What makes a language model cite
one source over another

Every load bearing claim below carries a confidence label and a recheck date.

What is the unit of retrieval in AI search?

The unit of retrieval in AI search is the passage, not the page. A retrieval system scores spans of text against a reformulated query and assembles the answer from the winners, so what competes for a citation is a section of your article.

So a section answering one narrow sub question completely can enter an answer the page it sits on would never have ranked for. And a page that answers a big question beautifully, where the middle needs the beginning, can rank well and never get lifted.

Why is the standard content brief the wrong shape?

The standard content brief is the wrong shape because it optimises a document for a system that scores documents, and retrieval scores passages. One page, one keyword, two thousand words, written to flow. Still correct for ranking, close to useless for extraction.

The fix is unglamorous. A 2,000 word article built as eight self contained 250 word sections beats a 2,000 word article that flows, even when the second is better writing. The cost is real: it makes articles read like reference material instead of essays, and if your readers came for the essay you are trading their pleasure for machine legibility.

The instinct comes from the sending side of the group, where we run two named motions. WideNET is high volume systematic angle testing across the full addressable market: you do not know which angle lands, so you test many in parallel and let the data pick. Spearhead is signal triggered against the hottest slice, used when you already know. Sections behave like WideNET. You cannot predict which of your eight gets pulled into an answer, so you write eight that could each carry one, then read the grounding queries in Bing Webmaster Tools to see which did. I have been wrong about which section would travel more often than I have been right.

What is a Quotable Unit?

A Quotable Unit is a section of a page that answers one complete question without needing anything around it. Five properties, each with a test that takes under a minute.

The fifth is the one almost nobody does and the one that pays. A passage saying "fetched on 6 August 2026 with the documented GPTBot user agent, byte count taken from the raw response body" travels with its origin attached, so whoever quotes it carries your name. "In our experience" does not survive extraction, because once it is separated from your byline nobody can tell whose experience it was.

How do you test whether a section is quotable?

Copy the section into a blank document, hand it to somebody who has not read the article, and run the five tests above in order. That is the whole procedure, and the order matters because a section that fails the first test cannot pass the rest in any meaningful way.

• Standalone first. If the reader has to ask what "this" refers to in the opening sentence, stop and rewrite the opening sentence.
• Then definitional. Ask them what the section claims. If they cannot say it back after reading only the first sentence, the definition is buried.
• Then numeric. Ask them what the number is divided by. If they cannot find the denominator on the page, the number is not usable and should either gain a denominator or come out.
• Then bounded. Ask them how many separate claims the section makes. More than one means it should be two sections with two headings.
• Then attributable. Ask them who measured it. If the only available answer is your byline, the section will lose its origin the moment it is extracted.

This is a heuristic built from the trait list in the 2026 research, not a measured predictor of citation, and we label it that way everywhere it appears.

Why does a number with no denominator get ignored?

Because careful retrieval drops it and careless retrieval misquotes it, and there is no third outcome. Both are worse than never publishing the figure.

I learned that on the sending side of this group, where every threshold that decides whether a campaign lives or dies is written as positive replies divided by emails sent, and where a row carrying a bare percentage either gets skipped or gets the wrong threshold applied to it. An answer engine does that same job with less context and no way to ask a follow up question. The thresholds themselves, and the arithmetic of rate definitions generally, belong to the group's go to market math property. This page stops at the retrieval consequence, which is the part that concerns us.

Here is a number from this site with its denominator sitting next to it. On 6 August 2026 the parent domain's homepage returned 5,179 bytes to GPTBot, against an empty shell measured on the same domain the same day at 5,179 bytes. Zero bytes of body copy. The blog index on the same fetch returned 34,123 bytes. Read 5,179 on its own and it is a file size. Read it against the shell and it is a complete finding, which is why the shell measurement travels in the same sentence every time we publish the byte count.

Do AI assistants prefer original data or listicles?

Both, through different mechanisms, and separating the evidence quality matters more than picking a winner. Ranked list formats dominate the largest published citation sample available: Evertune's May 2026 analysis of roughly 25,000 unique URLs and around 400 million citations across six surfaces found that listicles took a clear majority of citations, well out of proportion to their share of the URL set. That is a platform reporting its own telemetry, so the effect size carries the contested label and no percentage from it appears on this page. The sample description stays because it is what makes the source auditable. The direction is safe.

Which citation statistics will we not print, and why?

Five claims circulate in almost every pitch in this category. None can go on a page here, because none has a published method.

We name the claims and leave the digits off. That is a deliberate choice and it is the reason this table can be quoted safely: an answer engine lifting the left column of a table carries whatever is in it, so a refusal list printed with the numbers intact is a distribution channel for the numbers. Naming the claim is enough, because the argument is about the missing method and the digits add nothing to it.

What decides citation that has nothing to do with your page?

Whether something you do not own says the same thing about you. Multiple 2026 measurements agree on broad B2B category queries: the majority of citations go to third party sources instead of vendor sites, because an engine asked to evaluate a category reaches for something that looks like an evaluation and your homepage does not look like one however well it is written. Directories and other people's roundups do, whether or not they are better sources than you. The specific percentages circulating are vendor published and are not going on this page.

The implication runs against our own commercial interest, so here it is plainly. Owned properties are the weakest citation surface for exactly the commercial queries that convert, and a perfect page can lose one to a directory listing of a third of its quality. What owned media wins is the informational cluster, a smaller prize than most pitches imply.

Who should ignore this page, and what do we still not know?

Three kinds of company should ignore everything above. The first is by far the most common.

Companies that cannot publish anything original. Formatting improves the odds of a section being lifted. It does not create a reason to lift it.

Anyone who needs this working this quarter. On the outbound side I can put a build schedule in writing, because infrastructure warm up and onboarding both have known lengths and the client signs off on messaging and lead lists before a single email leaves. There is no equivalent for citation, from me or anyone. That asymmetry is worth sitting with before you sign anything in this category, because it is the difference between a discipline with dates and one without.

Four things nobody outside the platforms knows: how OpenAI ranks candidates once it has fetched them, whether ChatGPT runs its own index alongside Bing, which provider powers Claude's search, and whether the December 2024 rendering measurement still holds, since most 2026 sources re-cite it instead of re-running it. Ask any advisor those four. The answer you want is the one that says nobody knows.