Anthropic describes its goal for Claude in three words that turn out to explain a lot about citation: helpful, honest, and harmless. Honesty is the one that matters most here. A model built to avoid stating things it cannot support will be careful about which sources it repeats, and that caution shapes every decision about how Claude chooses sources. It is not looking for the flashiest page. It is looking for the one it can quote without risking an inaccuracy. Once you internalize that, the tactics for getting cited stop being tricks and start being a discipline.
Retrieval first: Claude can only cite what it can reach
Nothing else matters if Claude cannot find your page. When Claude answers with web search enabled, it retrieves live pages and cites what it uses. When it answers from training alone, it speaks from memory and names no URL. So the first question in how Claude chooses sources is simply whether your page is reachable at the moment it searches, which means indexed, relevant to the query, and surfaced by the search step Claude relies on.
This makes ordinary search visibility a prerequisite rather than an afterthought. A page that never surfaces cannot be cited no matter how good it is. But retrieval is only the entry ticket. Getting fetched puts you in the room; it does not win you the quote. Plenty of retrievable pages get read and set aside because they failed the signals that come next. Treat retrieval as the floor you have to clear, then spend most of your effort on what happens after the fetch.
There is a useful mental picture here. Think of retrieval as Claude assembling a short stack of candidate pages for a question, and everything after that as a contest among the stack. Getting into the stack is necessary but not sufficient, and the two failure modes look completely different. A page that never makes the stack has a visibility problem, solved by the ordinary work of ranking and indexing. A page that makes the stack every time and still never gets quoted has a quality-of-passage problem, solved by rewriting. Diagnosing which one you have saves you from applying an SEO fix to a writing problem, or the reverse.
What makes a passage quotable
Claude builds an answer by drawing specific passages from specific sources, so the unit of citation is the passage, not the page. A quotable passage states one clear idea plainly, answers a question the user actually asked, and can stand on its own without the surrounding paragraphs to prop it up. When a passage meets those conditions, Claude can lift it and attribute it cleanly, which is exactly what it wants to do.
The opposite is a passage that only makes sense in context, hedges its central claim, or takes several sentences to arrive at the point. Claude can still read and understand it, but it becomes a poor candidate to quote, because quoting it accurately would require rebuilding the context around it. Faced with that, Claude reaches for a cleaner source that said the same thing more directly. Quotability is not about writing simply. It is about writing so each important sentence can survive on its own.
The practical upshot is that quotability is a property you can engineer sentence by sentence, not a mysterious quality some pages happen to have. Every section is a chance to write one line that could stand alone as a correct answer, and the pages that get cited are simply the ones where more of those lines exist. Treat it as a craft rather than luck and your citation rate follows.
The six signals Claude weighs
Six signals do most of the sorting. The first is passage relevance, how precisely a chunk answers the exact question. The second is directness, whether the answer leads or hides. The third is accuracy, whether the claim holds up, because an honest model is wary of repeating shaky statements. The fourth is source credibility, the sense that the site and author know the subject and can be trusted on it.
The fifth is corroboration, whether other trusted sources agree, which lowers the risk of quoting the claim at all. The sixth is freshness, weighted heavily for topics that move and lightly for topics that do not. Notice that only one of the six, credibility, is about who you are. The other five are about how you wrote the page and how the wider web treats your claims. That is encouraging, because it means a smaller brand willing to write with precision and earn corroboration can compete with larger names on the signals that carry the most weight.
Why does Claude sometimes refuse to cite a ranking page?
You will see it happen: a page ranks well, Claude clearly retrieved it, and it still went uncited. The usual reason is that the page never stated the specific answer in a form worth quoting. It was relevant enough to fetch and too vague to lift. Claude read it, found nothing it could attribute with confidence, and moved on to a source that had committed to a clear claim.
The second reason is an accuracy or corroboration gap. If a page makes a bold statement that Claude cannot square with what other trusted sources say, the honesty constraint kicks in and the page gets skipped rather than repeated. This is why overreaching claims backfire in AI search. A page that promises more than it can support does not just fail to persuade, it disqualifies itself from citation, because the one thing an honest model will not do is repeat something it suspects is wrong.
