DeepSeek chooses sources by retrieving live pages for a question, then selecting the ones that answer it most directly and can be trusted enough to quote. That selection only happens when web search is active; without it, DeepSeek answers from trained knowledge and names nothing. So every question about how DeepSeek chooses sources is really a question about its search path, and on that path the model behaves like a careful librarian: it gathers candidates, keeps the ones that clearly answer the question, and prefers the ones the rest of the collection agrees with.
Six factors drive that decision, and they arrange themselves into a shape worth picturing. Rather than a flat checklist, think of a triangle whose corners are your page, the question being asked, and the wider web that can confirm or contradict you. DeepSeek is constantly measuring the tension between those three points. The factors below are the specific ways it takes that measurement, and seeing them as one geometry, rather than six unrelated rules, is what makes them usable.
Citation only happens on the search path
The first thing to fix in your head is that citation is conditional. DeepSeek has two modes: answer from memory, or search the web and answer from what it finds. Citations belong to the second mode. If you are trying to appear as a named source, you are optimizing for the search path, and everything below assumes that path is active. On the memory path, the work is different and slower, closer to long-run reputation than to page-level tactics.
This distinction saves you from chasing the wrong fix. A brand frustrated that DeepSeek “never cites us” is sometimes looking at memory-path answers, where nothing gets cited by design. The productive test is to ask a question with search enabled and see who gets quoted. Those results are the contest you can actually influence page by page, and they are the ones the six factors describe. Get clear on which mode you are studying, and the rest of the analysis stops being confusing.
The Corroboration Triangle
Here is the shape to hold onto. Call it the Corroboration Triangle: your page at one corner, the exact question at another, and the corroborating web at the third. A citation becomes likely when all three sides are short. Your page has to sit close to the question, meaning it answers precisely what was asked. Your page has to sit close to the web, meaning credible sources agree with your claims. And the question has to sit close to the web, meaning the topic is one the open web actually covers, so corroboration is possible at all.
The triangle is useful because it tells you which side to shorten. If your page answers the question but no other source backs the claim, the page-to-web side is long, and you get skipped for lack of corroboration. If credible sources agree with you but your page does not actually answer the specific question, the page-to-question side is long, and you lose on relevance. Most citation failures are one long side, not a general weakness. Find the long side and you know what to fix, which is far more efficient than improving everything at once.
Factor one: can the search reach you?
The first factor is retrievability, and it is binary. If DeepSeek’s search cannot crawl and index your page, you are not a candidate, and no other factor matters. This is where brands lock themselves out without realizing it, through robots rules that block crawlers, pages hidden behind interactions, or content that never gets indexed. The fix is unglamorous: confirm the relevant crawler is allowed, confirm your important pages are indexed, and confirm they surface for the queries you care about.
Retrievability is the corner of the triangle you must secure before drawing any sides. It is also the cheapest to fix and the most commonly neglected, because it feels too basic to be the problem. It often is the problem. Before analyzing relevance or corroboration, verify that the search can reach you at all, because a beautifully corroborated, perfectly relevant page that the crawler cannot see contributes exactly nothing to your citations.
Factor two: does the passage answer the question?
The second factor is relevance at the passage level. DeepSeek matches parts of pages to the specific question, so what matters is whether some passage on your page answers that question directly and near the surface. A page that is generally about your topic but never states the precise answer loses to a page that opens a section with exactly the answer being sought. This is the page-to-question side of the triangle, and shortening it is a writing task more than a technical one.
Write self-contained answers that lead their sections. Put the direct response in the first sentence, use a heading that matches the question in plain language, and resist burying the answer under setup. When your passage and the question line up cleanly, DeepSeek can match and lift it without effort, which is what the selection step rewards. Vague, meandering pages force the model to work, and a model with a dozen candidates does not work hard for any single one.
Factor three: is the claim verifiable?
