What does it mean for a sentence to extract?
A model composing an answer reuses the content of a source as small facts, each with a subject, a relation, and a value: ftes.ai meters work in tokens; a Digital FTE starts at Level 0 autonomy; the token wallet stops work at zero. Each of those can be lifted into an answer without the sentence that carried it. A sentence that cannot be reduced to that shape has nothing for the model to carry, however well it reads.
Entity density is simply how many such facts a passage contains per hundred words. It is not a score anyone publishes; it is a property a writer can see by trying to reduce each sentence and noticing which ones refuse.
Sentence one: the adjective that is not a value
"Our platform is powerful and flexible enough for any team." The subject is vague, the relation is a state of being, and the values are adjectives. Reduced, it yields "platform is powerful", which no assembling model can use, because it is a claim with no content and no way to check it. Sentences of this shape make up a large share of ordinary product copy and contribute nothing to retrieval or citation.
The rewrite names the entity and states something specific: "ftes.ai meters every AI call in tokens and charges three user tokens for each token of real model cost." That reduces to two usable facts, both checkable against the product.
Sentence two: the pronoun subject
"It works inside the tools you already use, so there is nothing to rip out." Inside a page this is clear. Extracted, the subject is unknown, and a fact with an unknown subject is discarded rather than guessed. The same sentence with its subject restored — "A Digital FTE works inside the tools a team already uses" — carries a subject that a model can attribute and a reader can verify.
This is why the ftes.ai standard forbids a pronoun as the subject of a section's opening sentence. Pronouns inside a sentence are fine; a pronoun that opens a passage takes the subject away exactly where it is most needed.
Sentence three: the relation that is a feeling
"Managers love how much time the reviews save them." The subject is a group we cannot name, the relation is an emotion, and the value is an unquantified saving. Beyond being unextractable it is also a claim about customers, which the ftes.ai content standard refuses outright unless it is something the platform can show. Copy of this shape is where invented statistics usually enter a library, because the next draft tends to attach a number to it.
The extractable version describes the mechanism instead: "The Approvals tab shows each proposal with its evidence, so a manager approves or sends back without opening the source system." That is a fact about the product, and a reader can confirm it in one click.
Sentence four: the mechanism with a name
"A Digital FTE holds any outward-facing work at an approval gate until its manager approves it, and the decision is written to an audit trail." The subject is named, the relation is concrete, and there are two values, both of them named mechanisms rather than adjectives. The sentence reduces to three facts and reads naturally. It is also the shape that gets cited, because named mechanisms are what an answer needs when it explains how something works.
Naming mechanisms — approval gate, fact gate, token wallet, autonomy ladder, morning brief — is the single highest-yield habit in this library. A named thing can be looked up, compared, and carried; a benefit cannot.
How much entity repetition is too much?
Repeat the entity as the subject of each section and wherever a new claim begins; use pronouns freely inside a sentence. That is roughly two or three full-name uses per passage, which reads as ordinary careful writing rather than as keyword stuffing. The failure mode to avoid is the opposite of density: a passage that names its entity once in the heading and then refers to "the platform", "the tool", and "the system" interchangeably, which leaves a model unsure whether three things are being discussed or one.
Consistency of naming matters more than frequency. One name per thing, used the same way every time.
What extraction shape cannot control
Writing extractable sentences does not decide whether a model uses them. Selection depends on the query, on the competing corpus, and on how the system weighs sources, none of which is visible from outside. The controllable part is whether a sentence contains a fact at all. A page of adjectives has nothing to be selected; a page of named mechanisms with specific values has something to offer every question it is retrieved for.
ftes.ai reports these separately for that reason: extraction shape is a property of the writing and is audited at draft time, while citation is an observed event reported only when it is observed.