LLM SEO: The Ranking Factors That Actually Get You Cited
LLM SEO ranking factors ranked by evidence: why classic search rank is the #1 input to AI citation, why llms.txt is useless, and a retrofit that proved it.
Classic search rank is the number-one controllable input to AI citation
If you want one LLM SEO ranking factor to fix first, it is your position in ordinary Google search. Not a manifest file, not schema, not a clever prompt. Rank.
The strongest evidence we have says so plainly. Cyrus Shepard’s Zyppy meta-analysis synthesized 54 experiments, studies, and patents across ChatGPT, Gemini, and Perplexity, then scored 23 factors by how repeatably they hold up, how strong the evidence is, and whether platform documentation backs them. Traditional search rank scored 9.4 out of 10, second only to whether the page is crawlable at all. Paired with the finding that 38% of AI Overview citations come from the top 10 Google results, the picture is not subtle: the answer engine mostly quotes pages that already win the blue links.
So the fastest path to getting cited is the least novel advice in this whole category. Rank the page. Everything else is a multiplier on top of a page that already ranks, not a substitute for one.
The single most controllable input to getting cited by AI is the least exciting one: rank in classic Google search first.
The AI citation ranking factors, scored and ranked
Here is the full Zyppy ranking. Read it as a priority list, not a checklist where every box weighs the same. The gap between the top of this table and the bottom is the difference between work that pays and work that feels productive.
| # | AI citation ranking factor | Score /10 |
|---|---|---|
| 1 | URL accessibility (crawlable) | 9.5 |
| 2 | Classic search rank | 9.4 |
| 3 | Fan-out rank | 9.3 |
| 4 | Preview control | 9.2 |
| 5 | Query-answer match | 9.2 |
| 6 | Intent-format match | 9.0 |
| 7 | Topic-cluster ranking | 8.9 |
| 8 | Answer near the top | 8.8 |
| 9 | AI-ready structure | 8.6 |
| 10 | Factually specific | 8.3 |
| 11 | Explicit phrasing | 8.1 |
| 12 | Cites sources | 8.0 |
| 13 | Self-contained passages | 8.0 |
| 14 | Content visibility | 7.6 |
| 15 | Freshness | 7.0 |
| 16 | Brand / entity trust | 6.8 |
| 17 | Length | 6.7 |
| 18 | Language | 6.3 |
| 19 | Entity consistency | 5.8 |
| 20 | Structured data | 5.6 |
| 21 | Known source | 5.4 |
| 22 | Domain authority | 5.0 |
| 23 | llms.txt | 2.0 |
The top of the list clusters around three ideas: be reachable, already rank, and answer the exact question cleanly. Notice that classic-search signals (rank, fan-out rank, topic-cluster ranking) occupy three of the top seven slots. The so-what: there is no separate “AI SEO” discipline you can run instead of SEO. It is the same muscle, pointed at longer, more specific questions.
What each factor means for a writer, not an engineer
A score is useless until it turns into an edit. Here is how the top factors translate into things you actually change on the page, drawn from retrofits I run on client blogs.
Be reachable and already ranking (factors 1 to 3). Confirm your AI-crawler user agents are not blocked, then treat the page’s Google rank as the real project. If it sits on page two, the citation work is premature. Get it into the top 10 for the head term and its fan-out variants first, because that top-10 pool is where 38% of citations are drawn from.
Match the question, and answer it near the top (factors 5, 8, 13). This is where most of my edits live. Take the exact question a buyer types, “how to measure employee engagement,” and make sure a clean, self-contained answer to it sits high on the page, not buried in an FAQ at the bottom. On one retrofit I moved the definitional question out of the FAQ and set it as an H2 with the answer directly beneath the heading. A passage a model can lift without needing the paragraph above it is a passage that gets quoted.
Match the format the query wants (factor 6). A comparison query wants a table. When I audited why a competitor kept getting pulled into answers for “best employee reward” queries and the client did not, the difference was structural: the competitor ran a standardized “strengths, weaknesses, best for” block for each option, and the models lifted it wholesale. We added a quick-look comparison chart in the same shape. Give the model the format it is looking for and you stop losing to a worse page with better structure.
Be specific and cite your sources (factors 10, 12). Vague, hedged copy does not get quoted, because there is nothing quotable in it. A named number, a dated result, a linked primary source: those are the sentences an answer engine can stand behind. Pages that cite get cited.
