What actually gets a video cited by AI search in 2026 (a 54-study meta-analysis)
Most advice about ranking in AI search is still guesswork dressed up as certainty. On May 7, 2026, SEO researcher Cyrus Shepard published something different: a synthesis of 54 experiments, patents, and case studies published over roughly two years, distilled into 23 scored ranking factors, each rated on how repeatable the finding is across independent studies, how strong the underlying evidence is, and whether a platform has officially documented it. It's the most evidence-grounded single view of this topic available, and it's worth reading against our own citation coverage rather than instead of it.
The top-scored factors
| Factor | Score (of 10) |
|---|---|
| URL accessibility | 9.5 |
| Search rank | 9.4 |
| Fan-out rank | 9.3 |
| Preview controls | 9.2 |
| Query-answer match | 9.2 |
| Intent-format match | 9.0 |
| Answer near the top | 8.8 |
| AI-ready structure | 8.6 |
Source: Cyrus Shepard, Zyppy Signal, May 7, 2026.
The two terms worth actually understanding
Fan-out rank is the one most creators haven't heard of and probably matters most: AI engines don't just answer the query typed in, they run supplementary searches to ground the response, and ranking well across those related questions affects citation odds as much as the original query does. That's a structural argument for the same topical depth our E-E-A-T guide already recommends, covering a subject from multiple angles rather than one narrow video per keyword. Preview controls is the more mechanical one: a "nosnippet" tag or an AI-scraper block reduces how much text an engine can preview, and less preview means fewer citations, full stop.
What the evidence doesn't yet support
llms.txt, the plaintext file this site maintains specifically for AI crawlers, scored lowest of the 23 factors, at 2.0. That's not an argument against having one, it costs almost nothing to maintain, but it is an honest signal that the file's actual effect on citation behavior remains unproven at the evidence level this analysis demands. Worth knowing before treating it as a bigger lever than the data currently supports.
A second study, pointing the same way
Ahrefs separately analyzed 76 million AI Overview citations and found brand mentions correlate with citation probability at roughly 0.664, against 0.218 for backlinks, a 3x gap on the specific job AI retrieval does. Read alongside Shepard's framework, the pattern is consistent: being talked about and being structurally accessible now matter more than the link-building effort classic SEO spent two decades optimizing.
Where Thothium fits
Thothium exports the metadata, chapters, and description structure that the highest-scored factors here actually measure, so a video is built against evidence rather than a guess about what AI search wants. It is in free alpha, and the form below gets you a key.
Frequently asked questions
Who actually did this research, and can it be trusted?
Cyrus Shepard, founder of the SEO firm Zyppy, published the analysis on his Zyppy Signal newsletter on May 7, 2026. He didn't run new experiments himself; he synthesized 54 existing experiments, patents, and case studies and scored each factor on repeatability, strength of evidence, and official platform support, explicitly labeling the result correlation, not proven causation.
What do "fan-out rank" and "preview controls" actually mean?
Fan-out rank refers to how well your content ranks across the supplementary, related searches an AI engine runs to ground its answer, not just the original query, so ranking well on the surrounding questions matters as much as the main one. Preview controls means technical settings like "nosnippet" tags or AI-scraper blocks; the research found that restricting how much text a crawler can preview measurably reduces citation odds.
Why did llms.txt score so low if it's specifically built for AI crawlers?
Because scoring here measures repeatable, evidenced impact, not intent. llms.txt is a young, voluntarily-adopted convention with little independent verification that following it changes citation behavior, versus a factor like URL accessibility, which has direct technical grounding in whether a crawler can retrieve a page at all. A low score means "unproven," not "worthless," and it's cheap enough to maintain regardless.
How does the Ahrefs brand-mention finding fit with the rest of this?
It's a separate, complementary data point: Ahrefs analyzed 76 million AI Overview citations and found brand mentions correlate with citation probability at roughly 3x the strength of backlinks. That reinforces one specific idea already in Shepard's framework, that AI systems appear to weigh being talked about differently than being linked to, without requiring the same technical link-building effort SEO has run on for two decades.
Last updated September 5, 2026. This is a synthesis of correlational research, not a guaranteed formula; AI search systems and their ranking behavior continue to change, and the factors and scores here reflect evidence available as of the study's publication.