How a generative engine builds an answer

The distinction to make before any other is whether the system went and read something for that question or not. A model can answer purely from what it absorbed in training: in that case no document is involved at answer time, so there is nothing to cite. Or it can retrieve pages from the web and build the answer on those: only in that second case does a source exist, and only there can you become one.

It is a difference that completely changes what you can do. In the first case your brand appears or does not appear based on how present it is in the training data, which is a closed set and months or years old. In the second it appears based on what was retrievable and pickable at that moment, which is a set you can work on now.

In practice engines decide which route to take themselves, question by question. A request about a general concept often does not trigger retrieval; a request about a vendor, a price, a recent development, or a comparison almost always does. This explains an observation that confuses many people: the same company appears for some questions and vanishes for others with no apparent logic. The logic exists, and it sits upstream of visibility.

When retrieval kicks in, the chain is made of distinct steps filtering on different criteria. They are worth knowing separately, because failing at the second is a different problem from failing at the fourth, and they are fixed by work with nothing in common.

Because an answer has to fit in a few lines, and a system producing one does not need your whole page: it needs the part that answers. So documents get split into portions, and it is the portions that are evaluated and used.

The consequence is the most counterintuitive thing about the whole subject: you do not compete with your page, you compete with your paragraphs. A page that is excellent overall, where the answer to a specific question is spread across three distant sections, is less likely to be used than a mediocre page where that answer sits compact in five lines.

Hence a writing rule that looks stylistic and is actually structural: every passage carrying useful information has to be readable alone. If understanding it requires the previous paragraph, the extracted portion arrives mutilated, and a system checking whether the text supports the claim will tend to drop it.

It also works in reverse, as a diagnosis. If you notice you are cited often on one topic and never on another you cover on the same page, the second topic is very likely written in a way that does not survive the cut.

Because citation is a separate decision, taken afterwards. The system may have read your page, used it to form the answer, and then named someone else. That is not a malfunction: choosing whom to cite rewards different criteria than retrieval does, and tends toward the sources that are most specific, most attributable, and most corroborated by others.

This is why measurement needs two separate columns. Recording only whether the brand appears in the answer leaves two opposite situations indistinguishable: not having been read, and having been read and dropped at the citation step. The first is a problem of accessibility or of the content not existing; the second is a problem of specificity and authority. Treating them the same leads to spending on the wrong side.

The most useful operational signal I have found is this: if one of your pages appears among the retrieved sources but your name does not appear in the answer text, the content is there and is not attributable enough. Usually it means it says things that are true and generic, which the system can take without having to say who from. A figure only you hold, or a statement that only makes sense attributed, changes that behaviour.

A large one, and in a way that penalises being alone. When several independent documents say the same thing, verification passes more easily and the claim enters the answer with more confidence. When only the interested company's own site says it, the system is looking at an uncorroborated claim made by a party with a stake in it.

That does not mean it gets discarded. It means it weighs less, and that where the system has to pick a few sources it tends to prefer the ones that confirm each other. It is why presence in third-party sources produces effects no amount of work on your own site replicates: it is not about links, it is that the same claim exists in two places that do not control each other.

There is an uncomfortable flip side too. If a piece of outdated information about you is repeated across several external sources and the correct one exists only on your site, the wrong version is the better-corroborated of the two, and it will win. Fixing external sources is not a courtesy, it is maintenance.

Directly on three of the five stages, only indirectly on the other two. Knowing which is which is what separates a plan from a list of good intentions.

Worth stating, because a lot of unearned confidence circulates on this topic. That the stages exist, and that some systems ground the answer in retrieved documents and return references, is documented by the vendors. Exactly how each one selects, weighs, and orders sources is not: these are proprietary systems, and nobody publishes the criterion.

So the level of certainty differs per claim. That retrieval happens at the level of portions rather than whole documents is consistent with all observable behaviour and with how these systems are built, but it is not a published figure. That citation is a decision distinct from retrieval is observed directly, because you do see a source used and not named.

The rule I follow is to treat the mechanism as a useful model, not as a specification. It serves to choose where to act and to avoid conflating two different problems. It does not serve to promise a result, and anyone using it to promise is describing a system they do not have the documentation for.