The third factor is verifiability, the page-to-web side of the triangle. DeepSeek prefers to quote claims it can confirm, so a specific, checkable assertion beats a vague or promotional one, and a claim echoed across credible sources beats a claim that lives only on your domain. This is corroboration doing its work. When the model can look around and find agreement, quoting you is low-risk. When your claim stands alone, quoting you is a gamble the model would rather not take.
This factor is where earned coverage pays off directly. Placing your key positions in reputable third-party sources shortens the page-to-web side, converting lonely claims into corroborated ones. It is also why concrete beats grand: a precise, verifiable statement invites confirmation, while an inflated one invites doubt. If you want DeepSeek to trust your claims, make them the kind that can be checked, and make sure that when the model checks, the credible web agrees.
Factors four and five: identity and consistency
The fourth and fifth factors are who is behind the page and whether your identity holds together across the web. A named author with real expertise and a publisher with a track record lower the risk of citation, because the model can attribute the source to a known, credible entity. Consistency multiplies this: when your brand is described the same way across many places, the model resolves you to a clear identity rather than a fuzzy one, and clear identities are easier to trust and cite.
These factors reward coherence over cleverness. Put real authors with real credentials on substantive pages, keep your descriptions of what you do consistent, and avoid the scattered, contradictory presence that makes an entity hard to pin down. You are giving DeepSeek the evidence it needs to treat you as a specific, credible source rather than an anonymous page it stumbled on. Identity is slow to build and hard to fake, which is exactly why it works as a trust signal.
Factor six: freshness where it counts
The sixth factor is freshness, and it matters only for questions whose answers change. For a time-sensitive topic, DeepSeek favors a current page over a stale one, because an outdated answer is worse than a slightly less authoritative fresh one. For evergreen questions, freshness barely moves the decision. The skill is targeting: keep the pages that cover moving topics genuinely current, and stop worrying about freshness on pages where the answer is stable.
Freshness is targeted maintenance, not a universal chore. Spending effort updating evergreen pages that did not need it is waste; letting a time-sensitive page go stale is a self-inflicted loss to any competitor who kept theirs current. Know which of your pages sit on moving ground, maintain those, and let the rest win on the factors that actually decide them. Applied precisely, freshness is how you beat a stronger but outdated source on exactly the questions that reward being up to date.
A worked example of the triangle
Picture a mid-size software brand that wants DeepSeek to cite its page on reducing onboarding drop-off. The page is well written and lives on a fast, indexed site, so retrievability is fine and the page-to-question side is short: it answers the exact question near the top. Yet it never gets cited. Walking the triangle, the long side is obvious once you look: the brand’s central claim, that a specific onboarding change lifts retention, appears nowhere except its own blog. There is no corroboration for DeepSeek to find, so quoting the claim is a risk the model declines.
The fix is not more content, it is shortening the page-to-web side. The brand gets its onboarding data and point of view covered in a couple of reputable industry publications and referenced in a few credible roundups. Now when DeepSeek retrieves the page and checks the claim, the web agrees, and the citation becomes safe to make. Nothing about the page itself changed. What changed is that the claim stopped being lonely, which is exactly the shift the Corroboration Triangle predicts will move a page from retrieved-but-skipped to retrieved-and-cited. Most stuck pages are this example with the names swapped.
Reading the triangle in your own results
Put the six factors back into the triangle and you have a diagnostic you can run on your own pages. Ask a question with search on, see who gets cited, and compare their pages to yours across the three corners. Are they more retrievable, more precisely matched to the question, or more corroborated across the web? Usually one side of your triangle is clearly longer than theirs, and that side is your assignment. This beats generic advice because it points at your specific weak link rather than a list of everything you could improve.
For most brands, the long side is corroboration, because they built strong pages on their own domain and never seeded their claims across the trusted web. Shortening that side, through earned coverage and consistent presence, tends to improve several factors at once: your claims become verifiable, your identity becomes clearer, and the web starts agreeing with you when DeepSeek checks. Understand how DeepSeek chooses sources as a matter of geometry, find your longest side, and fix that, and citations follow from the structure rather than from luck.