How Do AI Systems Choose Which Websites to Cite?

An AI system cites a website because, at the moment it composed a particular answer, that site’s page was retrieved for the question, matched what the answer needed, and was usable as support for a specific claim. There is no master list of “citable websites,” no application process, and no published ranking of trusted domains that a business can get itself added to.

That is the honest headline, and it is worth stating before anything else: no AI platform publishes its citation selection function. OpenAI, Google, Anthropic, Microsoft, and Perplexity describe pieces of how their systems find and use information, but none of them publishes the weights, thresholds, or rules by which a given page becomes a citation in a given answer. Anyone who claims to sell you that formula — a “citation score,” a guaranteed placement method, a proprietary list of what ChatGPT “wants” — is inventing it. What follows is what can actually be said: the mechanics that are documented, the behavior that can be observed, and the inferences that are reasonable, labeled as which is which.

Step one: a page must be retrieved before it can be cited

The most underappreciated fact about AI citations is that citation is the last step, not the first. When an AI system answers a question using outside information, the sequence is roughly: the question is interpreted, a retrieval step searches an index (or the live web, depending on the product), a set of candidate pages or passages comes back, and the model composes an answer drawing on some of them. The citations attached to the answer point at sources the system actually used — or, in some products, sources closely related to what it used.

This sequence has a hard consequence: a page that is never retrieved can never be cited. It does not matter how accurate, well-written, or authoritative the page is. If the retrieval step does not surface it for the question being asked, it is invisible at the only moment that matters. This is documented behavior in the general sense — retrieval-augmented generation works this way by design, and the platforms’ own documentation describes search and grounding steps feeding answer generation. The specific ranking functions inside each product’s retrieval step are not published.

So the practical question splits in two. First: does your page get found when the question is asked? Second: once found, does it get used? Most businesses obsess over the second question and never examine the first.

What “getting found” depends on

Retrieval systems — whether a search index behind an AI product or a dedicated retrieval index — work by matching a question to text. Reasonable inferences from how these systems demonstrably behave, consistent with documented search and embedding technology:

The page must be crawlable and indexed. If a crawler cannot access the page, or the page never entered the index the system draws on, retrieval has nothing to find. This is a binary gate, and it fails silently — no one notifies you.
The page’s text must actually contain the substance. Retrieval matches questions to passages. A page that is about your services in a marketing sense but never states them in plain text — what you do, where, for whom, under what conditions — gives the retriever little to match against a specific question.
Relevance is question-specific, not site-wide. A domain does not get cited; a page gets cited, for a particular question, because a particular passage matched. The same website may be cited for one question and absent for a near-identical one, because the retrieval step returned different candidates. This is observed behavior anyone can reproduce by asking variations of the same question.
The text must be extractable. Content locked inside images, rendered only by scripts the crawler does not execute, or buried in formats the indexer handles poorly, may exist for human visitors and effectively not exist for retrieval. Documented crawler guidance from the platforms themselves emphasizes accessible text content.

What happens after retrieval: use, corroboration, and selection

Once a set of candidate pages is in front of the model, a second selection happens: which of them actually shape the answer and get cited. This stage is the least documented of all, and honesty requires saying so. What can be said:

Corroboration is observable. When several independent sources state the same fact, answers tend to reflect it confidently. When one page states something that many other sources contradict, answers tend to follow the majority record or hedge. Whether this happens through retrieval ranking, model behavior, or both varies by product and is not published. But the practical implication is not inference — it is arithmetic: a claim that appears on one page, contradicted or unsupported everywhere else, is structurally weak no matter how well the page is written.

Specificity gets used. Passages that directly answer the question — with the actual fact, figure, definition, or instruction — are more useful to a composing model than passages that gesture at a topic. This is a reasonable inference from how these systems behave, and it matches the platforms’ published guidance to publishers, which consistently emphasizes clear, substantive, well-structured content.

Citations are not endorsements, and not every statement gets one. Products differ: some cite nearly every factual claim, some cite sparingly, some cite sources for an answer as a block. A page can inform an answer without being cited, and a citation can appear for a narrow claim while the rest of the answer comes from elsewhere — including the model’s own training. Being uncited in an answer does not prove your page played no role; being cited does not prove the system “trusts” your site in general.

A concrete example

Consider two plumbers in the same town. The first has a one-page website: a logo, a phone number, a photo of a van, and the words “Quality plumbing services — call today!” The second has pages that state, in plain text, the services offered, the neighborhoods served, the conditions handled (“burst pipes, water heater replacement, sewer line inspection”), pricing approach, and licensing — and the same facts appear, consistently, on its Google Business Profile, two directories, and a local news article about a community project.

A customer asks an AI assistant: “Who installs tankless water heaters in [town], and what does it typically cost?” The retrieval step looks for passages about tankless water heater installation in that town. The first plumber’s page contains almost no matchable text — it may not be retrieved at all. The second plumber’s pages match directly, and the local article corroborates that the business exists and does this work. The answer names the second plumber and cites its pages. Nothing about this required the system to have an opinion about which business is “better.” It cited what it could find, read, and support.

Misconceptions

“There is a list of approved sources.” No platform has published one, and observed citation behavior — small local sites cited alongside national publications — contradicts a fixed allowlist for general questions.

“Domain authority determines citations.” Authority-style metrics are third-party SEO inventions. Established sites are cited often, but largely because they publish retrievable, specific, heavily corroborated text at scale — the mechanism, not the metric, is doing the work.

“Schema markup gets you cited.” Structured data helps machines identify entities and facts; it does not substitute for retrievable text and corroboration, and no platform documents it as a citation trigger.

“Once cited, always cited.” Answers vary between runs, sessions, and product updates. A citation is an event, not a status.

What this means for your business

You cannot apply for citations, and you should distrust anyone selling the formula. What you can do is remove the failure points: make sure your pages are crawlable and indexed, state your facts in plain extractable text, keep those facts consistent everywhere they appear, and earn independent corroboration. That is the entire documented and observable basis on which citations happen.