an llms.txt file so AI can find your content. Both sound technical, both are easy to productise, and both can be generated in minutes. Neither, on the controlled evidence now available, measurably moves AI citations on pages that are already visible to AI systems.

This is not an argument that these tactics are worthless. Schema has genuine upstream value. llms.txt has a real, narrow use case. The argument is narrower and more useful than that: the specific claim that adding schema or publishing llms.txt will increase your AI citations does not survive contact with controlled data. Three separate research efforts, using different methods and different datasets, now point the same way.

The three studies, in one place

Three studies, three methods, one direction. A controlled experiment on schema. A 137,000-domain traffic analysis on llms.txt. And the largest platform's own published guidance covering both. When independent research using different methodologies converges like this, the conclusion is no longer a matter of opinion or vendor positioning.

The llms.txt finding deserves a closer look

The schema study and Google's guidance have been discussed at length elsewhere, including in our own deep-dives linked below. The newest study, the 137,000-domain llms.txt analysis published on 15 June 2026, contains the single most striking data point of the three, and it has not yet been widely absorbed.

The 28% adoption figure tells you how effectively the tactic has been sold. More than one in four domains in the study have published an llms.txt file, despite the fact that no major AI platform has ever committed to reading it. Adoption has been driven by speculation that AI platforms might start consuming the file, not by any confirmation that they do.

The 97% figure tells you what that adoption produced. Of roughly 38,000 domains with a valid file, almost none received a single request for it in the measured month. And the small fraction that did get read were not read by the bots site owners are hoping for. When Ahrefs classified every user agent fetching llms.txt into twelve categories, the bot most associated with AI search visibility came in last.

Highlighted rows are AI bot categories. The bottom row, AI retrieval bots, is the category most directly tied to AI search citation. Combined, all four AI categories make up 19.5% of requests, but that is 19.5% of the 3% of files read at all, meaning roughly 0.6% of published llms.txt files ever receive any AI bot request whatsoever. "Fetched" also does not mean "read", so every figure is a ceiling on actual consumption.

Two details make the picture unambiguous. First, Slackbot, a chat app's link-preview bot, fetched llms.txt files more often than PerplexityBot did. Perplexity is one of the AI search engines the file was supposedly designed to help. Second, Ahrefs found zero requests from AI bots for llms.txt files that don't exist. AI bots never go looking for the file opportunistically; they only ever fetch one that happens to be in a path they were already visiting.

The one genuine use case the data supports is the agentic one. AI agents and agentic infrastructure, the category that includes coding agents like Claude-Code, were the largest single AI consumer at 10.5%. This lines up with what Google's John Mueller said when pressed on the contradiction between Search Central's guidance and Chrome's Lighthouse audit: llms.txt is, in his words, a temporary crutch for AI coding tools parsing developer documentation, not something most sites need for search.

Why these tactics get sold anyway

If the controlled evidence is this clear, why does the category keep leading with schema and llms.txt? The answer is not conspiracy. It is convenience. Both tactics share three properties that make them easy to sell and hard to resist.

They're easy to measure. A tool can check in milliseconds whether your site has schema or an llms.txt file, assign a pass or fail, and put it on a dashboard. A binary checkbox feels like progress. They're easy to generate. Both can be produced automatically, which makes them an ideal productised deliverable: the tool finds the gap and the tool fills it, in one motion. And they correlate with citation in observational data, because sites that bother implementing them tend to also invest in content, links, and maintenance. Strip the schema out and the rest of the quality stack still carries the page. The correlation is real; the causation isn't.

None of that makes the tools acting in bad faith. It makes them optimised for what is packageable rather than for what the controlled evidence says moves citations. The distinction matters because a buyer who spends their AI visibility budget on the packageable layer is spending it on eligibility signals while their actual citation problem, which lives in the content, goes untouched.

What the evidence says does drive citations

If schema and llms.txt are not the levers, what is? The same body of research that disconfirms the markup-first tactics points clearly at the alternative. The Princeton and KDD GEO benchmarks, the peer-reviewed foundation under most credible AI citation research, measured which interventions actually moved citation rates in production AI systems. The top three were citing sources (+40.6%), adding quotations (+35.1%), and adding statistics (+32.9%).

Notice what those three have in common. None of them is a technical file or a markup format. All three are properties of the content itself: the density of verifiable, attributed, specific information inside the passage. That is the signal AI retrieval systems are actually responding to, because retrieval-augmented generation, the mechanism Google confirmed in its own May 2026 guidance, works by retrieving and synthesising individual passages, not by reading markup or index files.

This is the basis for the Citation Probability Score® framework. CPS® scores content at the block level, the 134 to 167 word chunk that production retrieval systems actually embed and evaluate, across five pillars that map directly onto the research: Content Structure, Fact Density, Answer Architecture, Self-Containment, and Freshness. It does not score schema presence or llms.txt presence as citation drivers, because the controlled evidence says they aren't. It scores the content qualities that the GEO benchmarks, the Ahrefs studies, and Google's own mechanism documentation all point to.

Go deeper on each study

This page is the synthesis. Each of the three studies has its own dedicated analysis where the methodology, the caveats, and the per-platform implications are worked through in full.

What to do with this

The practical takeaway is a reallocation, not a teardown. If your AI visibility programme currently leads with schema scoring and llms.txt generation as the headline deliverables, those belong in the technical readiness layer, where they do useful work, not at the top of your citation strategy. Keep them. Stop expecting them to move citations.

Then move your attention, and your budget, to where the controlled evidence points: the content itself, scored at the block level. Read the first paragraph of your most important pages. Does it open with a declarative answer or with brand narrative? Is it dense with verifiable facts, or thin and descriptive? Does it stand on its own, or does it depend on the heading above it that the retrieval system never sees? Those are the questions that determine whether you get cited. No file at your domain root changes the answer to any of them.