Why does structure decide citation?
Structure decides citation because language models and the retrieval systems in front of them work on passages. When an engine answers a question, it retrieves candidate chunks of text, typically a section or a paragraph, judges each on how well it answers, and cites the ones it used. A page is never read as a whole at that moment.
So the question for every section is whether it would make sense, and answer something, if it were the only text the engine saw. Content structured that way gets cited; content that depends on the surrounding page does not.
How should the opening be written?
The opening should be a direct answer to the primary question in 40 to 60 words, starting with a definition or a verdict sentence: a Digital FTE is, or the answer is. It should be liftable verbatim into an AI Overview and natural when read aloud by a voice assistant.
It should not begin with in this guide, or with context the reader did not ask for. On ftes.ai this opening is the tldr field, and the SPEC-035 standard fixes its length and form for every piece.
How should sections and headings be written?
Each heading should be a question the section answers or a claim the section supports, such as how does the approval gate work or the approval gate makes autonomy reversible. Bare labels such as overview or benefits tell a retriever nothing about what the section contains.
Each section's first sentence should restate the entity and the topic, so that the section is self-contained: the approval gate on ftes.ai requires, rather than it requires. Nothing in a section should say as mentioned above or as we will see, because the reader of that chunk has not seen either. Keep sections to roughly 120 to 220 words, which is long enough to answer and short enough to lift.
- Heading: a question or a claim
- First sentence: entity and topic restated
- Body: the answer, then the reasoning, then the specifics
- No cross-references to other parts of the page
- 120 to 220 words
How should facts be stated?
Facts should be stated as explicit triples: the entity, the attribute, and the value, in one sentence. A Digital FTE starts at Level 0 on the autonomy ladder is a triple an engine can extract and attribute. The platform is designed with safety in mind is not; it has no attribute and no value.
Named mechanisms are triples waiting to happen. Naming the approval gate, the fact gate, the token wallet, and the morning brief, and stating what each does, gives an engine specific claims to cite. Generic benefit language gives it nothing.
How do you cover the fan-out and the voice?
Cover the fan-out by listing, before writing, the five to nine sub-questions an engine would split the primary query into: what it is, how it works, what it costs, what the risks are, how it compares, when to use it, who it is for. Answer each in its own section or FAQ entry, so that whichever sub-question the engine runs, a section of yours answers it.
End with FAQs whose answers are short, at least two of them under 30 words, and natural when spoken. Those are what a voice assistant reads aloud and what an answer box lifts. Link three related pieces so the engine sees the page as part of a cluster rather than an orphan.
What does a section written to the method look like beside one that is not?
As a labelled example, a section not written to the method might be headed Benefits and begin: as noted earlier on this page, this makes things much easier for teams, with several advantages worth considering. Retrieved alone, that chunk names no entity, answers no question, and depends on text the engine does not have. A section written to the method is headed how does the approval gate keep an AI employee safe and begins: the approval gate on ftes.ai requires a named person to approve every action a Digital FTE proposes before it happens, and each proposal carries its evidence.
The second chunk names its entity in the first sentence, states a triple, answers the question its heading asks, and could be quoted alone in an AI answer without confusing anyone. It continues with the reasoning and the specifics, runs to roughly 150 words, and never refers to anything elsewhere on the page. The test is to read each with the rest of the page hidden: the first fails all three questions about entity, question, and standalone use; the second passes them, which is the whole method in one comparison.