AI Citation Factors: What Actually Decides If AI Cites Your Site

What counts as an AI citation (and what does not)

Three different things get called the same thing, and conflating them is why so much advice contradicts itself.

Everything in this article is about the first one. It is the only one measurable on a per-page basis, the only one that sends traffic, and the only one with a real evidence base behind it.

Is a citation worth chasing? Seer Interactive's 2026 analysis found that being cited inside a Google AI Overview is associated with roughly 120% more organic clicks per impression and a 41% lift in paid clicks compared with appearing on the same query uncited. That is the honest case for doing this work: the AI answer is going to appear either way, so the question is whether your URL is inside it.

The 23 citation factors, ranked by strength of evidence

Shepard's scoring is the most useful artefact in this field right now because it weights each factor by three things — how often the finding repeats across studies, how strong the underlying data is, and whether official documentation or a patent backs it. A 50-million-query study outranks a 10-query blog experiment. Scores run 9.5 down to 2.0.

The grouping into gates is ours. It is what turns a list into a work order: the gates run in sequence, and effort spent at gate three while gate one is broken is wasted.

Read the top of that list carefully. Four of the five strongest factors are things an SEO was already doing in 2019. The distinctively "AI" items — schema, llms.txt, entity work — cluster at the bottom. That is not an argument for ignoring them; it is an argument about sequence.

Gate 1: can the engine actually reach the page?

The highest-scoring factor in the entire dataset is the least glamorous one: the URL has to be fetchable when the engine asks. This is where most invisible sites are actually losing, and it is the fastest thing to fix.

Three failure modes account for nearly all of it:

• Bot rules that block the retrieval agents. Every major vendor now runs at least two user agents with different jobs — one that gathers training data and one that fetches pages live to answer a question. Blocking the training crawler is a defensible business decision. Blocking the retrieval agent is opting out of citations. Our guide to robots.txt for AI crawlers lists which agent does which job.
• Edge protection that returns a challenge instead of HTML. WAF rules, JavaScript interstitials and managed bot-fighting modes frequently answer AI fetchers with a 202, a 403 or a challenge page. Your browser sees the article; the agent sees a wall. This is the single most common cause we find on client sites, and it is documented in why AI doesn't cite your website.
• Preview controls, set years ago and forgotten. This one scores 9.2 and almost nobody checks it. A nosnippet directive, a max-snippet limit or a data-nosnippet attribute around your key paragraph tells Google how much text it may show — and by extension how much it can lift into an AI surface. Publishers who added these during the scraping panic of 2024 are now paying for it in citations.

The uncomfortable trade-off, stated plainly: content protection and AI citation pull in opposite directions. You can block the crawlers and keep your content out of models, or allow the retrieval agents and be quotable. There is no configuration that gets both. Decide deliberately rather than by default, and revisit the decision as licensing deals change.

Gate 2: you get cited largely because you rank

Search rank scores 9.4 and fan-out rank 9.3 — the two highest content-side factors in the set. Ahrefs found that 38% of AI Overview citations come from pages in Google's organic top 10, with the overlap increasing as you widen the window. The relationship is not perfect, and it has loosened as engines pull deeper into the index, but it is the strongest lever available.

The nuance worth internalising is fan-out. Google's own documentation describes it: the model generates a set of concurrent related queries and retrieves against all of them, so a question about fixing a weedy lawn also fires searches for herbicides, chemical-free removal and prevention. You are not competing for one query. You are competing for eight, and a page that answers only the headline question loses to one that covers the fan.

This is why topic-cluster ranking (8.9) shows up as its own factor. Ranking mid-table for six related queries beats ranking first for one, because six chances to be retrieved beats one chance. It is also why Google warns, in the same guide, against spinning up a separate thin page for every fan-out variant — that trips the scaled content abuse policy. Depth on one page, breadth across a genuine cluster; not a page per query.

Gate 3: write passages a model can lift without editing

Engines do not read your page the way a person does. They retrieve chunks of it, and there is a cap on how much of any single URL makes it into the context window. Dan Petrovic's work on Gemini's grounding snippets showed a strict retrieval cap per URL, with content near the top of the page far likelier to make the cut. So the game is: put the citable claim early, and make it stand on its own.

Four rewrites that map directly onto the highest-scoring extraction factors:

Note what is not on that list: chunking your article into fragments, or rewriting it in a robotic register for machines. Google's mythbusting section addresses both directly — there is no requirement to break content into tiny pieces, and no need to write in a special way for AI. Structure means headings, sections and tables. It does not mean shredding the page.