The Citable Sentence Test
Here is a test you can run on any page in five minutes. Call it the Citable Sentence Test. Read each section and ask: is there a single sentence here that Claude could quote, word for word, as a correct and complete answer to a real question, without needing the rest of the paragraph to make sense of it? If yes, that section is citation-ready. If no, that section is invisible to citation no matter how well it ranks.
Most pages fail the test in most sections, and the failures cluster in predictable places: introductions that warm up instead of answering, paragraphs that qualify before they claim, and conclusions that summarize without stating anything new. The fix is mechanical. Move the answer to the front of the section, make it a complete sentence, and let the nuance follow. Run the Citable Sentence Test across a page, rewrite every section that fails it, and you will have done more for how Claude chooses sources in your favor than any amount of keyword tuning.
Accuracy and the honesty constraint
It is worth dwelling on accuracy, because it is the signal most people underweight. Claude’s design pushes it toward statements it can support and away from ones it cannot. For a source, that means accuracy is not just good practice, it is a ranking factor. A precise, verifiable claim is safe to repeat. A vague or inflated one is a liability the model routes around.
The practical implication is to write claims you can back. Prefer specific, checkable statements over sweeping ones. When you make a strong claim, make it the kind that other sources also support, so Claude finds agreement when it looks. This is the opposite of the puffery that fills a lot of marketing copy, and that contrast is your opening. In a field where most pages overreach, the page that states exactly what is true, and no more, becomes the one an honest model trusts enough to quote.
Corroboration lowers the risk of quoting you
Corroboration is the quiet multiplier. When Claude considers repeating a claim, the existence of other credible sources saying the same thing makes the claim safer to attribute. A statement that appears only on your own site is a statement with no second opinion, and an honest model treats it more cautiously than one echoed across the trusted web. This is where earned coverage does work that on-page optimization cannot.
The move is to get your key claims out of your own domain and into places Claude’s retrieval will reach: credible publications, industry references, and third-party coverage that repeats your position. Each independent source that agrees with you raises the odds that Claude will treat your claim as established rather than unverified. Corroboration is slow to build and hard for competitors to erase, which is precisely what makes it worth building. It turns your best claims from assertions into consensus, and consensus is what an honest model quotes without hesitation.
The training layer you cannot edit but can influence
Everything above concerns retrieval, the mode where Claude searches and cites live pages. But Claude also answers from training, the knowledge it absorbed before any given conversation, and that layer works differently. You cannot edit what a model already learned, and you cannot make it cite a live URL when it is answering from memory. What you can do is shape what future training tends to absorb about you, which is a slower and more indirect game than optimizing a page, but a real one.
The mechanism is presence and consistency across the open web over time. Models learn from vast amounts of public text, and a brand that appears often, clearly, and consistently in that text, associated repeatedly with specific topics, becomes part of what a model understands about the world. A brand that is absent, inconsistent, or contradictory in how it presents itself gives future training nothing stable to latch onto. This is why durable reputation, the kind built through sustained coverage and a coherent story, eventually shows up in how a model talks about your category even without searching.
The practical takeaway is that retrieval and training reward the same underlying behavior from different angles. Clear, credible, consistent presence makes you quotable when Claude searches and memorable when it does not. You optimize the retrieval layer with sharp pages you can edit today, and you influence the training layer with a body of public work you build over months and years. Neither replaces the other. Understanding how Claude chooses sources means understanding that some of the game is won on the page in front of you and some of it is won across the whole web over time, and the brands that show up in both are the ones a model both quotes and remembers.
Turn the signals into an editing pass
None of this is useful as theory. Make it an editing pass you run on real pages. Confirm the page is retrievable. Run the Citable Sentence Test and rewrite the sections that fail. Check every strong claim for accuracy and for whether it is corroborated elsewhere. Update anything time-sensitive so freshness works for you. Then test by asking Claude the questions your page should answer and watching which sources it names. Do that pass on your ten most important pages and you will feel the difference in how Claude chooses sources, because you will have stopped hoping to be quotable and started engineering it, one clear sentence at a time.