Someone scored the guesswork. Cyrus Shepard gathered 54 experiments, patents, and case studies covering ChatGPT, Gemini, and Perplexity, then ranked 23 factors associated with earning an AI citation by how repeatable the evidence was and whether the platforms themselves confirmed it. The resulting order is not the order the industry talks about.
The top five are URL accessibility at 9.5, search rank at 9.4, fan-out rank at 9.3, preview control at 9.2, and query-answer match at 9.2. Near the bottom sit structured data at 5.6, domain authority at 5.0, and an llms.txt file at 2.0, which is last of the 23.
Do meta titles and descriptions matter for AI citation?
Two of them do, and they are likely not the two you have been editing. Preview control scored 9.2 and sits fourth, which covers the nosnippet and data-nosnippet directives that tell search engines how much of a page they may display, and limiting how much text an engine can show reduces how likely it is to cite you. That makes it the rare meta tag that can remove you from consideration without any warning, and nobody watches it because it lives in a template, out of sight of the editor where people look.
Query-answer match also scored 9.2, and it works on the semantic closeness between your page titles, your subheads, and both the question asked and the answer the engine assembles. So the title earns its place on this list, where its job is classification.
What does not appear anywhere in the top factors is the meta description as a piece of salesmanship. It still does real work once an engine picks you, because that is the text frequently quoted back when you get cited, and writing meta tags for SEO and AI search covers that side of it. Being described well and being chosen are two different jobs, and only one of them is decided by your description.
Why is URL accessibility the highest-scoring factor?
Because everything else is moot if the crawler cannot reach the page. Accessibility scored 9.5, ahead of every content and authority signal on the list, which puts the least glamorous work at the top of the priority order. Blocked paths, login walls, JavaScript that renders the substance client-side, and robots rules written for a different era all remove a page from the pool before any judgment about quality happens. Robots rules are the cheapest of those to fix, and the robots.txt and llms.txt Punch List names the AI search crawlers to allow, GPTBot, OAI-SearchBot, ClaudeBot, and PerplexityBot among them, with a template that keeps the training-only crawlers out without blocking the ones that cite you.
That makes AI visibility a technical problem first, and accessibility as an AI search strategy goes further into why the plumbing decides the outcome. Pair it with monitoring index coverage, since a page missing from the index is missing from the answer.
Does classic SEO still matter for AI answers?
More than the discourse suggests. Search rank came second at 9.4 and fan-out rank third at 9.3, and 38% of AI Overview citations come from the top ten Google results, so the work that earns a good ranking is the same work that earns consideration for an answer. Fan-out rank is the newer idea worth understanding, because an engine breaks one question into several related queries behind the scenes and pulls from what surfaces across all of them, which rewards depth across a topic, so one page tuned to one phrase does poorly.
Topic cluster ranking scored 8.9 for the same reason. A cluster of connected pages that each rank well beats one heroic page, which is an argument for internal linking and for covering a subject across several pages. That is the pillar model in the AI-Ready workshop plan, where the 30 day gameplan starts with four pillar topics rather than keywords, plans one per week with one blog and a set of social posts built from it, so the cluster fills in from several directions.
What content shape gets cited?
The shape that matches what was asked. Intent-format match scored 9.0, and best-of questions favor listicles and comparison tables while how-to questions favor step-by-step guides, so the format is part of the answer and not decoration around it. Answer near the top scored 8.8, which rewards putting the answer in the first line of a section, ahead of any build-up.
Then the qualities that make a passage quotable. Factually specific scored 8.3, explicit phrasing 8.1, citing your sources 8.0, and self-contained passages 8.0, which together describe writing that an engine can lift one paragraph out of without the paragraph collapsing. Vague, hedged prose that only makes sense in sequence gives a model nothing it can safely quote.
Which popular tactics score lowest?
An llms.txt file scored 2.0 and finished last of the 23 factors, despite the attention it gets as a quick fix. Domain authority scored 5.0, so being a big site helps less than being the site that answers the question. Structured data scored 5.6, lower than its reputation suggests among people who treat schema as the whole AEO play. Structured data still guides search engines and earns rich results, so it belongs in the work, and the score says it will not carry the citation on its own.
Brand and entity trust scored 6.8, above domain authority, and entity consistency scored 5.8. So a recognizable, consistently described business does better than a large one, which is a slower and more durable kind of advantage.
Where does freshness fit?
Freshness scored 7.0 with a caveat that changes how you apply it, because the value varies by query. A question about a current event needs recent data and a question about a durable concept does not, so a blanket policy of republishing everything on a schedule spends effort where no engine is asking for it. Update the pages whose answers move and leave the ones whose answers hold.
How do you turn this into a priority order?
Work down the scores. Confirm the page is reachable and indexed, check that no preview directive is suppressing your text, phrase titles and subheads the way your buyer phrases the question, match the format to the intent behind it, put the answer at the top of each section, make each passage specific enough to stand alone, and cite your sources. Treat schema, freshness, and an llms.txt file as tail-end work.
That order also explains why AI visibility rarely comes from one clever move, since the highest-scoring factors are unglamorous and cumulative. Our AEO and SEO checklist walks the same ground in checklist form, and what AEO means and how it differs from SEO covers the vocabulary if the terms are new.
The short answer to the question in the title is that engines cite pages they can reach, that already rank, whose headings match the question, and whose paragraphs can be quoted without repair. Meta tags belong in that list twice, once as a directive that can exclude you outright and once as the title that helps a machine classify you. Content Cucumber writes SEO content built for both the results page and the engines answering without one.
Let us help you get found in AI search results
AI Search Optimization →Keep Reading
Frequently Asked Questions
URL accessibility, at 9.5 out of the 23 factors scored. If a crawler cannot reach the page, no content or authority signal compensates, so blocked paths, login walls, client-side rendering of the substance, and stale robots rules remove a page from the pool before quality is judged.
Yes, heavily. Search rank scored 9.4 and came second of 23 factors, fan-out rank scored 9.3 and came third, and 38% of AI Overview citations come from the top ten Google results. The work that earns a good ranking is the same work that earns consideration for an AI answer.
An engine breaks one question into several related queries behind the scenes, then pulls from what surfaces across all of them. It scored 9.3, third of 23, and it rewards covering a topic in depth across connected pages, since one page tuned to one phrase surfaces for fewer of those queries.
It scored 2.0 and finished last of the 23 factors, despite the attention it gets as a quick fix. Treat it as tail-end work, well below page accessibility, ranking, and how your headings match the question.
Structured data scored 5.6, which places it low among the 23 factors and below its reputation among people who treat schema as the whole AEO play. It still guides search engines and earns rich results, so it belongs in the work, and the score says it will not carry the citation by itself.
Writing an engine can lift one paragraph from without the paragraph collapsing. Answer near the top scored 8.8, factually specific 8.3, explicit phrasing 8.1, citing your sources 8.0, and self-contained passages 8.0. Vague, hedged prose that only makes sense in sequence gives a model nothing it can safely quote.