One more that gets missed: intent-format match (9.0). A "best X" query pulls comparison tables and lists; a "how to X" query pulls ordered steps; a "what is X" query pulls a definition in the first hundred words. If your format fights the query type, the engine reaches past you for something shaped correctly.

Gate 4: the off-site signals you cannot fix on your own site

This is the layer that makes practitioners uncomfortable, because none of it lives in your CMS. Ahrefs analysed 75,000 brands against AI Overview appearances and found the three strongest correlations were all off-site:

The distribution is harsher than the correlations suggest. Brands in the top quartile for web mentions averaged 169 AI Overview mentions; the quartile below averaged 14. The bottom half averaged between zero and three. Roughly 26% of the brands studied had no AI Overview presence at all.

Two caveats that vendors selling "mention packages" tend to leave out. First, the study filtered for domains with a Domain Rating above 40 and took each one's top-ranking keyword — so it describes how established brands separate from each other, not how a new site breaks in. Second, and stated plainly by the authors: correlation is not causation. Brands that are widely discussed are also brands with budgets, PR teams and products people talk about. Buying unlinked mentions is not the mechanism, and Google explicitly names seeking inauthentic mentions as something to ignore.

What the data does support is duller and more durable: be a real brand that real publications, forums and videos have reason to reference, and use your own name consistently everywhere so the engine can resolve you to one entity.

What barely moves the needle

Three items sit at the bottom of the evidence table, and all three are heavily marketed.

• llms.txt (2.0). No credible experiment shows an effect on citations, and Google states outright that Search ignores the file. We covered the server-log evidence in detail in does llms.txt help SEO. Ship it only if it costs you a build step.
• Structured data (5.6). The honest position is more interesting than either camp's version. Google says schema is not required for generative AI features and there is no special markup to add. Yet nearly every study that looks at schema and citations finds a small positive relationship, consistently. Keep it for rich results and entity clarity; do not expect it to carry an AI strategy.
• Domain authority (5.0). A third-party link metric, weakly related to citations, and not something any engine consumes. Useful as a rough proxy for the brand signals above, useless as a target.

If a proposal leads with these three, it is selling the parts of the map that are easiest to invoice for, not the parts that move. The labels those proposals arrive under are worth understanding on their own terms: we cover what answer engine optimization actually means, and what the research behind generative engine optimization did and did not measure.

A 30-minute self-audit, in gate order

Work the gates in sequence and stop at the first failure — there is no point tuning passage structure on a page an agent cannot fetch.

1. Fetch your own page as an agent. Request it with a retrieval agent's user-agent string from outside your network and check for a 200 with real HTML. A 202, 403, or a challenge page is your answer.
2. Grep for preview controls. Search the source and headers for nosnippet, max-snippet, data-nosnippet and noarchive. Anything you did not deliberately add today should come off.
3. Disable JavaScript and reload. Whatever disappears is content the engine may never see.
4. Check rank for the fan-out, not the head term. List the eight questions a person would ask around your topic, and check where you sit on each.
5. Read your first 150 words as a stranger. Is there one self-contained, specific, quotable sentence that answers the query? If the answer arrives in paragraph nine, it is not being retrieved.
6. Search your brand name and read what the web says about you, not what you say about yourself.

Items one to three are where the surprises live, and they are also the ones a site owner is least equipped to find alone, because everything looks fine from a logged-in browser on a whitelisted IP. If you want the fetch tests, preview-control sweep and rank picture run for you and returned as a prioritised list, that is exactly what our $20 audit and roadmap delivers inside 24–48 hours.

What this adds up to

The evidence points somewhere unfashionable: there is no separate AI-citation discipline waiting to be bought. The factors that decide citations are crawl access, search rank across a query fan, extractable writing and a brand the engine has heard of. Three of those four are ordinary SEO executed with more attention to the fan-out and to passage design; the fourth is public relations.

What genuinely is new is the engineering layer at gate one — retrieval-agent access, preview controls, render-side visibility, and the agent-facing plumbing that sits alongside it. That layer is invisible in every ranking tool, it breaks silently, and it gates everything else. Where it sits relative to SEO, AEO and GEO is mapped out in our pillar guide to AI search visibility.

If your pages rank and still never get cited, the failure is almost always at gate one or gate three, and both are diagnosable in an afternoon. Book a free AI-visibility consultation and we will tell you which gate you are stuck at before you spend anything.