Prompts for SEO work that are worth keeping

Prompts are not a secret and there is no magic phrasing. A prompt that works carries the same four things a brief for a freelancer carries.

These three turn things you can see into things you can act on. Nothing in them asks the model to know anything.

PROMPT 1: the brief from a results page

Here are the top ten results for “[query]” in [country]:
[paste each title and its domain, one per line]

And the People also ask questions shown:
[paste the questions]

Tell me:
1) what the searcher wants, in one sentence
2) the sections that appear on most of these pages
3) the questions none of these pages answers
4) what kind of page ranks here (guide, comparison, product page, forum)
Use only what I pasted. Do not suggest anything you cannot point to in it.

For: turning a results page into an outline in five minutes. Cannot: tell you the volume of the query, the difficulty, or whether you can rank. Those come from a keyword tool and from your own site's strength.

PROMPT 2: the questions a buyer asks

Our product: [two sentences, what it does and for whom].
It does not: [the three things people wrongly assume].
Price: [what it costs, or “not public”].

List the twenty questions a buyer asks before paying, ordered by how early
in the decision they ask them. Phrase each one the way a person would type
it into Google, not as a marketing headline. No answers, only questions.

For: the FAQ of a pricing page, the headings of a comparison page, the objections a landing page has to answer. Cannot: know which of those questions people actually search, which is what Search Console and a keyword tool are for.

PROMPT 3: the outline from the brief

SEARCH: [query]
READER: [who they are, what they know]
ANGLE: [two sentences on what this page must contain that others do not]
MUST COVER: [the sub-questions from prompt 1]

Give me six to nine section headings, each phrased as the reader would ask
it, in the order that answers the search fastest. Under each heading, one
line saying what it holds and whether it is prose, a list, steps or a table.
No introductions, no conclusion section.

For: the skeleton, before any prose exists. Cannot: decide the angle. If you leave that line vague, the outline comes back as the same page everybody else has published.

PROMPT 4: the draft from facts

Write the section “[heading]” for the page above.

FACTS YOU MAY USE:
- [fact]
- [fact]
- [fact]
Nothing else. Where you need a fact I have not given, write [FACT NEEDED].

CONSTRAINTS: second person, British spelling, sentences under twenty-five
words, no exclamation marks, no em-dashes, no marketing register.
SHAPE: one short paragraph, then five numbered steps of one sentence each.
LENGTH: about 200 words.

For: a first draft whose facts are yours. Cannot: notice the fact you forgot. It will write around the gap or mark it, which is why you check the draft against your own list rather than against the draft's confidence.

PROMPT 5: the register rewrite

Rewrite the text below. Rules:
- plain second person, British spelling
- cut every adjective that carries no information
- no marketing register: no leverage, unlock, seamless, robust, empower,
  game-changer, dive in, elevate
- keep every figure exactly as written, do not round, do not add any
- keep it under [n] words
Return three versions, shortest first.

[paste the text]

For: a draft that reads like a brochure, yours or someone else's. Cannot: tell whether a figure is right. It will preserve a wrong number with great care, which is the correct behaviour and a reason to check separately.

PROMPT 6: the check pass

Read the page below as a hostile fact-checker.

List, as a table:
1) every number, date and named figure, with the sentence it appears in
2) every claim about what a company, tool or search engine does
3) every sentence that hedges without deciding (“may”, “can vary”, “some say”)

Do not fix anything. Do not comment on style. Only list what a reader could
challenge.

[paste the page]

For: the last read before publishing, when your own eyes have stopped working. Cannot: verify any of it. The table is a list of things for you to go and check, and half the value is seeing how many there are.

1. 1

   Search in a clean window, in the right country

   A private window, and the country you sell to. Your history and location change the page, so a page read through your own logged-in browser is not the page your buyers see.

2. 2

   Copy the titles with their domains

   The domain matters as much as the title: ten vendor pages mean a commercial query, ten guides mean an informational one, a forum in the top five means people want an opinion.

3. 3

   Copy the People also ask questions

   These are the sub-questions Google associates with the query. They become your headings more often than anything the model invents.

4. 4

   Note whether an AI overview opened, and who it cited

   The cited pages are the ones an assistant found quotable on this query. Their shape is worth copying: the answer early, in a form that can be lifted.

5. 5

   Paste all of it, then ask one question at a time

   A paste plus one instruction beats a paste plus five. Ask for the intent, then the gaps, then the outline, in separate turns.

A model does not sound unsure when it is inventing. These are the shapes to be suspicious of, in order of how often they cause damage.

• A round number attached to a behaviour: a share of searches, a percentage of users, an average anything. If you did not supply it, treat it as fiction.
• A named study with no link, or a link you have not opened.
• A feature or a limit of a tool, stated in the present tense. Product pages change every quarter and the model's picture is old.
• A policy paraphrased rather than quoted. The drift is small and it is always in the direction of what sounds reasonable.
• A date. Dates are the easiest thing in the world to get wrong and the easiest for a reader to check.
• Any sentence you would be embarrassed to defend if a customer emailed about it.

The test is not whether the claim is plausible. Every invention is plausible; that is what makes it dangerous. The test is whether you can put your finger on where it came from.

Three things sit outside the reach of any wording, and pretending otherwise is how thin pages get made.

• Your own data. Search volumes, your positions, your traffic, your conversion rate: the model cannot see any of it, and a prompt asking harder produces a better-phrased guess.
• First-hand experience. What broke when you shipped it, what your customers keep getting wrong, what you decided and why.
• The angle. Deciding what this page says that the ten ranking pages do not is the judgement you are paid for, and it has to exist before the prompt does.

Everything above assumes those three are yours. The prompts are for the typing.