How Do LLMs Choose Their Sources?

Three gates, not one decision

A source passes three separate gates before its name appears in an answer, and they are governed by different things. Retrieval decides which pages are considered. Selection decides which passage gets used. Attribution decides whether the name travels with the claim.

Treating this as one decision is the root of most bad advice in the field. Guidance about writing clearly is guidance about the second gate. It is good guidance and it is useless to a site failing the first, which is most sites that are entirely absent.

The gates also fail identically from outside. Whether you were never fetched, fetched and passed over, or used without credit, what you see is a paragraph that mentions somebody else. There is no error, no notification and no log you can read.

So the useful question is never "how do I get chosen". It is "which gate am I failing", and that is answerable with a few hours of work.

Gate one: retrieval

Retrieval is the question of whether your page is in the set being considered at all, and it is decided by things that have nothing to do with how well the page is written.

The first is permission. Answer crawlers are separately named clients, governed by rules that are frequently set by somebody solving an unrelated problem. A site can be fully open to search and closed to answers, with nobody having decided that and no notification sent.

The second is whether the content exists in the fetched document. A page whose substance appears only once scripts have run may deliver nothing to whatever is collecting it. This is invisible in a browser, where everything looks fine, and it affects a whole category of contemporary site.

The third is subject weight. Reaching for a source is a judgment about whether that source is a reasonable thing to reach for on this subject, and one page is thin evidence. This is where the real cost lives and where most sites quietly fail without any technical fault at all.

Gate two: selection

Once your page is in the candidate set, the competition happens at the level of the passage rather than the page. This is the gate most writing advice is about, and the advice is broadly right.

A passage that can be used states something clearly, states it early, and makes sense when removed from everything around it. Those three properties are all that separate a usable passage from an unusable one, and commercial writing habitually violates all three at once.

The most common violation is delay. Marketing pages set context, acknowledge complexity and then state the point. A reader going in order is well served by that. Anything assembling a short answer needs the point near the top.

The second is dependence. Sentences beginning with "this" or "it", comparisons whose other half was named two paragraphs ago, figures without their unit. Perfectly clear in place, meaningless once extracted.

The third is interchangeability. A sentence that could sit on any competitor's site adds nothing to an answer. Specificity is what makes a passage worth reaching for, and specificity means a real constraint, a real trade-off, or a real number you can stand behind.

Gate three: attribution

The third gate is the one you control least, and pretending otherwise produces disappointment.

What your page established can be used without your name attached. It can be blended with several other sources into a statement nobody is credited for. It can be attributed to a source that learned it from you. All of that is ordinary and none of it is within your control.

What appears to help is being the identifiable origin of something specific rather than one restatement among many. A distinctive framing, a number that traces to work you did, a distinction others do not draw. If a claim is recognizably yours, the name is more likely to travel with it.

This favors original work over comprehensive summary, which is awkward, because summary is cheaper to produce and performs perfectly well in search. A thorough overview of a settled subject is a good page and a weak citation candidate, precisely because nothing in it needs sourcing to anybody in particular.

What kinds of source get named most

Watching which sources appear across many answers shows a fairly stable pattern, and it is not flattering to brand-owned pages.

Places where people discuss things in their own words tend to appear often, because they contain phrasing that matches how questions are actually asked and because they carry the kind of specific, experiential detail that is useful in an answer.

Established publications appear often for the same reason they always have. They are long-lived, widely referenced, and their claims are easy to treat as settled.

Specialist independent sites do well on narrow subjects, which is the most encouraging part of the pattern for anybody building a cluster. Depth on a specific subject beats breadth and beats size.

Brand-owned pages appear least, and the reason is usually not bias against them. It is that they tend to be promotional, interchangeable, and written to persuade rather than to establish. A brand page that reads like a reference rather than a pitch does considerably better, and that is a writing decision rather than a structural disadvantage.

What does not appear to decide it

Several things people spend effort on do not seem to move this, and it is worth saying so plainly.

Keyword density does not. Repeating a phrase does not make a passage more usable and it makes it worse to read, which works against you at the second gate.

Total page count does not. A large site with shallow coverage of many subjects is competing subject by subject, and in each one it is small.

Recency alone does not. A recent page on a subject with no depth behind it does not outrank a well-established treatment. Recency matters where the subject genuinely changes and matters little where it does not.

Technical markup helps understanding but does not substitute for substance. Describing a page accurately in structured form is useful and it does not make an interchangeable page worth quoting.

