Most pages bury the answer. The heading asks a question, and the first paragraph clears its throat. An LLM can fix that in bulk, but only if you stop it from inventing things. Below are three prompts, a before-and-after example, and a review checklist. Everything here is an untested template: I did not run these prompts on a live model for this post. The example output is illustrative, written by hand.
For the evidence behind answer-first writing, and what Google does and does not say, read the AEO checklist first. This post is the editing step that follows.
Which sections should you rewrite first?
Do not rewrite everything. Rank sections by value and by how badly the answer is buried:
- Pages that already get impressions for a question-shaped query.
- Sections whose first paragraph is long or starts with background.
- Headings that are clearly a question a person would type.
The script in the AEO checklist prints the first-paragraph length under every H2. Sort by length, and start with the longest.
Prompt 1: find the answer that is already there
Use this before any rewrite. It forces the model to quote your text instead of composing new facts.
You are editing a web page section. Below is the heading and the section text.
Heading: [HEADING]
Section: [PASTE SECTION]
Task:
1. Say in one sentence what question the heading asks.
2. Quote, word for word, the sentence or sentences in the section that answer it.
If no sentence answers it, reply exactly: NO ANSWER IN TEXT.
3. List every number, date, name and product claim in the section.
Do not add information that is not in the section.
If the reply is NO ANSWER IN TEXT, the problem is content, not phrasing. Write the answer yourself, from a source.
Prompt 2: rewrite the opening only
Rewrite only the opening of this section so the first one or two sentences
answer the heading directly. Keep the rest of the section unchanged.
Rules:
- Use only facts that appear in the section. Do not add numbers, dates,
names, prices or claims.
- Keep any hedges ("may", "in my tests", "as of [date]") that the original has.
- Maximum 60 words for the opening.
- Plain language, no marketing words.
Heading: [HEADING]
Section: [PASTE SECTION]
Output: the new opening, then a list of any original sentence you moved or removed.
The last line of that prompt is the safety net. A model that has to list what it moved or removed is easier to audit than one that silently returns a new paragraph.
Prompt 3: pick the right shape for the answer
Given the heading and section, tell me whether the answer is best shown as:
(a) a short paragraph, (b) a numbered list, (c) a bulleted list, or (d) a table.
Give one sentence of reasoning. Then produce that version using only facts in the text.
My rule of thumb: a definition is a paragraph, a procedure is a numbered list, a comparison is a table. Google's featured snippets page did not give format guidance, so this is my judgement.
What does a good rewrite look like?
Illustrative example, written by hand, not model output.
Heading: How long does a robots.txt change take to apply?
Before:
Robots.txt files have been part of the web for a long time and are used by many sites to control crawlers. Different bots behave differently, and it is useful to understand how each one works. For OpenAI's crawlers, the documentation says changes can take around 24 hours.
After:
It can take about 24 hours for OpenAI's systems to adjust after you update robots.txt, according to OpenAI's bots documentation. Other crawlers may differ, so check each one's own documentation.
What changed: the answer moved to the front, the source stayed, the hedge about other crawlers stayed, and no new claim appeared. The cut sentences carried no information.
How do you review the output?
Never publish rewrites unread. Run this checklist on each one:
- Every number, date and name in the new opening exists in the original.
- No claim got stronger. "Can" did not become "will".
- The source attribution survived.
- The opening still makes sense with the heading removed.
- The rest of the section does not repeat the opening.
- A reader who stops after two sentences would not be misled.
If you have an evaluation habit, add a regression test. Keep a list of facts per section and diff them before and after. A simple approach: extract numbers with a regular expression from both versions and compare the sets. If a number appears only in the rewrite, reject it.
What not to do with these prompts
- Do not ask the model to add statistics "to make it more authoritative". That is how invented numbers get in.
- Do not rewrite headings into questions mechanically. A forced question reads badly.
- Do not remove caveats to shorten the answer.
- Do not skip the date on time-sensitive facts. Prices and policies change, so keep "as of" wording, and update it when you re-check.
More prompt patterns for planning content around these ideas are in AI SEO content strategy prompts, and the GEO guide explains which on-page edits showed any effect in the research I could read.