The proof: 8 retrofits, +18 points of AI presence in 30 days
None of the above is theory. On a B2B HR-tech account, I took eight blog pages that were underperforming in AI answers, applied the changes above, and re-measured their presence (the share of tracked AI responses to a target question where the brand appeared) 30 days later.
| Target question | Platform | Before | After |
|---|---|---|---|
| Best employee retention strategies | ChatGPT | 7% | 60% |
| Best employee reward ideas | AI Overview | 16% | 60% |
| Best employee feedback tools | ChatGPT | 0% | 20% |
| How to measure employee engagement | ChatGPT | 0% | 11% |
Across all eight pages, average presence rose from about 30% to about 48%, a lift of 18 points, and seven of the eight held or improved. The two most striking cases started at zero: pages the models never surfaced, now cited. Sit with that for a second. The content largely existed already. What changed was accessibility, answer placement, and format, the exact factors at the top of the table.
Here is the part that ties the whole argument together. The same retrofits moved the classic rankings too. The reward-ideas page climbed to position 3, up 13 spots, over the same window its AI presence quadrupled. I did not run separate campaigns for search and for AI. I improved the page, and both scoreboards moved.
I did not optimize for Google and then optimize for AI. I improved the page once, and the ranking and the citation rose together.
llms.txt and the busywork that feels like AI SEO
The bottom of that table is a to-don’t list. llms.txt scores 2.0, and Zyppy’s read is blunt: no credible evidence it influences citations at all. Structured data (5.6) helps a little, consistently, but the effect is small. Domain authority as a standalone metric (5.0) is weaker than the specific, page-level signals above it.
This matters because the busywork is seductive. Adding a file to your root feels like doing AI SEO, and it is measurable in the sense that you can confirm the file exists. It just does not move the outcome. Spend that hour rewriting the top of a page so it answers the query in the first sentence. That hour compounds. The manifest file does not.
On-page gets you into the running, off-site decides the answer
Even a perfectly retrofitted page is playing for a minority share of the sources an answer is built from. Across the six B2B SaaS accounts I track, 88% of the sources AI cites in a category are third-party, roughly 9% are competitors, and only about 3% are the brand’s own domain. On average a brand’s own site is the fourth most-cited source about its own category.
The so-what is a change in where your effort goes once the page is citable. On-page factors are necessary and entirely in your control, so do them first. But being present in the answer at all depends on the sources the model already trusts: the software-review directories, the community threads, the video and analyst coverage that show up in every vertical. I break the source data down in where ChatGPT gets its information, and why being cited is not the same as being recommended in AI citations are a vanity metric. Getting the on-page factors right earns you the ticket. It does not win the raffle by itself.
The whole playbook, on-page and off, is the generative engine optimization work, and the answer-box mechanics live in how to rank in AI Overviews. I run the ranking pulls and single-variable retrofits behind this piece with my own Claude and Search Console setup.
FAQ
What are the GEO ranking factors?
They are the signals that decide whether an answer engine cites your page, and the strongest evidence we have ranks them by weight. The top of the list, from Zyppy’s 54-study meta-analysis, is URL accessibility (9.5), classic search rank (9.4), fan-out rank (9.3), preview control (9.2), and query-answer match (9.2). The through-line is that traditional search performance and clean, on-topic answers dominate, while novelties like llms.txt (2.0) sit at the bottom.
Does classic SEO still matter for LLM SEO and AI search?
More than anything else you can control. Classic search rank scores 9.4 out of 10 as an AI citation factor, and 38% of AI Overview citations come from the top 10 Google results. AI search is not a replacement discipline; it is traditional SEO pointed at longer, more specific questions. If a page does not rank, optimizing it for AI is premature.
How do you get cited by ChatGPT?
Rank the page for the question and its fan-out variants, make sure it is crawlable, and put a clean, self-contained answer to the exact query near the top in the format the query wants. Then earn presence off-site, because most of what ChatGPT cites in a category is third-party. In a retrofit on a B2B HR-tech account, applying the on-page half of that moved several pages from zero presence to being cited within 30 days.
Does llms.txt help you get cited by AI?
No, on the current evidence. It scores 2.0 out of 10 in Zyppy’s ranking, the lowest of 23 factors, with no credible study showing it influences citations. Adding it is close to harmless and close to useless. The time is far better spent on query-answer match and answer placement.
What is the single most important AI ranking factor?
Being crawlable (9.5) technically tops the list, but that is table stakes: fix it once and move on. The most important factor you will keep working on is classic search rank (9.4). It is the one that compounds, and in my own retrofits the edits that lifted AI presence lifted Google rankings at the same time.