The short answer
Lead with the answer. Put the direct answer in the first two lines, phrase your heading as the exact question a person asks, and write short factual sentences a model can lift without editing. Add real specifics, name tools, dates and steps. Cut hype and filler. A page worth quoting reads like the answer already, not the wind-up.
Start with the answer, not the wind-up
Write the answer in the first two sentences, then explain. AI assistants scan for a passage that resolves the query on its own. If your opening paragraph sets the scene before it commits to a point, the model has nothing clean to lift, so it moves to a page that states the conclusion up front. Treat the first two lines as the quote you want to see in the answer box.
Match your heading to the exact question a person types or speaks. "How do I lower my energy bill" beats "Energy saving considerations". Models map a query to headings that mirror it word for word, so plain question phrasing wins over clever titles. Read your heading aloud. If it does not sound like something a real person would ask a search bar, rewrite it until it does.
Keep every sentence doing one job. A quotable sentence carries a single fact and needs no surrounding context to make sense. Long sentences with three clauses force the model to trim them, and a trimmed sentence often loses your meaning. Short declarative lines survive the copy and paste. Aim for sentences a reader could screenshot and send to a colleague without any explanation attached.
Give the model facts it can trust
Specifics get quoted, vague claims get ignored. Name the real thing: the crawler is GPTBot for ChatGPT, Google-Extended for Google, and PerplexityBot for Perplexity. Give a real timeframe, a real step count, a named tool. Models favour passages that contain concrete detail because concrete detail signals a source worth citing. A sentence with a number, a name or a date reads as evidence, while a sentence built on adjectives reads as marketing and gets skipped.
Do not invent numbers to look authoritative. A made up statistic can be checked, and once a model or a reader catches one error the whole page loses trust. If you do not have a figure, describe the mechanism instead of guessing a percentage. Accuracy is what earns repeat citation over months, because assistants revisit sources that held up. One wrong claim is enough to get a page quietly dropped from future answers.
Answer the follow up questions on the same page. When someone asks how to do a thing, they next ask how long it takes, what it costs them in effort, and what can go wrong. Cover those in their own short sections. A page that resolves the whole cluster of related questions becomes the single source a model returns to, rather than one of several it has to stitch together from different sites.
Structure the page so a machine can parse it
Use headings, lists and steps to break the answer into liftable blocks. A model reads structure as meaning. A numbered steps block tells it these are sequential actions, and a bulleted list tells it these are parallel options. Flat walls of text hide your best sentences inside paragraphs the model has to work to understand. Give each distinct point its own heading so the answer to each sub question sits in a clearly labelled place.
Put the summary near the top and the detail below. A short answer block after the heading lets the model grab the gist, while the sections underneath give it the depth to support a longer response. This mirrors how people read too. Nobody reads a help page top to bottom. They scan for the block that matches their question, so make each block self contained and titled with the question it answers.
Here is the order that works on almost any answer page.
Question as heading
State the exact question in the words a person uses.
Answer in two lines
Resolve the question before any context.
Detail in short sections
One sub question per heading, with facts and names.
Follow up questions
Answer the next things a reader will ask.
Cut the writing that gets a page skipped
Delete hype before you publish. Words like unlock, supercharge and game changing carry no information, so a model cannot quote them as an answer and a reader learns nothing from them. Every sentence that describes how great something is takes the place of a sentence that could have stated a fact. Read each line and ask what it teaches. If the answer is nothing, remove it and the page gets stronger and shorter at once.
Remove the throat clearing at the start of sections. Phrases like in this article we will explore delay the answer and waste the opening lines that matter most. Start each section with its point. The reader and the model both want the conclusion first, then the reasoning. A section that opens with its answer gives the assistant a clean block to quote and gives the human a reason to keep reading past the first line.
Check the page reads well when a single paragraph is pulled out alone. That is exactly how an assistant will use it. Copy any paragraph into a blank document and see if it still makes sense with no heading and no page around it. If it needs the rest of the page to be understood, rewrite it to stand on its own. Self contained paragraphs are the unit that assistants quote.