Why the same question gives different answers

Anybody who tests this properly notices quickly that asking the same question twice can produce different sources, and that fact has consequences for how you measure anything.

The process producing an answer is not deterministic. Two runs of the same question can retrieve overlapping but different sets, and can select different passages from the same set. This is a property of the system rather than a fault in your measurement.

The practical implication is that a single reading is an anecdote. If you ask once, see a competitor named, and reorganize a quarter around it, you have built a plan on one sample of something that varies.

The correction is to ask each question several times across several days and record the distribution rather than the answer. What you are looking for is how often your name appears across many readings, and whether that frequency moves. Our page on benchmarking covers how to structure that properly.

What to actually do about it

The three gates produce a clear order of work, and it is not the order most programs follow.

Confirm you can be fetched, by the right clients, at the site and at the edge. This takes an afternoon and explains most cases of total absence.

Confirm your content exists without scripts. Fetch a page as a plain document and read it. If the substance is missing, that is your constraint and no amount of writing fixes it.

Pick one subject and go deep rather than covering four adequately. This is the expensive commitment and it is the one that decides retrieval over the long run.

Then rewrite openings so the answer arrives in the first two sentences, and make claims that survive extraction. This is fast, it is high return, and it improves the page for people as well. Our piece on writing for LLMs covers the sentence-level habits.

The honest trade-off

The uncomfortable conclusion of all three gates is that the highest-return strategy is concentration, and concentration is risky.

Depth on one subject is what moves retrieval, and choosing one subject means not choosing the others. If you pick a subject buyers do not actually ask about, you will have done the work correctly and gained nothing, and you will not find out for a quarter.

The mitigation is to derive the subject from language you did not write. Support tickets, sales notes, the way customers describe the problem in their own words. That is a better guide to what gets asked than any internal view of what your category is called.

The alternative strategy, covering everything moderately, is safer and reliably invisible here. It is worth naming that as a choice rather than letting it happen by default, because it is what a content calendar produces if nobody intervenes.

How this differs from ranking

It is worth being precise about the differences, because assuming the two work the same way produces confident wrong decisions.

Ranking produces an ordered list and the order is relatively stable. Citation produces a name or no name, and the same question asked twice can differ. There is no position to hold, no gradual climb to observe, and no rank to report to anybody.

Ranking rewards the page as a whole. Citation rewards the passage. A page can be comprehensive, well linked and genuinely the best resource on a subject while containing no single paragraph that survives extraction, and it will lose to a thinner page with one clean statement in it.

Ranking is measurable in tooling you already own. Citation is measurable only by asking, because a person who reads your name and then searches for you arrives as direct traffic with the assistant leaving no trace at all.

And ranking rewards breadth reasonably well. A large site covering many subjects accumulates authority that helps each of them. Citation is decided per subject, so a large site covering forty subjects shallowly is competing as a small site forty times over.

The one place they agree strongly is that both reward being genuinely referenced by other people, and both reward real depth. That overlap is where the shared work sits, and it is the reason the two disciplines are worth running as one program rather than two.

The practical consequence of all this is that success looks different and should be reported differently. A ranking program reports positions moving. A citation program reports how often a name appears across a fixed set of questions asked repeatedly. Anybody presenting the second as though it were the first has either misunderstood the mechanism or is describing a number they cannot support.

Four errors this framing prevents

Understanding the gates is only useful if it changes a decision, so here are the four decisions it changes most often.

Commissioning writing before checking access. The commonest expensive mistake. A quarter of content produced for a site nothing can fetch teaches you nothing and costs everything, and the access check that would have prevented it takes an afternoon.

Reading one answer as a result. Asking once, seeing a rival named, and reorganizing around it. The process varies between runs, so one reading is a sample rather than a finding.

Treating absence as a writing problem. It usually is not. It is usually a reachability problem or a depth problem, and both are structural rather than editorial.

Assuming a good page is enough. A single excellent page on a subject is thin evidence that you are the thing to reach for. The gate is about weight of coverage, and one page does not carry weight regardless of quality.

Bottom line

Three gates, in order. Retrieval is mechanical and is where most sites fail. Selection is editorial and is where most advice is aimed. Attribution is partly out of your hands and rewards being the origin of something rather than a restatement of it.

Diagnose before you write. Confirm you can be fetched and that your content exists without scripts, then commit to depth on one subject, then fix the openings. Effort spent on a later gate while an earlier one is closed produces nothing at all.

And measure by asking repeatedly rather than by watching your analytics, because the event you care about leaves no trace there.